From a42d96dfe047e28a9cd5463b33246f4456e92cdb Mon Sep 17 00:00:00 2001
From: Zengwei Yao <yaozengwei@outlook.com>
Date: Mon, 20 Jun 2022 13:40:01 +0800
Subject: [PATCH 1/9] Fix warmup (#435)

* fix warmup when scan_pessimistic_batches_for_oom

* delete comments
---
 .../ASR/conv_emformer_transducer_stateless/train.py   |  7 +++----
 .../ASR/pruned_transducer_stateless2/train.py         |  7 +++----
 .../ASR/pruned_transducer_stateless3/train.py         |  7 +++----
 .../ASR/pruned_transducer_stateless4/train.py         |  7 +++----
 .../ASR/pruned_transducer_stateless5/train.py         |  7 +++----
 .../ASR/pruned_transducer_stateless6/train.py         | 11 +++++------
 6 files changed, 20 insertions(+), 26 deletions(-)

diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/train.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/train.py
index 106f3e511..acaf1397f 100755
--- a/egs/librispeech/ASR/conv_emformer_transducer_stateless/train.py
+++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/train.py
@@ -1018,6 +1018,7 @@ def run(rank, world_size, args):
             optimizer=optimizer,
             sp=sp,
             params=params,
+            warmup=0.0 if params.start_epoch == 1 else 1.0,
         )
 
     scaler = GradScaler(enabled=params.use_fp16)
@@ -1078,6 +1079,7 @@ def scan_pessimistic_batches_for_oom(
     optimizer: torch.optim.Optimizer,
     sp: spm.SentencePieceProcessor,
     params: AttributeDict,
+    warmup: float,
 ):
     from lhotse.dataset import find_pessimistic_batches
 
@@ -1088,9 +1090,6 @@ def scan_pessimistic_batches_for_oom(
     for criterion, cuts in batches.items():
         batch = train_dl.dataset[cuts]
         try:
-            # warmup = 0.0 is so that the derivs for the pruned loss stay zero
-            # (i.e. are not remembered by the decaying-average in adam), because
-            # we want to avoid these params being subject to shrinkage in adam.
             with torch.cuda.amp.autocast(enabled=params.use_fp16):
                 loss, _ = compute_loss(
                     params=params,
@@ -1098,7 +1097,7 @@ def scan_pessimistic_batches_for_oom(
                     sp=sp,
                     batch=batch,
                     is_training=True,
-                    warmup=0.0,
+                    warmup=warmup,
                 )
             loss.backward()
             optimizer.step()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/train.py b/egs/librispeech/ASR/pruned_transducer_stateless2/train.py
index 36ee7ca74..55f32e119 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless2/train.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/train.py
@@ -883,6 +883,7 @@ def run(rank, world_size, args):
             optimizer=optimizer,
             sp=sp,
             params=params,
+            warmup=0.0 if params.start_epoch == 0 else 1.0,
         )
 
     scaler = GradScaler(enabled=params.use_fp16)
@@ -973,6 +974,7 @@ def scan_pessimistic_batches_for_oom(
     optimizer: torch.optim.Optimizer,
     sp: spm.SentencePieceProcessor,
     params: AttributeDict,
+    warmup: float,
 ):
     from lhotse.dataset import find_pessimistic_batches
 
@@ -983,9 +985,6 @@ def scan_pessimistic_batches_for_oom(
     for criterion, cuts in batches.items():
         batch = train_dl.dataset[cuts]
         try:
-            # warmup = 0.0 is so that the derivs for the pruned loss stay zero
-            # (i.e. are not remembered by the decaying-average in adam), because
-            # we want to avoid these params being subject to shrinkage in adam.
             with torch.cuda.amp.autocast(enabled=params.use_fp16):
                 loss, _ = compute_loss(
                     params=params,
@@ -993,7 +992,7 @@ def scan_pessimistic_batches_for_oom(
                     sp=sp,
                     batch=batch,
                     is_training=True,
-                    warmup=0.0,
+                    warmup=warmup,
                 )
             loss.backward()
             optimizer.step()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/train.py b/egs/librispeech/ASR/pruned_transducer_stateless3/train.py
index 92eae78d1..be9fa8f8b 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless3/train.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/train.py
@@ -1001,6 +1001,7 @@ def run(rank, world_size, args):
             optimizer=optimizer,
             sp=sp,
             params=params,
+            warmup=0.0 if params.start_epoch == 0 else 1.0,
         )
 
     scaler = GradScaler(enabled=params.use_fp16)
@@ -1061,6 +1062,7 @@ def scan_pessimistic_batches_for_oom(
     optimizer: torch.optim.Optimizer,
     sp: spm.SentencePieceProcessor,
     params: AttributeDict,
+    warmup: float,
 ):
     from lhotse.dataset import find_pessimistic_batches
 
@@ -1071,9 +1073,6 @@ def scan_pessimistic_batches_for_oom(
     for criterion, cuts in batches.items():
         batch = train_dl.dataset[cuts]
         try:
-            # warmup = 0.0 is so that the derivs for the pruned loss stay zero
-            # (i.e. are not remembered by the decaying-average in adam), because
-            # we want to avoid these params being subject to shrinkage in adam.
             with torch.cuda.amp.autocast(enabled=params.use_fp16):
                 loss, _ = compute_loss(
                     params=params,
@@ -1081,7 +1080,7 @@ def scan_pessimistic_batches_for_oom(
                     sp=sp,
                     batch=batch,
                     is_training=True,
-                    warmup=0.0,
+                    warmup=warmup,
                 )
             loss.backward()
             optimizer.step()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/train.py b/egs/librispeech/ASR/pruned_transducer_stateless4/train.py
index 48c0e683d..0fece2464 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless4/train.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless4/train.py
@@ -932,6 +932,7 @@ def run(rank, world_size, args):
             optimizer=optimizer,
             sp=sp,
             params=params,
+            warmup=0.0 if params.start_epoch == 1 else 1.0,
         )
 
     scaler = GradScaler(enabled=params.use_fp16)
@@ -992,6 +993,7 @@ def scan_pessimistic_batches_for_oom(
     optimizer: torch.optim.Optimizer,
     sp: spm.SentencePieceProcessor,
     params: AttributeDict,
+    warmup: float,
 ):
     from lhotse.dataset import find_pessimistic_batches
 
@@ -1002,9 +1004,6 @@ def scan_pessimistic_batches_for_oom(
     for criterion, cuts in batches.items():
         batch = train_dl.dataset[cuts]
         try:
-            # warmup = 0.0 is so that the derivs for the pruned loss stay zero
-            # (i.e. are not remembered by the decaying-average in adam), because
-            # we want to avoid these params being subject to shrinkage in adam.
             with torch.cuda.amp.autocast(enabled=params.use_fp16):
                 loss, _ = compute_loss(
                     params=params,
@@ -1012,7 +1011,7 @@ def scan_pessimistic_batches_for_oom(
                     sp=sp,
                     batch=batch,
                     is_training=True,
-                    warmup=0.0,
+                    warmup=warmup,
                 )
             loss.backward()
             optimizer.step()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/train.py b/egs/librispeech/ASR/pruned_transducer_stateless5/train.py
index e77eb19ff..eaf893997 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless5/train.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/train.py
@@ -980,6 +980,7 @@ def run(rank, world_size, args):
             optimizer=optimizer,
             sp=sp,
             params=params,
+            warmup=0.0 if params.start_epoch == 1 else 1.0,
         )
 
     scaler = GradScaler(enabled=params.use_fp16)
@@ -1072,6 +1073,7 @@ def scan_pessimistic_batches_for_oom(
     optimizer: torch.optim.Optimizer,
     sp: spm.SentencePieceProcessor,
     params: AttributeDict,
+    warmup: float,
 ):
     from lhotse.dataset import find_pessimistic_batches
 
@@ -1082,9 +1084,6 @@ def scan_pessimistic_batches_for_oom(
     for criterion, cuts in batches.items():
         batch = train_dl.dataset[cuts]
         try:
-            # warmup = 0.0 is so that the derivs for the pruned loss stay zero
-            # (i.e. are not remembered by the decaying-average in adam), because
-            # we want to avoid these params being subject to shrinkage in adam.
             with torch.cuda.amp.autocast(enabled=params.use_fp16):
                 loss, _ = compute_loss(
                     params=params,
@@ -1092,7 +1091,7 @@ def scan_pessimistic_batches_for_oom(
                     sp=sp,
                     batch=batch,
                     is_training=True,
-                    warmup=0.0,
+                    warmup=warmup,
                 )
             loss.backward()
             optimizer.step()
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/train.py b/egs/librispeech/ASR/pruned_transducer_stateless6/train.py
index 315c01c8e..9e9fc1440 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless6/train.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless6/train.py
@@ -74,9 +74,9 @@ from conformer import Conformer
 from decoder import Decoder
 from joiner import Joiner
 from lhotse.cut import Cut, MonoCut
+from lhotse.dataset.collation import collate_custom_field
 from lhotse.dataset.sampling.base import CutSampler
 from lhotse.utils import fix_random_seed
-from lhotse.dataset.collation import collate_custom_field
 from model import Transducer
 from optim import Eden, Eve
 from torch import Tensor
@@ -376,7 +376,7 @@ def get_params() -> AttributeDict:
             "distillation_layer": 5,  # 0-based index
             # Since output rate of hubert is 50, while that of encoder is 8,
             # two successive codebook_index are concatenated together.
-            # Detailed in function Transducer::concat_sucessive_codebook_indexes.
+            # Detailed in function Transducer::concat_sucessive_codebook_indexes
             "num_codebooks": 16,  # used to construct distillation loss
         }
     )
@@ -988,6 +988,7 @@ def run(rank, world_size, args):
             optimizer=optimizer,
             sp=sp,
             params=params,
+            warmup=0.0 if params.start_epoch == 1 else 1.0,
         )
 
     scaler = GradScaler(enabled=params.use_fp16)
@@ -1048,6 +1049,7 @@ def scan_pessimistic_batches_for_oom(
     optimizer: torch.optim.Optimizer,
     sp: spm.SentencePieceProcessor,
     params: AttributeDict,
+    warmup: float,
 ):
     from lhotse.dataset import find_pessimistic_batches
 
@@ -1058,9 +1060,6 @@ def scan_pessimistic_batches_for_oom(
     for criterion, cuts in batches.items():
         batch = train_dl.dataset[cuts]
         try:
-            # warmup = 0.0 is so that the derivs for the pruned loss stay zero
-            # (i.e. are not remembered by the decaying-average in adam), because
-            # we want to avoid these params being subject to shrinkage in adam.
             with torch.cuda.amp.autocast(enabled=params.use_fp16):
                 loss, _ = compute_loss(
                     params=params,
@@ -1068,7 +1067,7 @@ def scan_pessimistic_batches_for_oom(
                     sp=sp,
                     batch=batch,
                     is_training=True,
-                    warmup=0.0,
+                    warmup=warmup,
                 )
             loss.backward()
             optimizer.step()

From 998091ef52ceb46e4efe57dfc28ebce3c4edc10d Mon Sep 17 00:00:00 2001
From: Mingshuang Luo <37799481+luomingshuang@users.noreply.github.com>
Date: Mon, 20 Jun 2022 14:57:08 +0800
Subject: [PATCH 2/9] do some changes for export.py (#437)

---
 .../ASR/pruned_transducer_stateless2/export.py           | 7 +++++--
 egs/aishell/ASR/transducer_stateless/export.py           | 7 +++++--
 .../ASR/transducer_stateless_modified-2/export.py        | 7 +++++--
 egs/aishell/ASR/transducer_stateless_modified/export.py  | 7 +++++--
 egs/aishell4/ASR/pruned_transducer_stateless5/export.py  | 7 +++++--
 .../ASR/pruned_transducer_stateless2/export.py           | 7 +++++--
 .../ASR/pruned_transducer_stateless2/export.py           | 7 +++++--
 .../ASR/pruned_transducer_stateless2/export.py           | 7 +++++--
 egs/tedlium3/ASR/pruned_transducer_stateless/export.py   | 7 +++++--
 egs/tedlium3/ASR/transducer_stateless/export.py          | 9 ++++++---
 .../ASR/pruned_transducer_stateless2/export.py           | 7 +++++--
 11 files changed, 56 insertions(+), 23 deletions(-)

diff --git a/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/export.py b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/export.py
index 43033e517..00b54c39f 100644
--- a/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/export.py
+++ b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/export.py
@@ -114,8 +114,6 @@ def main():
     args = get_parser().parse_args()
     args.exp_dir = Path(args.exp_dir)
 
-    assert args.jit is False, "Support torchscript will be added later"
-
     params = get_params()
     params.update(vars(args))
 
@@ -155,6 +153,11 @@ def main():
     model.eval()
 
     if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
         logging.info("Using torch.jit.script")
         model = torch.jit.script(model)
         filename = params.exp_dir / "cpu_jit.pt"
diff --git a/egs/aishell/ASR/transducer_stateless/export.py b/egs/aishell/ASR/transducer_stateless/export.py
index 591b333e0..4c6519b96 100755
--- a/egs/aishell/ASR/transducer_stateless/export.py
+++ b/egs/aishell/ASR/transducer_stateless/export.py
@@ -184,8 +184,6 @@ def main():
     args = get_parser().parse_args()
     args.exp_dir = Path(args.exp_dir)
 
-    assert args.jit is False, "Support torchscript will be added later"
-
     params = get_params()
     params.update(vars(args))
 
@@ -225,6 +223,11 @@ def main():
     model.eval()
 
     if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
         logging.info("Using torch.jit.script")
         model = torch.jit.script(model)
         filename = params.exp_dir / "cpu_jit.pt"
diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/export.py b/egs/aishell/ASR/transducer_stateless_modified-2/export.py
index d009de603..3bd2ceb11 100755
--- a/egs/aishell/ASR/transducer_stateless_modified-2/export.py
+++ b/egs/aishell/ASR/transducer_stateless_modified-2/export.py
@@ -182,8 +182,6 @@ def get_transducer_model(params: AttributeDict) -> nn.Module:
 def main():
     args = get_parser().parse_args()
 
-    assert args.jit is False, "torchscript support will be added later"
-
     params = get_params()
     params.update(vars(args))
 
@@ -223,6 +221,11 @@ def main():
     model.eval()
 
     if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
         logging.info("Using torch.jit.script")
         model = torch.jit.script(model)
         filename = params.exp_dir / "cpu_jit.pt"
diff --git a/egs/aishell/ASR/transducer_stateless_modified/export.py b/egs/aishell/ASR/transducer_stateless_modified/export.py
index 9a20fab6f..11335a834 100755
--- a/egs/aishell/ASR/transducer_stateless_modified/export.py
+++ b/egs/aishell/ASR/transducer_stateless_modified/export.py
@@ -182,8 +182,6 @@ def get_transducer_model(params: AttributeDict) -> nn.Module:
 def main():
     args = get_parser().parse_args()
 
-    assert args.jit is False, "torchscript support will be added later"
-
     params = get_params()
     params.update(vars(args))
 
@@ -223,6 +221,11 @@ def main():
     model.eval()
 
     if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
         logging.info("Using torch.jit.script")
         model = torch.jit.script(model)
         filename = params.exp_dir / "cpu_jit.pt"
diff --git a/egs/aishell4/ASR/pruned_transducer_stateless5/export.py b/egs/aishell4/ASR/pruned_transducer_stateless5/export.py
index f487a8ba5..f42a85373 100755
--- a/egs/aishell4/ASR/pruned_transducer_stateless5/export.py
+++ b/egs/aishell4/ASR/pruned_transducer_stateless5/export.py
@@ -149,8 +149,6 @@ def main():
     args = get_parser().parse_args()
     args.exp_dir = Path(args.exp_dir)
 
-    assert args.jit is False, "Support torchscript will be added later"
-
     params = get_params()
     params.update(vars(args))
 
@@ -252,6 +250,11 @@ def main():
     model.eval()
 
     if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
         logging.info("Using torch.jit.script")
         model = torch.jit.script(model)
         filename = params.exp_dir / "cpu_jit.pt"
diff --git a/egs/alimeeting/ASR/pruned_transducer_stateless2/export.py b/egs/alimeeting/ASR/pruned_transducer_stateless2/export.py
index 0a69e0a57..8beec1b8a 100644
--- a/egs/alimeeting/ASR/pruned_transducer_stateless2/export.py
+++ b/egs/alimeeting/ASR/pruned_transducer_stateless2/export.py
@@ -114,8 +114,6 @@ def main():
     args = get_parser().parse_args()
     args.exp_dir = Path(args.exp_dir)
 
-    assert args.jit is False, "Support torchscript will be added later"
-
     params = get_params()
     params.update(vars(args))
 
@@ -155,6 +153,11 @@ def main():
     model.eval()
 
     if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
         logging.info("Using torch.jit.script")
         model = torch.jit.script(model)
         filename = params.exp_dir / "cpu_jit.pt"
diff --git a/egs/gigaspeech/ASR/pruned_transducer_stateless2/export.py b/egs/gigaspeech/ASR/pruned_transducer_stateless2/export.py
index 6b3a7a9ff..cff9c7377 100755
--- a/egs/gigaspeech/ASR/pruned_transducer_stateless2/export.py
+++ b/egs/gigaspeech/ASR/pruned_transducer_stateless2/export.py
@@ -131,8 +131,6 @@ def main():
     args = get_parser().parse_args()
     args.exp_dir = Path(args.exp_dir)
 
-    assert args.jit is False, "Support torchscript will be added later"
-
     params = get_params()
     params.update(vars(args))
 
@@ -191,6 +189,11 @@ def main():
     model.eval()
 
     if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
         logging.info("Using torch.jit.script")
         model = torch.jit.script(model)
         filename = params.exp_dir / "cpu_jit.pt"
diff --git a/egs/spgispeech/ASR/pruned_transducer_stateless2/export.py b/egs/spgispeech/ASR/pruned_transducer_stateless2/export.py
index 6119ecf2c..77faa3c0e 100755
--- a/egs/spgispeech/ASR/pruned_transducer_stateless2/export.py
+++ b/egs/spgispeech/ASR/pruned_transducer_stateless2/export.py
@@ -130,8 +130,6 @@ def main():
     args = get_parser().parse_args()
     args.exp_dir = Path(args.exp_dir)
 
-    assert args.jit is False, "Support torchscript will be added later"
-
     params = get_params()
     params.update(vars(args))
 
@@ -178,6 +176,11 @@ def main():
     model.eval()
 
     if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
         logging.info("Using torch.jit.script")
         model = torch.jit.script(model)
         filename = params.exp_dir / "cpu_jit.pt"
diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless/export.py b/egs/tedlium3/ASR/pruned_transducer_stateless/export.py
index 1e6edbb99..a1c3bcea3 100644
--- a/egs/tedlium3/ASR/pruned_transducer_stateless/export.py
+++ b/egs/tedlium3/ASR/pruned_transducer_stateless/export.py
@@ -117,8 +117,6 @@ def main():
     args = get_parser().parse_args()
     args.exp_dir = Path(args.exp_dir)
 
-    assert args.jit is False, "Support torchscript will be added later"
-
     params = get_params()
     params.update(vars(args))
 
@@ -161,6 +159,11 @@ def main():
     model.eval()
 
     if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
         logging.info("Using torch.jit.script")
         model = torch.jit.script(model)
         filename = params.exp_dir / "cpu_jit.pt"
diff --git a/egs/tedlium3/ASR/transducer_stateless/export.py b/egs/tedlium3/ASR/transducer_stateless/export.py
index f2bfa2ec9..c32b1d002 100644
--- a/egs/tedlium3/ASR/transducer_stateless/export.py
+++ b/egs/tedlium3/ASR/transducer_stateless/export.py
@@ -1,7 +1,7 @@
 #!/usr/bin/env python3
 #
 # Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang
-# 					     Mingshuang Luo)
+# 					                         Mingshuang Luo)
 #
 # See ../../../../LICENSE for clarification regarding multiple authors
 #
@@ -185,8 +185,6 @@ def main():
     args = get_parser().parse_args()
     args.exp_dir = Path(args.exp_dir)
 
-    assert args.jit is False, "Support torchscript will be added later"
-
     params = get_params()
     params.update(vars(args))
 
@@ -229,6 +227,11 @@ def main():
     model.eval()
 
     if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
         logging.info("Using torch.jit.script")
         model = torch.jit.script(model)
         filename = params.exp_dir / "cpu_jit.pt"
diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/export.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/export.py
index 8c4f92c81..345792a3c 100644
--- a/egs/wenetspeech/ASR/pruned_transducer_stateless2/export.py
+++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/export.py
@@ -114,8 +114,6 @@ def main():
     args = get_parser().parse_args()
     args.exp_dir = Path(args.exp_dir)
 
-    assert args.jit is False, "Support torchscript will be added later"
-
     params = get_params()
     params.update(vars(args))
 
@@ -155,6 +153,11 @@ def main():
     model.eval()
 
     if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
         logging.info("Using torch.jit.script")
         model = torch.jit.script(model)
         filename = params.exp_dir / "cpu_jit.pt"

From 91b2765cfd39fea1783cc07bb391e27bc2d5224e Mon Sep 17 00:00:00 2001
From: 2xwwx2 <wwx007121@sina.com>
Date: Mon, 20 Jun 2022 16:41:04 +0800
Subject: [PATCH 3/9] Fixs spelling mistake (#438)

---
 egs/aishell/ASR/prepare.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/egs/aishell/ASR/prepare.sh b/egs/aishell/ASR/prepare.sh
index 26324b0af..da0a1470e 100755
--- a/egs/aishell/ASR/prepare.sh
+++ b/egs/aishell/ASR/prepare.sh
@@ -18,7 +18,7 @@ stop_stage=10
 #      This directory contains the language model downloaded from
 #        https://huggingface.co/pkufool/aishell_lm
 #
-#        - 3-gram.unpruned.apra
+#        - 3-gram.unpruned.arpa
 #
 #  - $dl_dir/musan
 #      This directory contains the following directories downloaded from

From d3daeaf5cd6712fca63b98db960ee71ff211a59e Mon Sep 17 00:00:00 2001
From: Zengwei Yao <yaozengwei@outlook.com>
Date: Tue, 21 Jun 2022 19:16:59 +0800
Subject: [PATCH 4/9] Upload extracted codebook indexes (#429)

* save only vq-related info to manifest

* support to join manifest files

* support using extracted codebook indexes

* fix doc

* minor fix

* add enable-distillation argument option, fix monir typos

* fix style

* fix typo
---
 .../ASR/distillation_with_hubert.sh           | 119 +++++++++++++-----
 .../pruned_transducer_stateless6/decode.py    |   9 +-
 .../extract_codebook_index.py                 |  19 ++-
 .../ASR/pruned_transducer_stateless6/train.py |  22 ++--
 .../pruned_transducer_stateless6/vq_utils.py  |  80 +++++++++---
 5 files changed, 189 insertions(+), 60 deletions(-)
 mode change 100644 => 100755 egs/librispeech/ASR/distillation_with_hubert.sh

diff --git a/egs/librispeech/ASR/distillation_with_hubert.sh b/egs/librispeech/ASR/distillation_with_hubert.sh
old mode 100644
new mode 100755
index e18ba8f55..3d4c4856a
--- a/egs/librispeech/ASR/distillation_with_hubert.sh
+++ b/egs/librispeech/ASR/distillation_with_hubert.sh
@@ -1,3 +1,5 @@
+#!/usr/bin/env bash
+#
 # A short introduction about distillation framework.
 #
 # A typical traditional distillation method is
@@ -14,15 +16,15 @@
 #    teacher embeddings.
 # 3. a middle layer 6(1-based) out of total 6 layers is used to extract
 #    student embeddings.
-
-# This is an example to do distillation with librispeech clean-100 subset.
-# run with command:
-# bash distillation_with_hubert.sh [0|1|2|3|4]
 #
-# For example command
-# bash distillation_with_hubert.sh 0
-# will download hubert model.
-stage=$1
+# To directly download the extracted codebook indexes for model distillation, you can
+# set stage=2, stop_stage=4, use_extracted_codebook=True
+#
+# To start from scratch, you can
+# set stage=0, stop_stage=4, use_extracted_codebook=False
+
+stage=0
+stop_stage=4
 
 # Set the GPUs available.
 # This script requires at least one GPU.
@@ -33,10 +35,35 @@ stage=$1
 # export CUDA_VISIBLE_DEVICES="0"
 #
 # Suppose GPU 2,3,4,5 are available.
-export CUDA_VISIBLE_DEVICES="2,3,4,5"
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
 
+exp_dir=./pruned_transducer_stateless6/exp
+mkdir -p $exp_dir
 
-if [ $stage -eq 0 ]; then
+# full_libri can be "True" or "False"
+#   "True" -> use full librispeech dataset for distillation
+#   "False" -> use train-clean-100 subset for distillation
+full_libri=False
+
+# use_extracted_codebook can be "True" or "False"
+#   "True" -> stage 0 and stage 1 would be skipped,
+#     and directly download the extracted codebook indexes for distillation
+#   "False" -> start from scratch
+use_extracted_codebook=False
+
+# teacher_model_id can be one of
+#   "hubert_xtralarge_ll60k_finetune_ls960" -> fine-tuned model, it is the one we currently use.
+#   "hubert_xtralarge_ll60k" -> pretrained model without fintuing
+teacher_model_id=hubert_xtralarge_ll60k_finetune_ls960
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ] && [ ! "$use_extracted_codebook" == "True" ]; then
+  log "Stage 0: Download HuBERT model"
   # Preparation stage.
 
   # Install fairseq according to:
@@ -45,7 +72,7 @@ if [ $stage -eq 0 ]; then
   # commit 806855bf660ea748ed7ffb42fe8dcc881ca3aca0 is used.
   has_fairseq=$(python3 -c "import importlib; print(importlib.util.find_spec('fairseq') is not None)")
   if [ $has_fairseq == 'False' ]; then
-    echo "Please install fairseq before running following stages"
+    log "Please install fairseq before running following stages"
     exit 1
   fi
 
@@ -56,42 +83,41 @@ if [ $stage -eq 0 ]; then
 
   has_quantization=$(python3 -c "import importlib; print(importlib.util.find_spec('quantization') is not None)")
   if [ $has_quantization == 'False' ]; then
-    echo "Please install quantization before running following stages"
+    log "Please install quantization before running following stages"
     exit 1
   fi
 
-  echo "Download hubert model."
+  log "Download HuBERT model."
   # Parameters about model.
-  exp_dir=./pruned_transducer_stateless6/exp/
-  model_id=hubert_xtralarge_ll60k_finetune_ls960
   hubert_model_dir=${exp_dir}/hubert_models
-  hubert_model=${hubert_model_dir}/${model_id}.pt
+  hubert_model=${hubert_model_dir}/${teacher_model_id}.pt
   mkdir -p ${hubert_model_dir}
   # For more models refer to: https://github.com/pytorch/fairseq/tree/main/examples/hubert
   if [ -f ${hubert_model} ]; then
-    echo "hubert model alread exists."
+    log "HuBERT model alread exists."
   else
-    wget -c https://dl.fbaipublicfiles.com/hubert/${model_id} -P ${hubert_model}
+    wget -c https://dl.fbaipublicfiles.com/hubert/${teacher_model_id}.pt -P ${hubert_model_dir}
     wget -c wget https://dl.fbaipublicfiles.com/fairseq/wav2vec/dict.ltr.txt -P ${hubert_model_dir}
   fi
 fi
 
 if [ ! -d ./data/fbank ]; then
-  echo "This script assumes ./data/fbank is already generated by prepare.sh"
+  log "This script assumes ./data/fbank is already generated by prepare.sh"
   exit 1
 fi
 
-if [ $stage -eq 1 ]; then
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ] && [ ! "$use_extracted_codebook" == "True" ]; then
+  log "Stage 1: Verify that the downloaded HuBERT model is correct."
   # This stage is not directly used by codebook indexes extraction.
   # It is a method to "prove" that the downloaed hubert model
   # is inferenced in an correct way if WERs look like normal.
   # Expect WERs:
   # [test-clean-ctc_greedy_search] %WER 2.04% [1075 / 52576, 92 ins, 104 del, 879 sub ]
   # [test-other-ctc_greedy_search] %WER 3.71% [1942 / 52343, 152 ins, 126 del, 1664 sub ]
-  ./pruned_transducer_stateless6/hubert_decode.py
+  ./pruned_transducer_stateless6/hubert_decode.py --exp-dir $exp_dir
 fi
 
-if [ $stage -eq 2 ]; then
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
   # Analysis of disk usage:
   # With num_codebooks==8, each teacher embedding is quantized into
   # a sequence of eight 8-bit integers, i.e. only eight bytes are needed.
@@ -113,25 +139,61 @@ if [ $stage -eq 2 ]; then
   # During quantizer's training data(teacher embedding) and it's training,
   # only the first ONE GPU is used.
   # During codebook indexes extraction, ALL GPUs set by CUDA_VISIBLE_DEVICES are used.
+
+  if [ "$use_extracted_codebook" == "True" ]; then
+    if [ ! "$teacher_model_id" == "hubert_xtralarge_ll60k_finetune_ls960" ]; then
+      log "Currently we only uploaded codebook indexes from teacher model hubert_xtralarge_ll60k_finetune_ls960"
+      exit 1
+    fi
+    mkdir -p $exp_dir/vq
+    codebook_dir=$exp_dir/vq/$teacher_model_id
+    mkdir -p codebook_dir
+    codebook_download_dir=$exp_dir/download_codebook
+    if [ -d $codebook_download_dir ]; then
+      log "$codebook_download_dir exists, you should remove it first."
+      exit 1
+    fi
+    log "Downloading extracted codebook indexes to $codebook_download_dir"
+    # Make sure you have git-lfs installed (https://git-lfs.github.com)
+    git lfs install
+    git clone https://huggingface.co/Zengwei/pruned_transducer_stateless6_hubert_xtralarge_ll60k_finetune_ls960 $codebook_download_dir
+
+    mkdir -p data/vq_fbank
+    mv $codebook_download_dir/*.jsonl.gz data/vq_fbank/
+    mkdir -p $codebook_dir/splits4
+    mv $codebook_download_dir/*.h5 $codebook_dir/splits4/
+    log "Remove $codebook_download_dir"
+    rm -rf $codebook_download_dir
+  fi
+
   ./pruned_transducer_stateless6/extract_codebook_index.py \
-    --full-libri False
+    --full-libri $full_libri \
+    --exp-dir $exp_dir \
+    --embedding-layer 36 \
+    --num-utts 1000 \
+    --num-codebooks 8 \
+    --max-duration 100 \
+    --teacher-model-id $teacher_model_id \
+    --use-extracted-codebook $use_extracted_codebook
 fi
 
-if [ $stage -eq 3 ]; then
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
   # Example training script.
   # Note: it's better to set spec-aug-time-warpi-factor=-1
   WORLD_SIZE=$(echo ${CUDA_VISIBLE_DEVICES} | awk '{n=split($1, _, ","); print n}')
   ./pruned_transducer_stateless6/train.py \
     --manifest-dir ./data/vq_fbank \
     --master-port 12359 \
-    --full-libri False \
+    --full-libri $full_libri \
     --spec-aug-time-warp-factor -1 \
     --max-duration 300 \
     --world-size ${WORLD_SIZE} \
-    --num-epochs 20
+    --num-epochs 20 \
+    --exp-dir $exp_dir \
+    --enable-distillation True
 fi
 
-if [ $stage -eq 4 ]; then
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
   # Results should be similar to:
   # errs-test-clean-beam_size_4-epoch-20-avg-10-beam-4.txt:%WER = 5.67
   # errs-test-other-beam_size_4-epoch-20-avg-10-beam-4.txt:%WER = 15.60
@@ -140,5 +202,6 @@ if [ $stage -eq 4 ]; then
     --epoch 20 \
     --avg 10 \
     --max-duration 200 \
-    --exp-dir ./pruned_transducer_stateless6/exp
+    --exp-dir $exp_dir \
+    --enable-distillation True
 fi
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless6/decode.py
index 4739a6526..701cad73c 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless6/decode.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless6/decode.py
@@ -128,7 +128,7 @@ def get_parser():
     parser.add_argument(
         "--use-averaged-model",
         type=str2bool,
-        default=False,
+        default=True,
         help="Whether to load averaged model. Currently it only supports "
         "using --epoch. If True, it would decode with the averaged model "
         "over the epoch range from `epoch-avg` (excluded) to `epoch`."
@@ -143,6 +143,13 @@ def get_parser():
         help="The experiment dir",
     )
 
+    parser.add_argument(
+        "--enable-distillation",
+        type=str2bool,
+        default=True,
+        help="Whether to eanble distillation.",
+    )
+
     parser.add_argument(
         "--bpe-model",
         type=str,
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/extract_codebook_index.py b/egs/librispeech/ASR/pruned_transducer_stateless6/extract_codebook_index.py
index c5c172ff2..21409287c 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless6/extract_codebook_index.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless6/extract_codebook_index.py
@@ -24,7 +24,7 @@ import torch
 from vq_utils import CodebookIndexExtractor
 from asr_datamodule import LibriSpeechAsrDataModule
 from hubert_xlarge import HubertXlargeFineTuned
-from icefall.utils import AttributeDict
+from icefall.utils import AttributeDict, str2bool
 
 
 def get_parser():
@@ -38,6 +38,13 @@ def get_parser():
         help="The experiment dir",
     )
 
+    parser.add_argument(
+        "--use-extracted-codebook",
+        type=str2bool,
+        default=False,
+        help="Whether to use the extracted codebook indexes.",
+    )
+
     return parser
 
 
@@ -71,9 +78,13 @@ def main():
     params.world_size = world_size
 
     extractor = CodebookIndexExtractor(params=params)
-    extractor.extract_and_save_embedding()
-    extractor.train_quantizer()
-    extractor.extract_codebook_indexes()
+    if not params.use_extracted_codebook:
+        extractor.extract_and_save_embedding()
+        extractor.train_quantizer()
+        extractor.extract_codebook_indexes()
+
+    extractor.reuse_manifests()
+    extractor.join_manifests()
 
 
 if __name__ == "__main__":
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/train.py b/egs/librispeech/ASR/pruned_transducer_stateless6/train.py
index 9e9fc1440..b904e1e59 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless6/train.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless6/train.py
@@ -41,7 +41,7 @@ export CUDA_VISIBLE_DEVICES="0,1,2,3"
   --full-libri 1 \
   --max-duration 550
 
-# For distiallation with codebook_indexes:
+# For distillation with codebook_indexes:
 
 ./pruned_transducer_stateless6/train.py \
   --manifest-dir ./data/vq_fbank \
@@ -300,6 +300,13 @@ def get_parser():
         help="Whether to use half precision training.",
     )
 
+    parser.add_argument(
+        "--enable-distillation",
+        type=str2bool,
+        default=True,
+        help="Whether to eanble distillation.",
+    )
+
     return parser
 
 
@@ -372,7 +379,6 @@ def get_params() -> AttributeDict:
             "model_warm_step": 3000,  # arg given to model, not for lrate
             "env_info": get_env_info(),
             # parameters for distillation with codebook indexes.
-            "enable_distiallation": True,
             "distillation_layer": 5,  # 0-based index
             # Since output rate of hubert is 50, while that of encoder is 8,
             # two successive codebook_index are concatenated together.
@@ -394,7 +400,7 @@ def get_encoder_model(params: AttributeDict) -> nn.Module:
         dim_feedforward=params.dim_feedforward,
         num_encoder_layers=params.num_encoder_layers,
         middle_output_layer=params.distillation_layer
-        if params.enable_distiallation
+        if params.enable_distillation
         else None,
     )
     return encoder
@@ -433,9 +439,7 @@ def get_transducer_model(params: AttributeDict) -> nn.Module:
         decoder_dim=params.decoder_dim,
         joiner_dim=params.joiner_dim,
         vocab_size=params.vocab_size,
-        num_codebooks=params.num_codebooks
-        if params.enable_distiallation
-        else 0,
+        num_codebooks=params.num_codebooks if params.enable_distillation else 0,
     )
     return model
 
@@ -615,7 +619,7 @@ def compute_loss(
     y = k2.RaggedTensor(y).to(device)
 
     info = MetricsTracker()
-    if is_training and params.enable_distiallation:
+    if is_training and params.enable_distillation:
         codebook_indexes, _ = extract_codebook_indexes(batch)
         codebook_indexes = codebook_indexes.to(device)
     else:
@@ -645,7 +649,7 @@ def compute_loss(
             params.simple_loss_scale * simple_loss
             + pruned_loss_scale * pruned_loss
         )
-        if is_training and params.enable_distiallation:
+        if is_training and params.enable_distillation:
             assert codebook_loss is not None
             loss += params.codebook_loss_scale * codebook_loss
 
@@ -661,7 +665,7 @@ def compute_loss(
     info["loss"] = loss.detach().cpu().item()
     info["simple_loss"] = simple_loss.detach().cpu().item()
     info["pruned_loss"] = pruned_loss.detach().cpu().item()
-    if is_training and params.enable_distiallation:
+    if is_training and params.enable_distillation:
         info["codebook_loss"] = codebook_loss.detach().cpu().item()
 
     return loss, info
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/vq_utils.py b/egs/librispeech/ASR/pruned_transducer_stateless6/vq_utils.py
index c4935f921..e3dcd039b 100644
--- a/egs/librispeech/ASR/pruned_transducer_stateless6/vq_utils.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless6/vq_utils.py
@@ -37,6 +37,7 @@ from icefall.utils import (
     setup_logger,
 )
 from lhotse import CutSet, load_manifest
+from lhotse.cut import MonoCut
 from lhotse.features.io import NumpyHdf5Writer
 
 
@@ -62,16 +63,15 @@ class CodebookIndexExtractor:
         setup_logger(f"{self.vq_dir}/log-vq_extraction")
 
     def init_dirs(self):
-        # vq_dir is the root dir for quantizer:
-        #    training data/ quantizer / extracted codebook indexes
+        # vq_dir is the root dir for quantization, containing:
+        # training data, trained quantizer, and extracted codebook indexes
         self.vq_dir = (
             self.params.exp_dir / f"vq/{self.params.teacher_model_id}/"
         )
         self.vq_dir.mkdir(parents=True, exist_ok=True)
 
-        # manifest_dir for :
-        # splited original manifests,
-        # extracted codebook indexes and their related manifests
+        # manifest_dir contains:
+        # splited original manifests, extracted codebook indexes with related manifests # noqa
         self.manifest_dir = self.vq_dir / f"splits{self.params.world_size}"
         self.manifest_dir.mkdir(parents=True, exist_ok=True)
 
@@ -135,6 +135,7 @@ class CodebookIndexExtractor:
             logging.warn(warn_message)
             return
 
+        logging.info("Start to extract embeddings for training the quantizer.")
         total_cuts = 0
         with NumpyHdf5Writer(self.embedding_file_path) as writer:
             for batch_idx, batch in enumerate(self.quantizer_train_dl):
@@ -187,14 +188,15 @@ class CodebookIndexExtractor:
             return
 
         assert self.embedding_file_path.exists()
+        logging.info("Start to train quantizer.")
         trainer = quantization.QuantizerTrainer(
             dim=self.params.embedding_dim,
             bytes_per_frame=self.params.num_codebooks,
             device=self.params.device,
         )
         train, valid = quantization.read_hdf5_data(self.embedding_file_path)
-        B = 512  # Minibatch size, this is very arbitrary, it's close to what we used
-        # when we tuned this method.
+        B = 512  # Minibatch size, this is very arbitrary,
+        # it's close to what we used when we tuned this method.
 
         def minibatch_generator(data: torch.Tensor, repeat: bool):
             assert 3 * B < data.shape[0]
@@ -222,18 +224,50 @@ class CodebookIndexExtractor:
         """
         for subset in self.params.subsets:
             logging.info(f"About to split {subset}.")
-            ori_manifest = f"./data/fbank/cuts_train-{subset}.json.gz"
+            ori_manifest = (
+                f"./data/fbank/librispeech_cuts_train-{subset}.jsonl.gz"
+            )
             split_cmd = f"lhotse split {self.params.world_size} {ori_manifest} {self.manifest_dir}"
             os.system(f"{split_cmd}")
 
+    def join_manifests(self):
+        """
+        Join the vq manifest to the original manifest according to cut id.
+        """
+        logging.info("Start to join manifest files.")
+        for subset in self.params.subsets:
+            vq_manifest_path = (
+                self.dst_manifest_dir
+                / f"librispeech_cuts_train-{subset}-vq.jsonl.gz"
+            )
+            ori_manifest_path = (
+                self.ori_manifest_dir
+                / f"librispeech_cuts_train-{subset}.jsonl.gz"
+            )
+            dst_vq_manifest_path = (
+                self.dst_manifest_dir
+                / f"librispeech_cuts_train-{subset}.jsonl.gz"
+            )
+            cuts_vq = load_manifest(vq_manifest_path)
+            cuts_ori = load_manifest(ori_manifest_path)
+            cuts_vq = cuts_vq.sort_like(cuts_ori)
+            for cut_idx, (cut_vq, cut_ori) in enumerate(zip(cuts_vq, cuts_ori)):
+                assert cut_vq.id == cut_ori.id
+                cut_ori.codebook_indexes = cut_vq.codebook_indexes
+
+            CutSet.from_cuts(cuts_ori).to_jsonl(dst_vq_manifest_path)
+            logging.info(f"Processed {subset}.")
+            logging.info(f"Saved to {dst_vq_manifest_path}.")
+
     def merge_vq_manifests(self):
         """
         Merge generated vq included manfiests and storage to self.dst_manifest_dir.
         """
         for subset in self.params.subsets:
-            vq_manifests = f"{self.manifest_dir}/with_codebook_indexes-cuts_train-{subset}*.json.gz"
+            vq_manifests = f"{self.manifest_dir}/with_codebook_indexes-librispeech-cuts_train-{subset}*.jsonl.gz"
             dst_vq_manifest = (
-                self.dst_manifest_dir / f"cuts_train-{subset}.json.gz"
+                self.dst_manifest_dir
+                / f"librispeech_cuts_train-{subset}-vq.jsonl.gz"
             )
             if 1 == self.params.world_size:
                 merge_cmd = f"cp {vq_manifests} {dst_vq_manifest}"
@@ -273,7 +307,6 @@ class CodebookIndexExtractor:
                 os.symlink(ori_manifest_path, dst_manifest_path)
 
     def create_vq_fbank(self):
-        self.reuse_manifests()
         self.merge_vq_manifests()
 
     @cached_property
@@ -294,11 +327,13 @@ class CodebookIndexExtractor:
 
     def load_ori_dl(self, subset):
         if self.params.world_size == 1:
-            ori_manifest_path = f"./data/fbank/cuts_train-{subset}.json.gz"
+            ori_manifest_path = (
+                f"./data/fbank/librispeech_cuts_train-{subset}.jsonl.gz"
+            )
         else:
             ori_manifest_path = (
                 self.manifest_dir
-                / f"cuts_train-{subset}.{self.params.manifest_index}.json.gz"
+                / f"librispeech_cuts_train-{subset}.{self.params.manifest_index}.jsonl.gz"  # noqa
             )
 
         cuts = load_manifest(ori_manifest_path)
@@ -311,6 +346,7 @@ class CodebookIndexExtractor:
         torch.cuda.empty_cache()
 
     def extract_codebook_indexes(self):
+        logging.info("Start to extract codebook indexes.")
         if self.params.world_size == 1:
             self.extract_codebook_indexes_imp()
         else:
@@ -333,7 +369,7 @@ class CodebookIndexExtractor:
     def extract_codebook_indexes_imp(self):
         for subset in self.params.subsets:
             num_cuts = 0
-            cuts = []
+            new_cuts = []
             if self.params.world_size == 1:
                 manifest_file_id = f"{subset}"
             else:
@@ -356,15 +392,23 @@ class CodebookIndexExtractor:
                     assert len(cut_list) == codebook_indexes.shape[0]
                     assert all(c.start == 0 for c in supervisions["cut"])
 
+                    new_cut_list = []
                     for idx, cut in enumerate(cut_list):
-                        cut.codebook_indexes = writer.store_array(
+                        new_cut = MonoCut(
+                            id=cut.id,
+                            start=cut.start,
+                            duration=cut.duration,
+                            channel=cut.channel,
+                        )
+                        new_cut.codebook_indexes = writer.store_array(
                             key=cut.id,
                             value=codebook_indexes[idx][: num_frames[idx]],
                             frame_shift=0.02,
                             temporal_dim=0,
                             start=0,
                         )
-                    cuts += cut_list
+                        new_cut_list.append(new_cut)
+                    new_cuts += new_cut_list
                     num_cuts += len(cut_list)
                     message = f"Processed {num_cuts} cuts from {subset}"
                     if self.params.world_size > 1:
@@ -373,9 +417,9 @@ class CodebookIndexExtractor:
 
                 json_file_path = (
                     self.manifest_dir
-                    / f"with_codebook_indexes-cuts_train-{manifest_file_id}.json.gz"
+                    / f"with_codebook_indexes-librispeech-cuts_train-{manifest_file_id}.jsonl.gz"  # noqa
                 )
-                CutSet.from_cuts(cuts).to_json(json_file_path)
+                CutSet.from_cuts(new_cuts).to_jsonl(json_file_path)
 
 
 @torch.no_grad()

From 7100c33820c8c478e07d3435e25e4f1543b6eec7 Mon Sep 17 00:00:00 2001
From: Fangjun Kuang <csukuangfj@gmail.com>
Date: Tue, 21 Jun 2022 21:17:22 +0800
Subject: [PATCH 5/9] Add pruned RNN-T for aishell. (#436)

* Add pruned RNN-T for aishell.

* support torch script.

* Update CI.

* Minor fixes.

* Add links to sherpa.
---
 ...pruned-transducer-stateless3-2022-06-20.sh |   86 ++
 .github/workflows/run-aishell-2022-06-20.yml  |  119 ++
 egs/aishell/ASR/README.md                     |    3 +
 egs/aishell/ASR/RESULTS.md                    |   85 +-
 .../aidatatang_200zh.py                       |    1 +
 .../pruned_transducer_stateless3/aishell.py   |    1 +
 .../asr_datamodule.py                         |    1 +
 .../beam_search.py                            |    1 +
 .../pruned_transducer_stateless3/conformer.py |    1 +
 .../pruned_transducer_stateless3/decode.py    |  637 +++++++++
 .../pruned_transducer_stateless3/decoder.py   |    1 +
 .../encoder_interface.py                      |    1 +
 .../exp-context-size-1                        |    1 +
 .../pruned_transducer_stateless3/export.py    |  277 ++++
 .../pruned_transducer_stateless3/joiner.py    |    1 +
 .../ASR/pruned_transducer_stateless3/model.py |  236 ++++
 .../ASR/pruned_transducer_stateless3/optim.py |    1 +
 .../pretrained.py                             |  337 +++++
 .../pruned_transducer_stateless3/scaling.py   |    1 +
 .../ASR/pruned_transducer_stateless3/train.py | 1229 +++++++++++++++++
 .../transducer_stateless_modified-2/train.py  |    2 +-
 .../pruned_transducer_stateless2/decoder.py   |    3 +
 .../pruned_transducer_stateless5/conformer.py |   39 +-
 .../pruned_transducer_stateless5/export.py    |    9 +-
 24 files changed, 3055 insertions(+), 18 deletions(-)
 create mode 100755 .github/scripts/run-aishell-pruned-transducer-stateless3-2022-06-20.sh
 create mode 100644 .github/workflows/run-aishell-2022-06-20.yml
 create mode 120000 egs/aishell/ASR/pruned_transducer_stateless3/aidatatang_200zh.py
 create mode 120000 egs/aishell/ASR/pruned_transducer_stateless3/aishell.py
 create mode 120000 egs/aishell/ASR/pruned_transducer_stateless3/asr_datamodule.py
 create mode 120000 egs/aishell/ASR/pruned_transducer_stateless3/beam_search.py
 create mode 120000 egs/aishell/ASR/pruned_transducer_stateless3/conformer.py
 create mode 100755 egs/aishell/ASR/pruned_transducer_stateless3/decode.py
 create mode 120000 egs/aishell/ASR/pruned_transducer_stateless3/decoder.py
 create mode 120000 egs/aishell/ASR/pruned_transducer_stateless3/encoder_interface.py
 create mode 120000 egs/aishell/ASR/pruned_transducer_stateless3/exp-context-size-1
 create mode 100755 egs/aishell/ASR/pruned_transducer_stateless3/export.py
 create mode 120000 egs/aishell/ASR/pruned_transducer_stateless3/joiner.py
 create mode 100644 egs/aishell/ASR/pruned_transducer_stateless3/model.py
 create mode 120000 egs/aishell/ASR/pruned_transducer_stateless3/optim.py
 create mode 100755 egs/aishell/ASR/pruned_transducer_stateless3/pretrained.py
 create mode 120000 egs/aishell/ASR/pruned_transducer_stateless3/scaling.py
 create mode 100755 egs/aishell/ASR/pruned_transducer_stateless3/train.py

diff --git a/.github/scripts/run-aishell-pruned-transducer-stateless3-2022-06-20.sh b/.github/scripts/run-aishell-pruned-transducer-stateless3-2022-06-20.sh
new file mode 100755
index 000000000..cf35f711b
--- /dev/null
+++ b/.github/scripts/run-aishell-pruned-transducer-stateless3-2022-06-20.sh
@@ -0,0 +1,86 @@
+#!/usr/bin/env bash
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+cd egs/aishell/ASR
+
+git lfs install
+
+fbank_url=https://huggingface.co/csukuangfj/aishell-test-dev-manifests
+log "Downloading pre-commputed fbank from $fbank_url"
+
+git clone https://huggingface.co/csukuangfj/aishell-test-dev-manifests
+ln -s $PWD/aishell-test-dev-manifests/data .
+
+log "Downloading pre-trained model from $repo_url"
+repo_url=https://huggingface.co/csukuangfj/icefall-aishell-pruned-transducer-stateless3-2022-06-20
+git clone $repo_url
+repo=$(basename $repo_url)
+
+log "Display test files"
+tree $repo/
+soxi $repo/test_wavs/*.wav
+ls -lh $repo/test_wavs/*.wav
+
+pushd $repo/exp
+ln -s pretrained-epoch-29-avg-5-torch-1.10.pt pretrained.pt
+popd
+
+for sym in 1 2 3; do
+  log "Greedy search with --max-sym-per-frame $sym"
+
+  ./pruned_transducer_stateless3/pretrained.py \
+    --method greedy_search \
+    --max-sym-per-frame $sym \
+    --checkpoint $repo/exp/pretrained.pt \
+    --lang-dir $repo/data/lang_char \
+    $repo/test_wavs/BAC009S0764W0121.wav \
+    $repo/test_wavs/BAC009S0764W0122.wav \
+    $rep/test_wavs/BAC009S0764W0123.wav
+done
+
+for method in modified_beam_search beam_search fast_beam_search; do
+  log "$method"
+
+  ./pruned_transducer_stateless3/pretrained.py \
+    --method $method \
+    --beam-size 4 \
+    --checkpoint $repo/exp/pretrained.pt \
+    --lang-dir $repo/data/lang_char \
+    $repo/test_wavs/BAC009S0764W0121.wav \
+    $repo/test_wavs/BAC009S0764W0122.wav \
+    $rep/test_wavs/BAC009S0764W0123.wav
+done
+
+echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
+echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
+if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode"  ]]; then
+  mkdir -p pruned_transducer_stateless3/exp
+  ln -s $PWD/$repo/exp/pretrained.pt pruned_transducer_stateless3/exp/epoch-999.pt
+  ln -s $PWD/$repo/data/lang_char data/
+
+  ls -lh data
+  ls -lh pruned_transducer_stateless3/exp
+
+  log "Decoding test and dev"
+
+  # use a small value for decoding with CPU
+  max_duration=100
+
+  for method in greedy_search fast_beam_search modified_beam_search; do
+    log "Decoding with $method"
+
+    ./pruned_transducer_stateless3/decode.py \
+      --decoding-method $method \
+      --epoch 999 \
+      --avg 1 \
+      --max-duration $max_duration \
+      --exp-dir pruned_transducer_stateless3/exp
+  done
+
+  rm pruned_transducer_stateless3/exp/*.pt
+fi
diff --git a/.github/workflows/run-aishell-2022-06-20.yml b/.github/workflows/run-aishell-2022-06-20.yml
new file mode 100644
index 000000000..e684e598e
--- /dev/null
+++ b/.github/workflows/run-aishell-2022-06-20.yml
@@ -0,0 +1,119 @@
+# Copyright      2022  Fangjun Kuang (csukuangfj@gmail.com)
+
+# See ../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: run-aishell-2022-06-20
+# pruned RNN-T + reworked model with random combiner
+# https://huggingface.co/csukuangfj/icefall-aishell-pruned-transducer-stateless3-2022-06-20
+
+on:
+  push:
+    branches:
+      - master
+  pull_request:
+    types: [labeled]
+
+  schedule:
+    # minute (0-59)
+    # hour (0-23)
+    # day of the month (1-31)
+    # month (1-12)
+    # day of the week (0-6)
+    # nightly build at 15:50 UTC time every day
+    - cron: "50 15 * * *"
+
+jobs:
+  run_aishell_2022_06_20:
+    if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-18.04]
+        python-version: [3.7, 3.8, 3.9]
+
+      fail-fast: false
+
+    steps:
+      - uses: actions/checkout@v2
+        with:
+          fetch-depth: 0
+
+      - name: Setup Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v2
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: 'pip'
+          cache-dependency-path: '**/requirements-ci.txt'
+
+      - name: Install Python dependencies
+        run: |
+          grep -v '^#' ./requirements-ci.txt  | xargs -n 1 -L 1 pip install
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf
+
+      - name: Cache kaldifeat
+        id: my-cache
+        uses: actions/cache@v2
+        with:
+          path: |
+            ~/tmp/kaldifeat
+          key: cache-tmp-${{ matrix.python-version }}
+
+      - name: Install kaldifeat
+        if: steps.my-cache.outputs.cache-hit != 'true'
+        shell: bash
+        run: |
+          .github/scripts/install-kaldifeat.sh
+
+      - name: Inference with pre-trained model
+        shell: bash
+        env:
+          GITHUB_EVENT_NAME: ${{ github.event_name }}
+          GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
+        run: |
+          sudo apt-get -qq install git-lfs tree sox
+          export PYTHONPATH=$PWD:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
+          export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
+
+          .github/scripts/run-aishell-pruned-transducer-stateless3-2022-06-20.sh
+
+      - name: Display decoding results for aishell pruned_transducer_stateless3
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        shell: bash
+        run: |
+          cd egs/aishell/ASR/
+          tree ./pruned_transducer_stateless3/exp
+
+          cd pruned_transducer_stateless3
+          echo "results for pruned_transducer_stateless3"
+          echo "===greedy search==="
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for test" {} + | sort -n -k2
+          find exp/greedy_search -name "log-*" -exec grep -n --color "best for dev" {} + | sort -n -k2
+
+          echo "===fast_beam_search==="
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test" {} + | sort -n -k2
+          find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for dev" {} + | sort -n -k2
+
+          echo "===modified beam search==="
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test" {} + | sort -n -k2
+          find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for dev" {} + | sort -n -k2
+
+      - name: Upload decoding results for aishell pruned_transducer_stateless3
+        uses: actions/upload-artifact@v2
+        if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
+        with:
+          name: aishell-torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-pruned_transducer_stateless3-2022-06-20
+          path: egs/aishell/ASR/pruned_transducer_stateless3/exp/
diff --git a/egs/aishell/ASR/README.md b/egs/aishell/ASR/README.md
index d0a0c1829..75fc6326e 100644
--- a/egs/aishell/ASR/README.md
+++ b/egs/aishell/ASR/README.md
@@ -4,6 +4,8 @@
 Please refer to <https://icefall.readthedocs.io/en/latest/recipes/aishell/index.html>
 for how to run models in this recipe.
 
+
+
 # Transducers
 
 There are various folders containing the name `transducer` in this folder.
@@ -14,6 +16,7 @@ The following table lists the differences among them.
 | `transducer_stateless`             | Conformer | Embedding + Conv1d | with `k2.rnnt_loss`                                                               |
 | `transducer_stateless_modified`    | Conformer | Embedding + Conv1d | with modified transducer from `optimized_transducer`                     |
 | `transducer_stateless_modified-2`  | Conformer | Embedding + Conv1d | with modified transducer from `optimized_transducer` + extra data      |
+| `pruned_transducer_stateless3`     | Conformer (reworked) | Embedding + Conv1d | pruned RNN-T + reworked model with random combiner + using aidatatang_20zh as extra data|
 
 The decoder in `transducer_stateless` is modified from the paper
 [Rnn-Transducer with Stateless Prediction Network](https://ieeexplore.ieee.org/document/9054419/).
diff --git a/egs/aishell/ASR/RESULTS.md b/egs/aishell/ASR/RESULTS.md
index ecc93c21b..b420a1982 100644
--- a/egs/aishell/ASR/RESULTS.md
+++ b/egs/aishell/ASR/RESULTS.md
@@ -1,10 +1,93 @@
 ## Results
-### Aishell training result(Transducer-stateless)
+
+### Aishell training result(Stateless Transducer)
+
+#### Pruned transducer stateless 3
+
+See <https://github.com/k2-fsa/icefall/pull/436>
+
+
+[./pruned_transducer_stateless3](./pruned_transducer_stateless3)
+
+It uses pruned RNN-T.
+
+|                        | test | dev  | comment                               |
+|------------------------|------|------|---------------------------------------|
+| greedy search          | 5.39 | 5.09 | --epoch 29 --avg 5 --max-duration 600 |
+| modified beam search   | 5.05 | 4.79 | --epoch 29 --avg 5 --max-duration 600 |
+| fast beam search       | 5.13 | 4.91 | --epoch 29 --avg 5 --max-duration 600 |
+
+Training command is:
+
+```bash
+./prepare.sh
+./prepare_aidatatang_200zh.sh
+
+export CUDA_VISIBLE_DEVICES="4,5,6,7"
+
+./pruned_transducer_stateless3/train.py \
+  --exp-dir ./pruned_transducer_stateless3/exp-context-size-1 \
+  --world-size 4 \
+  --max-duration 200 \
+  --datatang-prob 0.5 \
+  --start-epoch 1 \
+  --num-epochs 30 \
+  --use-fp16 1 \
+  --num-encoder-layers 12 \
+  --dim-feedforward 2048 \
+  --nhead 8 \
+  --encoder-dim 512 \
+  --context-size 1 \
+  --decoder-dim 512 \
+  --joiner-dim 512 \
+  --master-port 12356
+```
+
+**Caution**: It uses `--context-size=1`.
+
+The tensorboard log is available at
+<https://tensorboard.dev/experiment/OKKacljwR6ik7rbDr5gMqQ>
+
+The decoding command is:
+
+```bash
+for epoch in 29; do
+  for avg in 5; do
+    for m in greedy_search modified_beam_search fast_beam_search; do
+      ./pruned_transducer_stateless3/decode.py \
+        --exp-dir ./pruned_transducer_stateless3/exp-context-size-1 \
+        --epoch $epoch \
+        --avg $avg \
+        --use-averaged-model 1 \
+        --max-duration 600 \
+        --decoding-method $m \
+        --num-encoder-layers 12 \
+        --dim-feedforward 2048 \
+        --nhead 8 \
+        --context-size 1 \
+        --encoder-dim 512 \
+        --decoder-dim 512 \
+        --joiner-dim 512
+    done
+  done
+done
+```
+
+Pretrained models, training logs, decoding logs, and decoding results
+are available at
+<https://huggingface.co/csukuangfj/icefall-aishell-pruned-transducer-stateless3-2022-06-20>
+
+We have a tutorial in [sherpa](https://github.com/k2-fsa/sherpa) about how
+to use the pre-trained model for non-streaming ASR. See
+<https://k2-fsa.github.io/sherpa/offline_asr/conformer/aishell.html>
 
 #### 2022-03-01
 
 [./transducer_stateless_modified-2](./transducer_stateless_modified-2)
 
+It uses [optimized_transducer](https://github.com/csukuangfj/optimized_transducer)
+for computing RNN-T loss.
+
 Stateless transducer + modified transducer + using [aidatatang_200zh](http://www.openslr.org/62/) as extra training data.
 
 
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/aidatatang_200zh.py b/egs/aishell/ASR/pruned_transducer_stateless3/aidatatang_200zh.py
new file mode 120000
index 000000000..9a799406b
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/aidatatang_200zh.py
@@ -0,0 +1 @@
+../transducer_stateless_modified-2/aidatatang_200zh.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/aishell.py b/egs/aishell/ASR/pruned_transducer_stateless3/aishell.py
new file mode 120000
index 000000000..1b5f38a54
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/aishell.py
@@ -0,0 +1 @@
+../transducer_stateless_modified-2/aishell.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/asr_datamodule.py b/egs/aishell/ASR/pruned_transducer_stateless3/asr_datamodule.py
new file mode 120000
index 000000000..ae3bdd1e0
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/asr_datamodule.py
@@ -0,0 +1 @@
+../transducer_stateless_modified-2/asr_datamodule.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/beam_search.py b/egs/aishell/ASR/pruned_transducer_stateless3/beam_search.py
new file mode 120000
index 000000000..e24eca39f
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/beam_search.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/beam_search.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/conformer.py b/egs/aishell/ASR/pruned_transducer_stateless3/conformer.py
new file mode 120000
index 000000000..c7c1a4b6e
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/conformer.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless5/conformer.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/decode.py b/egs/aishell/ASR/pruned_transducer_stateless3/decode.py
new file mode 100755
index 000000000..f686174f3
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/decode.py
@@ -0,0 +1,637 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
+#                                                 Zengwei Yao)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+(1) greedy search
+./pruned_transducer_stateless3/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless3/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless3/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search
+./pruned_transducer_stateless3/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 4 \
+    --max-contexts 4 \
+    --max-states 8
+"""
+
+
+import argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import torch
+import torch.nn as nn
+from aishell import AIShell
+from asr_datamodule import AsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=30,
+        help="""It specifies the checkpoint to use for decoding.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=False,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless3/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --decoding-method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --decoding-method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --decoding-method is
+        fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=1,
+        help="The context size in the decoder. 1 means bigram; "
+        "2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="""Maximum number of symbols per frame.
+        Used only when --decoding_method is greedy_search""",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    token_table: k2.SymbolTable,
+    batch: dict,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[List[str]]]:
+    """Decode one batch and return the result in a dict. The dict has the
+    following format:
+
+        - key: It indicates the setting used for decoding. For example,
+               if greedy_search is used, it would be "greedy_search"
+               If beam search with a beam size of 7 is used, it would be
+               "beam_7"
+        - value: It contains the decoding result. `len(value)` equals to
+                 batch size. `value[i]` is the decoding result for the i-th
+                 utterance in the given batch.
+    Args:
+      params:
+        It's the return value of :func:`get_params`.
+      model:
+        The neural model.
+      token_table:
+        It maps token ID to a string.
+      batch:
+        It is the return value from iterating
+        `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
+        for the format of the `batch`.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return the decoding result. See above description for the format of
+      the returned dict.
+    """
+    device = next(model.parameters()).device
+    feature = batch["inputs"]
+    assert feature.ndim == 3
+
+    feature = feature.to(device)
+    # at entry, feature is (N, T, C)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    encoder_out, encoder_out_lens = model.encoder(
+        x=feature, x_lens=feature_lens
+    )
+
+    if params.decoding_method == "fast_beam_search":
+        hyp_tokens = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+    elif (
+        params.decoding_method == "greedy_search"
+        and params.max_sym_per_frame == 1
+    ):
+        hyp_tokens = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+    elif params.decoding_method == "modified_beam_search":
+        hyp_tokens = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+    else:
+        hyp_tokens = []
+        batch_size = encoder_out.size(0)
+        for i in range(batch_size):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.decoding_method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.decoding_method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.decoding_method}"
+                )
+            hyp_tokens.append(hyp)
+
+    hyps = [[token_table[t] for t in tokens] for tokens in hyp_tokens]
+
+    if params.decoding_method == "greedy_search":
+        return {"greedy_search": hyps}
+    elif params.decoding_method == "fast_beam_search":
+        return {
+            (
+                f"beam_{params.beam}_"
+                f"max_contexts_{params.max_contexts}_"
+                f"max_states_{params.max_states}"
+            ): hyps
+        }
+    else:
+        return {f"beam_size_{params.beam_size}": hyps}
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    params: AttributeDict,
+    model: nn.Module,
+    token_table: k2.SymbolTable,
+    decoding_graph: Optional[k2.Fsa] = None,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The neural model.
+      token_table:
+        It maps a token ID to a string.
+      decoding_graph:
+        The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
+        only when --decoding_method is fast_beam_search.
+    Returns:
+      Return a dict, whose key may be "greedy_search" if greedy search
+      is used, or it may be "beam_7" if beam size of 7 is used.
+      Its value is a list of tuples. Each tuple contains two elements:
+      The first is the reference transcript, and the second is the
+      predicted result.
+    """
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    if params.decoding_method == "greedy_search":
+        log_interval = 50
+    else:
+        log_interval = 20
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            token_table=token_table,
+            decoding_graph=decoding_graph,
+            batch=batch,
+        )
+
+        for name, hyps in hyps_dict.items():
+            this_batch = []
+            assert len(hyps) == len(texts)
+            for hyp_words, ref_text in zip(hyps, texts):
+                ref_words = ref_text.split()
+                this_batch.append((ref_words, hyp_words))
+
+            results[name].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % log_interval == 0:
+            batch_str = f"{batch_idx}/{num_batches}"
+
+            logging.info(
+                f"batch {batch_str}, cuts processed until now is {num_cuts}"
+            )
+    return results
+
+
+def save_results(
+    params: AttributeDict,
+    test_set_name: str,
+    results_dict: Dict[str, List[Tuple[List[int], List[int]]]],
+):
+    test_set_wers = dict()
+    for key, results in results_dict.items():
+        recog_path = (
+            params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        store_transcripts(filename=recog_path, texts=results)
+        logging.info(f"The transcripts are stored in {recog_path}")
+
+        # The following prints out WERs, per-word error statistics and aligned
+        # ref/hyp pairs.
+        errs_filename = (
+            params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
+        )
+        # we compute CER for aishell dataset.
+        results_char = []
+        for res in results:
+            results_char.append((list("".join(res[0])), list("".join(res[1]))))
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results_char, enable_log=True
+            )
+            test_set_wers[key] = wer
+
+        logging.info("Wrote detailed error stats to {}".format(errs_filename))
+
+    test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
+    errs_info = (
+        params.res_dir
+        / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
+    )
+    with open(errs_info, "w") as f:
+        print("settings\tCER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, CER of different settings are:\n".format(test_set_name)
+    note = "\tbest for {}".format(test_set_name)
+    for key, val in test_set_wers:
+        s += "{}\t{}{}\n".format(key, val, note)
+        note = ""
+    logging.info(s)
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    params = get_params()
+    params.update(vars(args))
+
+    assert params.decoding_method in (
+        "greedy_search",
+        "beam_search",
+        "fast_beam_search",
+        "modified_beam_search",
+    )
+    params.res_dir = params.exp_dir / params.decoding_method
+
+    if params.iter > 0:
+        params.suffix = f"iter-{params.iter}-avg-{params.avg}"
+    else:
+        params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
+
+    if "fast_beam_search" in params.decoding_method:
+        params.suffix += f"-beam-{params.beam}"
+        params.suffix += f"-max-contexts-{params.max_contexts}"
+        params.suffix += f"-max-states-{params.max_states}"
+    elif "beam_search" in params.decoding_method:
+        params.suffix += (
+            f"-{params.decoding_method}-beam-size-{params.beam_size}"
+        )
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    if params.use_averaged_model:
+        params.suffix += "-use-averaged-model"
+
+    setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
+    logging.info("Decoding started")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(
+                params.exp_dir, iteration=-params.iter
+            )[: params.avg]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device), strict=False
+            )
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints(filenames, device=device), strict=False
+            )
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(
+                params.exp_dir, iteration=-params.iter
+            )[: params.avg + 1]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                ),
+                strict=False,
+            )
+
+    model.to(device)
+    model.eval()
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    asr_datamodule = AsrDataModule(args)
+    aishell = AIShell(manifest_dir=args.manifest_dir)
+    test_cuts = aishell.test_cuts()
+    dev_cuts = aishell.valid_cuts()
+    test_dl = asr_datamodule.test_dataloaders(test_cuts)
+    dev_dl = asr_datamodule.test_dataloaders(dev_cuts)
+
+    test_sets = ["test", "dev"]
+    test_dls = [test_dl, dev_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dls):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            token_table=lexicon.token_table,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/decoder.py b/egs/aishell/ASR/pruned_transducer_stateless3/decoder.py
new file mode 120000
index 000000000..722e1c894
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/decoder.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/decoder.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/encoder_interface.py b/egs/aishell/ASR/pruned_transducer_stateless3/encoder_interface.py
new file mode 120000
index 000000000..f58253127
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/encoder_interface.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/encoder_interface.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/exp-context-size-1 b/egs/aishell/ASR/pruned_transducer_stateless3/exp-context-size-1
new file mode 120000
index 000000000..bcd4abc2f
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/exp-context-size-1
@@ -0,0 +1 @@
+/ceph-fj/fangjun/open-source/icefall-aishell/egs/aishell/ASR/pruned_transducer_stateless3/exp-context-size-1
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/export.py b/egs/aishell/ASR/pruned_transducer_stateless3/export.py
new file mode 100755
index 000000000..307895a76
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/export.py
@@ -0,0 +1,277 @@
+#!/usr/bin/env python3
+#
+# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+"""
+Usage:
+./pruned_transducer_stateless3/export.py \
+  --exp-dir ./pruned_transducer_stateless3/exp \
+  --jit 0 \
+  --epoch 29 \
+  --avg 5
+
+It will generate a file exp_dir/pretrained-epoch-29-avg-5.pt
+
+To use the generated file with `pruned_transducer_stateless3/decode.py`,
+you can do::
+
+    cd /path/to/exp_dir
+    ln -s pretrained-epoch-29-avg-5.pt epoch-9999.pt
+
+    cd /path/to/egs/aishell/ASR
+    ./pruned_transducer_stateless3/decode.py \
+        --exp-dir ./pruned_transducer_stateless3/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100 \
+        --lang-dir data/lang_char
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=29,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 1.
+        You can specify --avg to use more checkpoints for model averaging.""",
+    )
+
+    parser.add_argument(
+        "--iter",
+        type=int,
+        default=0,
+        help="""If positive, --epoch is ignored and it
+        will use the checkpoint exp_dir/checkpoint-iter.pt.
+        You can specify --avg to use more checkpoints for model averaging.
+        """,
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch' and '--iter'",
+    )
+
+    parser.add_argument(
+        "--use-averaged-model",
+        type=str2bool,
+        default=True,
+        help="Whether to load averaged model. Currently it only supports "
+        "using --epoch. If True, it would decode with the averaged model "
+        "over the epoch range from `epoch-avg` (excluded) to `epoch`."
+        "Actually only the models with epoch number of `epoch-avg` and "
+        "`epoch` are loaded for averaging. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=Path,
+        default=Path("pruned_transducer_stateless3/exp"),
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default=Path("data/lang_char"),
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=1,
+        help="The context size in the decoder. 1 means bigram; "
+        "2 means tri-gram",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if not params.use_averaged_model:
+        if params.iter > 0:
+            filenames = find_checkpoints(
+                params.exp_dir, iteration=-params.iter
+            )[: params.avg]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+        elif params.avg == 1:
+            load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        else:
+            start = params.epoch - params.avg + 1
+            filenames = []
+            for i in range(start, params.epoch + 1):
+                if i >= 1:
+                    filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+            logging.info(f"averaging {filenames}")
+            model.to(device)
+            model.load_state_dict(average_checkpoints(filenames, device=device))
+    else:
+        if params.iter > 0:
+            filenames = find_checkpoints(
+                params.exp_dir, iteration=-params.iter
+            )[: params.avg + 1]
+            if len(filenames) == 0:
+                raise ValueError(
+                    f"No checkpoints found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            elif len(filenames) < params.avg + 1:
+                raise ValueError(
+                    f"Not enough checkpoints ({len(filenames)}) found for"
+                    f" --iter {params.iter}, --avg {params.avg}"
+                )
+            filename_start = filenames[-1]
+            filename_end = filenames[0]
+            logging.info(
+                "Calculating the averaged model over iteration checkpoints"
+                f" from {filename_start} (excluded) to {filename_end}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+        else:
+            assert params.avg > 0, params.avg
+            start = params.epoch - params.avg
+            assert start >= 1, start
+            filename_start = f"{params.exp_dir}/epoch-{start}.pt"
+            filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
+            logging.info(
+                f"Calculating the averaged model over epoch range from "
+                f"{start} (excluded) to {params.epoch}"
+            )
+            model.to(device)
+            model.load_state_dict(
+                average_checkpoints_with_averaged_model(
+                    filename_start=filename_start,
+                    filename_end=filename_end,
+                    device=device,
+                )
+            )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = (
+            params.exp_dir / f"cpu_jit-epoch-{params.epoch}-avg-{params.avg}.pt"
+        )
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = (
+            params.exp_dir
+            / f"pretrained-epoch-{params.epoch}-avg-{params.avg}.pt"
+        )
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = (
+        "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    )
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/joiner.py b/egs/aishell/ASR/pruned_transducer_stateless3/joiner.py
new file mode 120000
index 000000000..9052f3cbb
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/joiner.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/joiner.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/model.py b/egs/aishell/ASR/pruned_transducer_stateless3/model.py
new file mode 100644
index 000000000..e150e8230
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/model.py
@@ -0,0 +1,236 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang, Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import Optional
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from scaling import ScaledLinear
+
+from icefall.utils import add_sos
+
+
+class Transducer(nn.Module):
+    """It implements https://arxiv.org/pdf/1211.3711.pdf
+    "Sequence Transduction with Recurrent Neural Networks"
+    """
+
+    def __init__(
+        self,
+        encoder: EncoderInterface,
+        decoder: nn.Module,
+        joiner: nn.Module,
+        encoder_dim: int,
+        decoder_dim: int,
+        joiner_dim: int,
+        vocab_size: int,
+        decoder_datatang: Optional[nn.Module] = None,
+        joiner_datatang: Optional[nn.Module] = None,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, encoder_dm) and
+            `logit_lens` of shape (N,).
+          decoder:
+            It is the prediction network in the paper. Its input shape
+            is (N, U) and its output shape is (N, U, decoder_dim).
+            It should contain one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, encoder_dim) and
+            (N, U, decoder_dim). Its output shape is (N, T, U, vocab_size).
+            Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+          encoder_dim:
+            Output dimension of the encoder network.
+          decoder_dim:
+            Output dimension of the decoder network.
+          joiner_dim:
+            Input dimension of the joiner network.
+          vocab_size:
+            Output dimension of the joiner network.
+          decoder_datatang:
+            Optional. The decoder network for the aidatatang_200zh dataset.
+          joiner_datatang:
+            Optional. The joiner network for the aidatatang_200zh dataset.
+        """
+        super().__init__()
+
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+        self.decoder_datatang = decoder_datatang
+        self.joiner_datatang = joiner_datatang
+
+        self.simple_am_proj = ScaledLinear(
+            encoder_dim, vocab_size, initial_speed=0.5
+        )
+        self.simple_lm_proj = ScaledLinear(decoder_dim, vocab_size)
+
+        if decoder_datatang is not None:
+            self.simple_am_proj_datatang = ScaledLinear(
+                encoder_dim, vocab_size, initial_speed=0.5
+            )
+            self.simple_lm_proj_datatang = ScaledLinear(decoder_dim, vocab_size)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        aishell: bool = True,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+        warmup: float = 1.0,
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C).
+          x_lens:
+            A 1-D tensor of shape (N,). It contains the number of frames in `x`
+            before padding.
+          y:
+            A ragged tensor with 2 axes [utt][label]. It contains labels of each
+            utterance.
+          aishell:
+            True to use the decoder and joiner for the aishell dataset.
+            False to use the decoder and joiner for the aidatatang_200zh
+            dataset.
+          prune_range:
+            The prune range for rnnt loss, it means how many symbols(context)
+            we are considering for each frame to compute the loss.
+          am_scale:
+            The scale to smooth the loss with am (output of encoder network)
+            part
+          lm_scale:
+            The scale to smooth the loss with lm (output of predictor network)
+            part
+          warmup:
+            A value warmup >= 0 that determines which modules are active, values
+            warmup > 1 "are fully warmed up" and all modules will be active.
+        Returns:
+          Return the transducer loss.
+
+        Note:
+           Regarding am_scale & lm_scale, it will make the loss-function one of
+           the form:
+              lm_scale * lm_probs + am_scale * am_probs +
+              (1-lm_scale-am_scale) * combined_probs
+        """
+        assert x.ndim == 3, x.shape
+        assert x_lens.ndim == 1, x_lens.shape
+        assert y.num_axes == 2, y.num_axes
+
+        assert x.size(0) == x_lens.size(0) == y.dim0
+
+        encoder_out, encoder_out_lens = self.encoder(x, x_lens, warmup=warmup)
+        assert torch.all(encoder_out_lens > 0)
+
+        if aishell:
+            decoder = self.decoder
+            simple_lm_proj = self.simple_lm_proj
+            simple_am_proj = self.simple_am_proj
+            joiner = self.joiner
+        else:
+            decoder = self.decoder_datatang
+            simple_lm_proj = self.simple_lm_proj_datatang
+            simple_am_proj = self.simple_am_proj_datatang
+            joiner = self.joiner_datatang
+
+        # Now for the decoder, i.e., the prediction network
+        row_splits = y.shape.row_splits(1)
+        y_lens = row_splits[1:] - row_splits[:-1]
+
+        blank_id = decoder.blank_id
+        sos_y = add_sos(y, sos_id=blank_id)
+
+        # sos_y_padded: [B, S + 1], start with SOS.
+        sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
+
+        # decoder_out: [B, S + 1, decoder_dim]
+        decoder_out = decoder(sos_y_padded)
+
+        # Note: y does not start with SOS
+        # y_padded : [B, S]
+        y_padded = y.pad(mode="constant", padding_value=0)
+
+        y_padded = y_padded.to(torch.int64)
+        boundary = torch.zeros(
+            (x.size(0), 4), dtype=torch.int64, device=x.device
+        )
+        boundary[:, 2] = y_lens
+        boundary[:, 3] = encoder_out_lens
+
+        lm = simple_lm_proj(decoder_out)
+        am = simple_am_proj(encoder_out)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+                lm=lm.float(),
+                am=am.float(),
+                symbols=y_padded,
+                termination_symbol=blank_id,
+                lm_only_scale=lm_scale,
+                am_only_scale=am_scale,
+                boundary=boundary,
+                reduction="sum",
+                return_grad=True,
+            )
+
+        # ranges : [B, T, prune_range]
+        ranges = k2.get_rnnt_prune_ranges(
+            px_grad=px_grad,
+            py_grad=py_grad,
+            boundary=boundary,
+            s_range=prune_range,
+        )
+
+        # am_pruned : [B, T, prune_range, encoder_dim]
+        # lm_pruned : [B, T, prune_range, decoder_dim]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=joiner.encoder_proj(encoder_out),
+            lm=joiner.decoder_proj(decoder_out),
+            ranges=ranges,
+        )
+
+        # logits : [B, T, prune_range, vocab_size]
+
+        # project_input=False since we applied the decoder's input projections
+        # prior to do_rnnt_pruning (this is an optimization for speed).
+        logits = joiner(am_pruned, lm_pruned, project_input=False)
+
+        with torch.cuda.amp.autocast(enabled=False):
+            pruned_loss = k2.rnnt_loss_pruned(
+                logits=logits.float(),
+                symbols=y_padded,
+                ranges=ranges,
+                termination_symbol=blank_id,
+                boundary=boundary,
+                reduction="sum",
+            )
+
+        return (simple_loss, pruned_loss)
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/optim.py b/egs/aishell/ASR/pruned_transducer_stateless3/optim.py
new file mode 120000
index 000000000..0a2f285aa
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/optim.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/optim.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/pretrained.py b/egs/aishell/ASR/pruned_transducer_stateless3/pretrained.py
new file mode 100755
index 000000000..5cda411bc
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/pretrained.py
@@ -0,0 +1,337 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                    Wei Kang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+
+(1) greedy search
+./pruned_transducer_stateless3/pretrained.py \
+  --checkpoint /path/to/pretrained.pt \
+  --lang-dir /path/to/lang_char \
+  --method greedy_search \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(2) beam search
+./pruned_transducer_stateless3/pretrained.py \
+  --checkpoint /path/to/pretrained.pt \
+  --lang-dir /path/to/lang_char \
+  --method beam_search \
+  --beam-size 4 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(3) modified beam search
+./pruned_transducer_stateless3/pretrained.py \
+  --checkpoint /path/to/pretrained.pt \
+  --lang-dir /path/to/lang_char \
+  --method modified_beam_search \
+  --beam-size 4 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+
+(4) fast beam search
+./pruned_transducer_stateless3/pretrained.py \
+  --checkpoint /path/to/pretrained.pt \
+  --lang-dir /path/to/lang_char \
+  --method fast_beam_search \
+  --beam-size 4 \
+  /path/to/foo.wav \
+  /path/to/bar.wav
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from typing import List
+
+import k2
+import kaldifeat
+import torch
+import torchaudio
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from torch.nn.utils.rnn import pad_sequence
+from train import add_model_arguments, get_params, get_transducer_model
+
+from icefall.lexicon import Lexicon
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--checkpoint",
+        type=str,
+        required=True,
+        help="Path to the checkpoint. "
+        "The checkpoint is assumed to be saved by "
+        "icefall.checkpoint.save_checkpoint().",
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default=Path("data/lang_char"),
+        help="The lang dir",
+    )
+
+    parser.add_argument(
+        "--method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_search
+          - beam_search
+          - modified_beam_search
+          - fast_beam_search
+        """,
+    )
+
+    parser.add_argument(
+        "sound_files",
+        type=str,
+        nargs="+",
+        help="The input sound file(s) to transcribe. "
+        "Supported formats are those supported by torchaudio.load(). "
+        "For example, wav and flac are supported. "
+        "The sample rate has to be 16kHz.",
+    )
+
+    parser.add_argument(
+        "--sample-rate",
+        type=int,
+        default=16000,
+        help="The sample rate of the input sound file",
+    )
+
+    parser.add_argument(
+        "--beam-size",
+        type=int,
+        default=4,
+        help="""An integer indicating how many candidates we will keep for each
+        frame. Used only when --method is beam_search or
+        modified_beam_search.""",
+    )
+
+    parser.add_argument(
+        "--beam",
+        type=float,
+        default=4,
+        help="""A floating point value to calculate the cutoff score during beam
+        search (i.e., `cutoff = max-score - beam`), which is the same as the
+        `beam` in Kaldi.
+        Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-contexts",
+        type=int,
+        default=4,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--max-states",
+        type=int,
+        default=8,
+        help="""Used only when --method is fast_beam_search""",
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=1,
+        help="The context size in the decoder. 1 means bigram; "
+        "2 means tri-gram",
+    )
+    parser.add_argument(
+        "--max-sym-per-frame",
+        type=int,
+        default=1,
+        help="Maximum number of symbols per frame. "
+        "Use only when --method is greedy_search",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def read_sound_files(
+    filenames: List[str], expected_sample_rate: float
+) -> List[torch.Tensor]:
+    """Read a list of sound files into a list 1-D float32 torch tensors.
+    Args:
+      filenames:
+        A list of sound filenames.
+      expected_sample_rate:
+        The expected sample rate of the sound files.
+    Returns:
+      Return a list of 1-D float32 torch tensors.
+    """
+    ans = []
+    for f in filenames:
+        wave, sample_rate = torchaudio.load(f)
+        assert sample_rate == expected_sample_rate, (
+            f"expected sample rate: {expected_sample_rate}. "
+            f"Given: {sample_rate}"
+        )
+        # We use only the first channel
+        ans.append(wave[0])
+    return ans
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    params = get_params()
+    params.update(vars(args))
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    checkpoint = torch.load(args.checkpoint, map_location="cpu")
+    model.load_state_dict(checkpoint["model"], strict=False)
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    logging.info("Constructing Fbank computer")
+    opts = kaldifeat.FbankOptions()
+    opts.device = device
+    opts.frame_opts.dither = 0
+    opts.frame_opts.snip_edges = False
+    opts.frame_opts.samp_freq = params.sample_rate
+    opts.mel_opts.num_bins = params.feature_dim
+
+    fbank = kaldifeat.Fbank(opts)
+
+    logging.info(f"Reading sound files: {params.sound_files}")
+    waves = read_sound_files(
+        filenames=params.sound_files, expected_sample_rate=params.sample_rate
+    )
+    waves = [w.to(device) for w in waves]
+
+    logging.info("Decoding started")
+    features = fbank(waves)
+    feature_lens = [f.size(0) for f in features]
+    feature_lens = torch.tensor(feature_lens, device=device)
+
+    features = pad_sequence(
+        features, batch_first=True, padding_value=math.log(1e-10)
+    )
+
+    encoder_out, encoder_out_lens = model.encoder(
+        x=features, x_lens=feature_lens
+    )
+
+    num_waves = encoder_out.size(0)
+    hyp_list = []
+    logging.info(f"Using {params.method}")
+
+    if params.method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+        hyp_list = fast_beam_search_one_best(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+        )
+    elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
+        hyp_list = greedy_search_batch(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+        )
+    elif params.method == "modified_beam_search":
+        hyp_list = modified_beam_search(
+            model=model,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam_size,
+        )
+    else:
+        for i in range(num_waves):
+            # fmt: off
+            encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
+            # fmt: on
+            if params.method == "greedy_search":
+                hyp = greedy_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    max_sym_per_frame=params.max_sym_per_frame,
+                )
+            elif params.method == "beam_search":
+                hyp = beam_search(
+                    model=model,
+                    encoder_out=encoder_out_i,
+                    beam=params.beam_size,
+                )
+            else:
+                raise ValueError(
+                    f"Unsupported decoding method: {params.method}"
+                )
+            hyp_list.append(hyp)
+
+    hyps = []
+    for hyp in hyp_list:
+        hyps.append([lexicon.token_table[i] for i in hyp])
+
+    s = "\n"
+    for filename, hyp in zip(params.sound_files, hyps):
+        words = " ".join(hyp)
+        s += f"{filename}:\n{words}\n\n"
+    logging.info(s)
+
+    logging.info("Decoding Done")
+
+
+if __name__ == "__main__":
+    formatter = (
+        "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    )
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/scaling.py b/egs/aishell/ASR/pruned_transducer_stateless3/scaling.py
new file mode 120000
index 000000000..c10cdfe12
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/scaling.py
@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/scaling.py
\ No newline at end of file
diff --git a/egs/aishell/ASR/pruned_transducer_stateless3/train.py b/egs/aishell/ASR/pruned_transducer_stateless3/train.py
new file mode 100755
index 000000000..02efe94fe
--- /dev/null
+++ b/egs/aishell/ASR/pruned_transducer_stateless3/train.py
@@ -0,0 +1,1229 @@
+#!/usr/bin/env python3
+# Copyright    2021-2022  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                       Wei Kang,
+#                                                       Mingshuang Luo,)
+#                                                       Zengwei Yao)
+# Copyright    2021                      (Pingfeng Luo)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Usage:
+
+./prepare.sh
+./prepare_aidatatang_200zh.sh
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+
+./pruned_transducer_stateless3/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 0 \
+  --exp-dir pruned_transducer_stateless3/exp \
+  --max-duration 300 \
+  --datatang-prob 0.2
+
+# For mix precision training:
+
+./pruned_transducer_stateless3/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 1 \
+  --use-fp16 1 \
+  --exp-dir pruned_transducer_stateless3/exp \
+  --max-duration 550
+"""
+
+
+import argparse
+import copy
+import logging
+import random
+import warnings
+from pathlib import Path
+from shutil import copyfile
+from typing import Any, Dict, Optional, Tuple, Union
+
+import k2
+import optim
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+from aidatatang_200zh import AIDatatang200zh
+from aishell import AIShell
+from asr_datamodule import AsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+from lhotse import CutSet, load_manifest
+from lhotse.cut import Cut
+from lhotse.dataset.sampling.base import CutSampler
+from lhotse.utils import fix_random_seed
+from model import Transducer
+from optim import Eden, Eve
+from torch import Tensor
+from torch.cuda.amp import GradScaler
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall import diagnostics
+from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler
+from icefall.checkpoint import load_checkpoint, remove_checkpoints
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.checkpoint import (
+    save_checkpoint_with_global_batch_idx,
+    update_averaged_model,
+)
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.lexicon import Lexicon
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+LRSchedulerType = Union[
+    torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler
+]
+
+
+def add_model_arguments(parser: argparse.ArgumentParser):
+    parser.add_argument(
+        "--num-encoder-layers",
+        type=int,
+        default=12,
+        help="Number of conformer encoder layers..",
+    )
+
+    parser.add_argument(
+        "--dim-feedforward",
+        type=int,
+        default=2048,
+        help="Feedforward dimension of the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--nhead",
+        type=int,
+        default=8,
+        help="Number of attention heads in the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--encoder-dim",
+        type=int,
+        default=512,
+        help="Attention dimension in the conformer encoder layer.",
+    )
+
+    parser.add_argument(
+        "--decoder-dim",
+        type=int,
+        default=512,
+        help="Embedding dimension in the decoder model.",
+    )
+
+    parser.add_argument(
+        "--joiner-dim",
+        type=int,
+        default=512,
+        help="""Dimension used in the joiner model.
+        Outputs from the encoder and decoder model are projected
+        to this dimension before adding.
+        """,
+    )
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=30,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=1,
+        help="""Resume training from this epoch. It should be positive.
+        If larger than 1, it will load checkpoint from
+        exp-dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--start-batch",
+        type=int,
+        default=0,
+        help="""If positive, --start-epoch is ignored and
+        it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless3/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--lang-dir",
+        type=str,
+        default="data/lang_char",
+        help="""The lang dir
+        It contains language related input files such as
+        "lexicon.txt"
+        """,
+    )
+
+    parser.add_argument(
+        "--initial-lr",
+        type=float,
+        default=0.003,
+        help="The initial learning rate.  This value should not need "
+        "to be changed.",
+    )
+
+    parser.add_argument(
+        "--lr-batches",
+        type=float,
+        default=5000,
+        help="""Number of steps that affects how rapidly the learning rate
+        decreases. We suggest not to change this.""",
+    )
+
+    parser.add_argument(
+        "--lr-epochs",
+        type=float,
+        default=6,
+        help="""Number of epochs that affects how rapidly the learning rate decreases.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=1,
+        help="The context size in the decoder. 1 means bigram; "
+        "2 means tri-gram",
+    )
+
+    parser.add_argument(
+        "--prune-range",
+        type=int,
+        default=5,
+        help="The prune range for rnnt loss, it means how many symbols(context)"
+        "we are using to compute the loss",
+    )
+
+    parser.add_argument(
+        "--lm-scale",
+        type=float,
+        default=0.25,
+        help="The scale to smooth the loss with lm "
+        "(output of prediction network) part.",
+    )
+
+    parser.add_argument(
+        "--am-scale",
+        type=float,
+        default=0.0,
+        help="The scale to smooth the loss with am (output of encoder network)"
+        "part.",
+    )
+
+    parser.add_argument(
+        "--simple-loss-scale",
+        type=float,
+        default=0.5,
+        help="To get pruning ranges, we will calculate a simple version"
+        "loss(joiner is just addition), this simple loss also uses for"
+        "training (as a regularization item). We will scale the simple loss"
+        "with this parameter before adding to the final loss.",
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    parser.add_argument(
+        "--print-diagnostics",
+        type=str2bool,
+        default=False,
+        help="Accumulate stats on activations, print them and exit.",
+    )
+
+    parser.add_argument(
+        "--save-every-n",
+        type=int,
+        default=4000,
+        help="""Save checkpoint after processing this number of batches"
+        periodically. We save checkpoint to exp-dir/ whenever
+        params.batch_idx_train % save_every_n == 0. The checkpoint filename
+        has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt'
+        Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the
+        end of each epoch where `xxx` is the epoch number counting from 0.
+        """,
+    )
+
+    parser.add_argument(
+        "--keep-last-k",
+        type=int,
+        default=30,
+        help="""Only keep this number of checkpoints on disk.
+        For instance, if it is 3, there are only 3 checkpoints
+        in the exp-dir with filenames `checkpoint-xxx.pt`.
+        It does not affect checkpoints with name `epoch-xxx.pt`.
+        """,
+    )
+
+    parser.add_argument(
+        "--average-period",
+        type=int,
+        default=100,
+        help="""Update the averaged model, namely `model_avg`, after processing
+        this number of batches. `model_avg` is a separate version of model,
+        in which each floating-point parameter is the average of all the
+        parameters from the start of training. Each time we take the average,
+        we do: `model_avg = model * (average_period / batch_idx_train) +
+            model_avg * ((batch_idx_train - average_period) / batch_idx_train)`.
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--datatang-prob",
+        type=float,
+        default=0.2,
+        help="The probability to select a batch from the "
+        "aidatatang_200zh dataset",
+    )
+
+    add_model_arguments(parser)
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters.
+
+    All training related parameters that are not passed from the commandline
+    are saved in the variable `params`.
+
+    Commandline options are merged into `params` after they are parsed, so
+    you can also access them via `params`.
+
+    Explanation of options saved in `params`:
+
+        - best_train_loss: Best training loss so far. It is used to select
+                           the model that has the lowest training loss. It is
+                           updated during the training.
+
+        - best_valid_loss: Best validation loss so far. It is used to select
+                           the model that has the lowest validation loss. It is
+                           updated during the training.
+
+        - best_train_epoch: It is the epoch that has the best training loss.
+
+        - best_valid_epoch: It is the epoch that has the best validation loss.
+
+        - batch_idx_train: Used to writing statistics to tensorboard. It
+                           contains number of batches trained so far across
+                           epochs.
+
+        - log_interval:  Print training loss if batch_idx % log_interval` is 0
+
+        - reset_interval: Reset statistics if batch_idx % reset_interval is 0
+
+        - valid_interval:  Run validation if batch_idx % valid_interval is 0
+
+        - feature_dim: The model input dim. It has to match the one used
+                       in computing features.
+
+        - subsampling_factor:  The subsampling factor for the model.
+
+        - encoder_dim: Hidden dim for multi-head attention model.
+
+        - num_decoder_layers: Number of decoder layer of transformer decoder.
+
+        - warm_step: The warm_step for Noam optimizer.
+    """
+    params = AttributeDict(
+        {
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
+            "reset_interval": 200,
+            "valid_interval": 1000,
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            # parameters for Noam
+            "model_warm_step": 3000,  # arg given to model, not for lrate
+            "env_info": get_env_info(),
+        }
+    )
+
+    return params
+
+
+def get_encoder_model(params: AttributeDict) -> nn.Module:
+    # TODO: We can add an option to switch between Conformer and Transformer
+    encoder = Conformer(
+        num_features=params.feature_dim,
+        subsampling_factor=params.subsampling_factor,
+        d_model=params.encoder_dim,
+        nhead=params.nhead,
+        dim_feedforward=params.dim_feedforward,
+        num_encoder_layers=params.num_encoder_layers,
+    )
+    return encoder
+
+
+def get_decoder_model(params: AttributeDict) -> nn.Module:
+    decoder = Decoder(
+        vocab_size=params.vocab_size,
+        decoder_dim=params.decoder_dim,
+        blank_id=params.blank_id,
+        context_size=params.context_size,
+    )
+    return decoder
+
+
+def get_joiner_model(params: AttributeDict) -> nn.Module:
+    joiner = Joiner(
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+    )
+    return joiner
+
+
+def get_transducer_model(params: AttributeDict) -> nn.Module:
+    encoder = get_encoder_model(params)
+    decoder = get_decoder_model(params)
+    joiner = get_joiner_model(params)
+
+    decoder_datatang = get_decoder_model(params)
+    joiner_datatang = get_joiner_model(params)
+
+    model = Transducer(
+        encoder=encoder,
+        decoder=decoder,
+        joiner=joiner,
+        encoder_dim=params.encoder_dim,
+        decoder_dim=params.decoder_dim,
+        joiner_dim=params.joiner_dim,
+        vocab_size=params.vocab_size,
+        decoder_datatang=decoder_datatang,
+        joiner_datatang=joiner_datatang,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    model_avg: nn.Module = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+) -> Optional[Dict[str, Any]]:
+    """Load checkpoint from file.
+
+    If params.start_batch is positive, it will load the checkpoint from
+    `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if
+    params.start_epoch is larger than 1, it will load the checkpoint from
+    `params.start_epoch - 1`.
+
+    Apart from loading state dict for `model` and `optimizer` it also updates
+    `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The scheduler that we are using.
+    Returns:
+      Return a dict containing previously saved training info.
+    """
+    if params.start_batch > 0:
+        filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt"
+    elif params.start_epoch > 1:
+        filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    else:
+        return None
+
+    assert filename.is_file(), f"{filename} does not exist!"
+
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        model_avg=model_avg,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    model_avg: Optional[nn.Module] = None,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[LRSchedulerType] = None,
+    sampler: Optional[CutSampler] = None,
+    scaler: Optional[GradScaler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+      model_avg:
+        The stored model averaged from the start of training.
+      optimizer:
+        The optimizer used in the training.
+      sampler:
+       The sampler for the training dataset.
+      scaler:
+        The scaler used for mix precision training.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        model_avg=model_avg,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        sampler=sampler,
+        scaler=scaler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def is_aishell(c: Cut) -> bool:
+    """Return True if this cut is from the AIShell dataset.
+
+    Note:
+      During data preparation, we set the custom field in
+      the supervision segment of aidatatang_200zh to
+      dict(origin='aidatatang_200zh')
+      See ../local/process_aidatatang_200zh.py.
+    """
+    return c.supervisions[0].custom is None
+
+
+def compute_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute RNN-T loss given the model and its inputs.
+
+    Args:
+      params:
+        Parameters for training. See :func:`get_params`.
+      model:
+        The model for training. It is an instance of Conformer in our case.
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      is_training:
+        True for training. False for validation. When it is True, this
+        function enables autograd during computation; when it is False, it
+        disables autograd.
+     warmup: a floating point value which increases throughout training;
+        values >= 1.0 are fully warmed up and have all modules present.
+    """
+    device = (
+        model.device
+        if isinstance(model, DDP)
+        else next(model.parameters()).device
+    )
+    feature = batch["inputs"]
+    # at entry, feature is (N, T, C)
+    assert feature.ndim == 3
+    feature = feature.to(device)
+
+    supervisions = batch["supervisions"]
+    feature_lens = supervisions["num_frames"].to(device)
+
+    aishell = is_aishell(supervisions["cut"][0])
+
+    texts = batch["supervisions"]["text"]
+    y = graph_compiler.texts_to_ids(texts)
+    y = k2.RaggedTensor(y).to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            aishell=aishell,
+            prune_range=params.prune_range,
+            am_scale=params.am_scale,
+            lm_scale=params.lm_scale,
+            warmup=warmup,
+        )
+        # after the main warmup step, we keep pruned_loss_scale small
+        # for the same amount of time (model_warm_step), to avoid
+        # overwhelming the simple_loss and causing it to diverge,
+        # in case it had not fully learned the alignment yet.
+        pruned_loss_scale = (
+            0.0
+            if warmup < 1.0
+            else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0)
+        )
+        loss = (
+            params.simple_loss_scale * simple_loss
+            + pruned_loss_scale * pruned_loss
+        )
+
+    assert loss.requires_grad == is_training
+
+    info = MetricsTracker()
+    with warnings.catch_warnings():
+        warnings.simplefilter("ignore")
+        info["frames"] = (
+            (feature_lens // params.subsampling_factor).sum().item()
+        )
+
+    # Note: We use reduction=sum while computing the loss.
+    info["loss"] = loss.detach().cpu().item()
+    info["simple_loss"] = simple_loss.detach().cpu().item()
+    info["pruned_loss"] = pruned_loss.detach().cpu().item()
+
+    return loss, info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process."""
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        loss, loss_info = compute_loss(
+            params=params,
+            model=model,
+            graph_compiler=graph_compiler,
+            batch=batch,
+            is_training=False,
+        )
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: Union[nn.Module, DDP],
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    datatang_train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    rng: random.Random,
+    scaler: GradScaler,
+    model_avg: Optional[nn.Module] = None,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+    rank: int = 0,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all frames is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      scheduler:
+        The learning rate scheduler, we call step() every step.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      scaler:
+        The scaler used for mix precision training.
+      model_avg:
+        The stored model averaged from the start of training.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+      rank:
+        The rank of the node in DDP training. If no DDP is used, it should
+        be set to 0.
+    """
+    model.train()
+
+    aishell_tot_loss = MetricsTracker()
+    datatang_tot_loss = MetricsTracker()
+    tot_loss = MetricsTracker()
+
+    # index 0: for LibriSpeech
+    # index 1: for GigaSpeech
+    # This sets the probabilities for choosing which datasets
+    dl_weights = [1 - params.datatang_prob, params.datatang_prob]
+
+    iter_aishell = iter(train_dl)
+    iter_datatang = iter(datatang_train_dl)
+
+    batch_idx = 0
+
+    while True:
+        idx = rng.choices((0, 1), weights=dl_weights, k=1)[0]
+        dl = iter_aishell if idx == 0 else iter_datatang
+
+        try:
+            batch = next(dl)
+        except StopIteration:
+            break
+        batch_idx += 1
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        aishell = is_aishell(batch["supervisions"]["cut"][0])
+
+        try:
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, loss_info = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                    warmup=(params.batch_idx_train / params.model_warm_step),
+                )
+            # summary stats
+            tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+            if aishell:
+                aishell_tot_loss = (
+                    aishell_tot_loss * (1 - 1 / params.reset_interval)
+                ) + loss_info
+                prefix = "aishell"  # for logging only
+            else:
+                datatang_tot_loss = (
+                    datatang_tot_loss * (1 - 1 / params.reset_interval)
+                ) + loss_info
+                prefix = "datatang"
+
+            # NOTE: We use reduction==sum and loss is computed over utterances
+            # in the batch and there is no normalization to it so far.
+            scaler.scale(loss).backward()
+            scheduler.step_batch(params.batch_idx_train)
+            scaler.step(optimizer)
+            scaler.update()
+            optimizer.zero_grad()
+        except:  # noqa
+            display_and_save_batch(
+                batch, params=params, graph_compiler=graph_compiler
+            )
+            raise
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        if (
+            rank == 0
+            and params.batch_idx_train > 0
+            and params.batch_idx_train % params.average_period == 0
+        ):
+            update_averaged_model(
+                params=params,
+                model_cur=model,
+                model_avg=model_avg,
+            )
+
+        if (
+            params.batch_idx_train > 0
+            and params.batch_idx_train % params.save_every_n == 0
+        ):
+            save_checkpoint_with_global_batch_idx(
+                out_dir=params.exp_dir,
+                global_batch_idx=params.batch_idx_train,
+                model=model,
+                model_avg=model_avg,
+                params=params,
+                optimizer=optimizer,
+                scheduler=scheduler,
+                sampler=train_dl.sampler,
+                scaler=scaler,
+                rank=rank,
+            )
+            remove_checkpoints(
+                out_dir=params.exp_dir,
+                topk=params.keep_last_k,
+                rank=rank,
+            )
+
+        if batch_idx % params.log_interval == 0:
+            cur_lr = scheduler.get_last_lr()[0]
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, {prefix}_loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], batch size: {batch_size}, "
+                f"aishell_tot_loss[{aishell_tot_loss}], "
+                f"datatang_tot_loss[{datatang_tot_loss}], "
+                f"batch size: {batch_size}, "
+                f"lr: {cur_lr:.2e}"
+            )
+
+            if tb_writer is not None:
+                tb_writer.add_scalar(
+                    "train/learning_rate", cur_lr, params.batch_idx_train
+                )
+
+                loss_info.write_summary(
+                    tb_writer,
+                    f"train/current_{prefix}_",
+                    params.batch_idx_train,
+                )
+                tot_loss.write_summary(
+                    tb_writer, "train/tot_", params.batch_idx_train
+                )
+                aishell_tot_loss.write_summary(
+                    tb_writer, "train/aishell_tot_", params.batch_idx_train
+                )
+                datatang_tot_loss.write_summary(
+                    tb_writer, "train/datatang_tot_", params.batch_idx_train
+                )
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                graph_compiler=graph_compiler,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+            logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}")
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def filter_short_and_long_utterances(cuts: CutSet) -> CutSet:
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 12 seconds
+        #
+        # Caution: There is a reason to select 12.0 here. Please see
+        # ../local/display_manifest_statistics.py
+        #
+        # You should use ../local/display_manifest_statistics.py to get
+        # an utterance duration distribution for your dataset to select
+        # the threshold
+        return 1.0 <= c.duration <= 12.0
+
+    return cuts
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+
+    fix_random_seed(params.seed)
+    rng = random.Random(params.seed)
+    if world_size > 1:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+    logging.info(f"Device: {device}")
+
+    lexicon = Lexicon(params.lang_dir)
+    graph_compiler = CharCtcTrainingGraphCompiler(
+        lexicon=lexicon,
+        device=device,
+        oov="<unk>",
+    )
+
+    params.blank_id = 0
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    assert params.save_every_n >= params.average_period
+    model_avg: Optional[nn.Module] = None
+    if rank == 0:
+        # model_avg is only used with rank 0
+        model_avg = copy.deepcopy(model)
+
+    assert params.start_epoch > 0, params.start_epoch
+    checkpoints = load_checkpoint_if_available(
+        params=params, model=model, model_avg=model_avg
+    )
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank], find_unused_parameters=True)
+
+    optimizer = Eve(model.parameters(), lr=params.initial_lr)
+
+    scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)
+
+    if checkpoints and "optimizer" in checkpoints:
+        logging.info("Loading optimizer state dict")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    if (
+        checkpoints
+        and "scheduler" in checkpoints
+        and checkpoints["scheduler"] is not None
+    ):
+        logging.info("Loading scheduler state dict")
+        scheduler.load_state_dict(checkpoints["scheduler"])
+
+    if params.print_diagnostics:
+        opts = diagnostics.TensorDiagnosticOptions(
+            2 ** 22
+        )  # allow 4 megabytes per sub-module
+        diagnostic = diagnostics.attach_diagnostics(model, opts)
+
+    aishell = AIShell(manifest_dir=args.manifest_dir)
+    train_cuts = aishell.train_cuts()
+    train_cuts = filter_short_and_long_utterances(train_cuts)
+
+    datatang = AIDatatang200zh(manifest_dir=args.manifest_dir)
+    train_datatang_cuts = datatang.train_cuts()
+    train_datatang_cuts = filter_short_and_long_utterances(train_datatang_cuts)
+    train_datatang_cuts = train_datatang_cuts.repeat(times=None)
+
+    if args.enable_musan:
+        cuts_musan = load_manifest(
+            Path(args.manifest_dir) / "musan_cuts.jsonl.gz"
+        )
+    else:
+        cuts_musan = None
+
+    asr_datamodule = AsrDataModule(args)
+
+    train_dl = asr_datamodule.train_dataloaders(
+        train_cuts,
+        on_the_fly_feats=False,
+        cuts_musan=cuts_musan,
+    )
+
+    datatang_train_dl = asr_datamodule.train_dataloaders(
+        train_datatang_cuts,
+        on_the_fly_feats=False,
+        cuts_musan=cuts_musan,
+    )
+
+    valid_cuts = aishell.valid_cuts()
+    valid_dl = asr_datamodule.valid_dataloaders(valid_cuts)
+
+    for dl in [
+        train_dl,
+        # datatang_train_dl
+    ]:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=dl,
+            optimizer=optimizer,
+            graph_compiler=graph_compiler,
+            params=params,
+        )
+
+    scaler = GradScaler(enabled=params.use_fp16)
+    if checkpoints and "grad_scaler" in checkpoints:
+        logging.info("Loading grad scaler state dict")
+        scaler.load_state_dict(checkpoints["grad_scaler"])
+
+    logging.info(f"start training from epoch {params.start_epoch}")
+    for epoch in range(params.start_epoch, params.num_epochs + 1):
+        scheduler.step_epoch(epoch - 1)
+        fix_random_seed(params.seed + epoch - 1)
+        train_dl.sampler.set_epoch(epoch - 1)
+        datatang_train_dl.sampler.set_epoch(epoch)
+
+        if tb_writer is not None:
+            tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            datatang_train_dl=datatang_train_dl,
+            valid_dl=valid_dl,
+            rng=rng,
+            scaler=scaler,
+            tb_writer=tb_writer,
+            world_size=world_size,
+            rank=rank,
+        )
+
+        if params.print_diagnostics:
+            diagnostic.print_diagnostics()
+            break
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            model_avg=model_avg,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            sampler=train_dl.sampler,
+            scaler=scaler,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if world_size > 1:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def display_and_save_batch(
+    batch: dict,
+    params: AttributeDict,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+) -> None:
+    """Display the batch statistics and save the batch into disk.
+
+    Args:
+      batch:
+        A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
+        for the content in it.
+      params:
+        Parameters for training. See :func:`get_params`.
+    """
+    from lhotse.utils import uuid4
+
+    filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
+    logging.info(f"Saving batch to {filename}")
+    torch.save(batch, filename)
+
+    supervisions = batch["supervisions"]
+    features = batch["inputs"]
+
+    logging.info(f"features shape: {features.shape}")
+
+    y = graph_compiler.texts_to_ids(supervisions["text"])
+    num_tokens = sum(len(i) for i in y)
+    logging.info(f"num tokens: {num_tokens}")
+
+
+def scan_pessimistic_batches_for_oom(
+    model: Union[nn.Module, DDP],
+    train_dl: torch.utils.data.DataLoader,
+    optimizer: torch.optim.Optimizer,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    params: AttributeDict,
+):
+    from lhotse.dataset import find_pessimistic_batches
+
+    logging.info(
+        "Sanity check -- see if any of the batches in epoch 1 would cause OOM."
+    )
+    batches, crit_values = find_pessimistic_batches(train_dl.sampler)
+    for criterion, cuts in batches.items():
+        batch = train_dl.dataset[cuts]
+        try:
+            # warmup = 0.0 is so that the derivs for the pruned loss stay zero
+            # (i.e. are not remembered by the decaying-average in adam), because
+            # we want to avoid these params being subject to shrinkage in adam.
+            with torch.cuda.amp.autocast(enabled=params.use_fp16):
+                loss, _ = compute_loss(
+                    params=params,
+                    model=model,
+                    graph_compiler=graph_compiler,
+                    batch=batch,
+                    is_training=True,
+                    warmup=0.0 if params.start_epoch == 1 else 1.0,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except Exception as e:
+            if "CUDA out of memory" in str(e):
+                logging.error(
+                    "Your GPU ran out of memory with the current "
+                    "max_duration setting. We recommend decreasing "
+                    "max_duration and trying again.\n"
+                    f"Failing criterion: {criterion} "
+                    f"(={crit_values[criterion]}) ..."
+                )
+            display_and_save_batch(
+                batch, params=params, graph_compiler=graph_compiler
+            )
+            raise
+
+
+def main():
+    parser = get_parser()
+    AsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lang_dir = Path(args.lang_dir)
+
+    assert 0 <= args.datatang_prob < 1, args.datatang_prob
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/train.py b/egs/aishell/ASR/transducer_stateless_modified-2/train.py
index 962fffdf5..225d0d709 100755
--- a/egs/aishell/ASR/transducer_stateless_modified-2/train.py
+++ b/egs/aishell/ASR/transducer_stateless_modified-2/train.py
@@ -405,7 +405,7 @@ def compute_loss(
     is_training: bool,
 ) -> Tuple[Tensor, MetricsTracker]:
     """
-    Compute CTC loss given the model and its inputs.
+    Compute RNN-T loss given the model and its inputs.
 
     Args:
       params:
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/decoder.py b/egs/librispeech/ASR/pruned_transducer_stateless2/decoder.py
index b6d94aaf1..1ddfce034 100644
--- a/egs/librispeech/ASR/pruned_transducer_stateless2/decoder.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/decoder.py
@@ -73,6 +73,9 @@ class Decoder(nn.Module):
                 groups=decoder_dim,
                 bias=False,
             )
+        else:
+            # It is to support torch script
+            self.conv = nn.Identity()
 
     def forward(self, y: torch.Tensor, need_pad: bool = True) -> torch.Tensor:
         """
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/conformer.py b/egs/librispeech/ASR/pruned_transducer_stateless5/conformer.py
index 6f7231f4b..bf3917df0 100644
--- a/egs/librispeech/ASR/pruned_transducer_stateless5/conformer.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/conformer.py
@@ -117,10 +117,7 @@ class Conformer(EncoderInterface):
         x, pos_emb = self.encoder_pos(x)
         x = x.permute(1, 0, 2)  # (N, T, C) -> (T, N, C)
 
-        with warnings.catch_warnings():
-            warnings.simplefilter("ignore")
-            # Caution: We assume the subsampling factor is 4!
-            lengths = ((x_lens - 1) // 2 - 1) // 2
+        lengths = (((x_lens - 1) >> 1) - 1) >> 1
         assert x.size(0) == lengths.max().item()
         mask = make_pad_mask(lengths)
 
@@ -293,8 +290,10 @@ class ConformerEncoder(nn.Module):
         )
         self.num_layers = num_layers
 
+        assert len(set(aux_layers)) == len(aux_layers)
+
         assert num_layers - 1 not in aux_layers
-        self.aux_layers = set(aux_layers + [num_layers - 1])
+        self.aux_layers = aux_layers + [num_layers - 1]
 
         num_channels = encoder_layer.norm_final.num_channels
         self.combiner = RandomCombine(
@@ -1154,7 +1153,7 @@ class RandomCombine(nn.Module):
         """
         num_inputs = self.num_inputs
         assert len(inputs) == num_inputs
-        if not self.training:
+        if not self.training or torch.jit.is_scripting():
             return inputs[-1]
 
         # Shape of weights: (*, num_inputs)
@@ -1162,8 +1161,22 @@ class RandomCombine(nn.Module):
         num_frames = inputs[0].numel() // num_channels
 
         mod_inputs = []
-        for i in range(num_inputs - 1):
-            mod_inputs.append(self.linear[i](inputs[i]))
+
+        if False:
+            # It throws the following error for torch 1.6.0 when using
+            # torch script.
+            #
+            # Expected integer literal for index. ModuleList/Sequential
+            # indexing is only supported with integer literals. Enumeration is
+            # supported, e.g. 'for index, v in enumerate(self): ...':
+            #  for i in range(num_inputs - 1):
+            #      mod_inputs.append(self.linear[i](inputs[i]))
+            assert False
+        else:
+            for i, linear in enumerate(self.linear):
+                if i < num_inputs - 1:
+                    mod_inputs.append(linear(inputs[i]))
+
         mod_inputs.append(inputs[num_inputs - 1])
 
         ndim = inputs[0].ndim
@@ -1181,11 +1194,13 @@ class RandomCombine(nn.Module):
         # ans: (num_frames, num_channels, 1)
         ans = torch.matmul(stacked_inputs, weights)
         # ans: (*, num_channels)
-        ans = ans.reshape(*tuple(inputs[0].shape[:-1]), num_channels)
 
-        if __name__ == "__main__":
-            # for testing only...
-            print("Weights = ", weights.reshape(num_frames, num_inputs))
+        ans = ans.reshape(inputs[0].shape[:-1] + [num_channels])
+
+        # The following if causes errors for torch script in torch 1.6.0
+        #  if __name__ == "__main__":
+        #      # for testing only...
+        #      print("Weights = ", weights.reshape(num_frames, num_inputs))
         return ans
 
     def _get_random_weights(
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/export.py b/egs/librispeech/ASR/pruned_transducer_stateless5/export.py
index f1269a4bd..936508900 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless5/export.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/export.py
@@ -146,8 +146,6 @@ def main():
     args = get_parser().parse_args()
     args.exp_dir = Path(args.exp_dir)
 
-    assert args.jit is False, "Support torchscript will be added later"
-
     params = get_params()
     params.update(vars(args))
 
@@ -246,12 +244,15 @@ def main():
                 )
             )
 
-    model.eval()
-
     model.to("cpu")
     model.eval()
 
     if params.jit:
+        # We won't use the forward() method of the model in C++, so just ignore
+        # it here.
+        # Otherwise, one of its arguments is a ragged tensor and is not
+        # torch scriptabe.
+        model.__class__.forward = torch.jit.ignore(model.__class__.forward)
         logging.info("Using torch.jit.script")
         model = torch.jit.script(model)
         filename = params.exp_dir / "cpu_jit.pt"

From dc89b61b808a22b64e7e38cb09771ea7a6bb64d2 Mon Sep 17 00:00:00 2001
From: Fangjun Kuang <csukuangfj@gmail.com>
Date: Wed, 22 Jun 2022 00:09:25 +0800
Subject: [PATCH 6/9] Add fast_beam_search_nbest. (#420)

* Add fast_beam_search_nbest.

* Fix CI errors.

* Fix CI errors.

* More fixes.

* Small fixes.

* Support using log_add in LG decoding with fast_beam_search.

* Support LG decoding in pruned_transducer_stateless

* Support LG for pruned_transducer_stateless2.

* Support LG for fast beam search.

* Minor fixes.
---
 ...pruned-transducer-stateless5-2022-05-13.sh |   6 +
 .github/workflows/test.yml                    |  12 +-
 .../beam_search.py                            | 196 ++++++++++++++
 .../ASR/pruned_transducer_stateless/decode.py | 204 ++++++++++-----
 .../beam_search.py                            | 200 +++++++++++++-
 .../pruned_transducer_stateless2/decode.py    | 202 ++++++++++++--
 .../pruned_transducer_stateless3/decode.py    | 246 +++++++++++++-----
 .../pruned_transducer_stateless4/decode.py    | 210 +++++++++++++--
 .../pruned_transducer_stateless5/decode.py    | 210 +++++++++++++--
 icefall/decode.py                             |   6 +-
 10 files changed, 1298 insertions(+), 194 deletions(-)

diff --git a/.github/scripts/run-librispeech-pruned-transducer-stateless5-2022-05-13.sh b/.github/scripts/run-librispeech-pruned-transducer-stateless5-2022-05-13.sh
index 3d0c4e2ef..61210ac6e 100755
--- a/.github/scripts/run-librispeech-pruned-transducer-stateless5-2022-05-13.sh
+++ b/.github/scripts/run-librispeech-pruned-transducer-stateless5-2022-05-13.sh
@@ -32,6 +32,12 @@ for sym in 1 2 3; do
     --max-sym-per-frame $sym \
     --checkpoint $repo/exp/pretrained.pt \
     --bpe-model $repo/data/lang_bpe_500/bpe.model \
+    --num-encoder-layers 18 \
+    --dim-feedforward 2048 \
+    --nhead 8 \
+    --encoder-dim 512 \
+    --decoder-dim 512 \
+    --joiner-dim 512
     $repo/test_wavs/1089-134686-0001.wav \
     $repo/test_wavs/1221-135766-0001.wav \
     $repo/test_wavs/1221-135766-0002.wav
diff --git a/.github/workflows/test.yml b/.github/workflows/test.yml
index f9dab7afe..1583926ec 100644
--- a/.github/workflows/test.yml
+++ b/.github/workflows/test.yml
@@ -33,13 +33,13 @@ jobs:
         # disable macOS test for now.
         os: [ubuntu-18.04]
         python-version: [3.7, 3.8]
-        torch: ["1.8.0", "1.10.0"]
-        torchaudio: ["0.8.0", "0.10.0"]
-        k2-version: ["1.9.dev20211101"]
+        torch: ["1.8.0", "1.11.0"]
+        torchaudio: ["0.8.0", "0.11.0"]
+        k2-version: ["1.15.1.dev20220427"]
         exclude:
           - torch: "1.8.0"
-            torchaudio: "0.10.0"
-          - torch: "1.10.0"
+            torchaudio: "0.11.0"
+          - torch: "1.11.0"
             torchaudio: "0.8.0"
 
       fail-fast: false
@@ -67,7 +67,7 @@ jobs:
           # numpy 1.20.x does not support python 3.6
           pip install numpy==1.19
           pip install torch==${{ matrix.torch }}+cpu -f https://download.pytorch.org/whl/cpu/torch_stable.html
-          if [[ ${{ matrix.torchaudio }} == "0.10.0" ]]; then
+          if [[ ${{ matrix.torchaudio }} == "0.11.0" ]]; then
             pip install torchaudio==${{ matrix.torchaudio }}+cpu -f https://download.pytorch.org/whl/cpu/torch_stable.html
           else
             pip install torchaudio==${{ matrix.torchaudio }}
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless/beam_search.py
index db23fd993..40c442e7a 100644
--- a/egs/librispeech/ASR/pruned_transducer_stateless/beam_search.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/beam_search.py
@@ -75,6 +75,202 @@ def fast_beam_search_one_best(
     return hyps
 
 
+def fast_beam_search_nbest_LG(
+    model: Transducer,
+    decoding_graph: k2.Fsa,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+    num_paths: int,
+    nbest_scale: float = 0.5,
+    use_double_scores: bool = True,
+) -> List[List[int]]:
+    """It limits the maximum number of symbols per frame to 1.
+
+    The process to get the results is:
+     - (1) Use fast beam search to get a lattice
+     - (2) Select `num_paths` paths from the lattice using k2.random_paths()
+     - (3) Unique the selected paths
+     - (4) Intersect the selected paths with the lattice and compute the
+           shortest path from the intersection result
+     - (5) The path with the largest score is used as the decoding output.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      decoding_graph:
+        Decoding graph used for decoding, may be a TrivialGraph or a HLG.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      encoder_out_lens:
+        A tensor of shape (N,) containing the number of frames in `encoder_out`
+        before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+      num_paths:
+        Number of paths to extract from the decoded lattice.
+      nbest_scale:
+        It's the scale applied to the lattice.scores. A smaller value
+        yields more unique paths.
+      use_double_scores:
+        True to use double precision for computation. False to use
+        single precision.
+    Returns:
+      Return the decoded result.
+    """
+    lattice = fast_beam_search(
+        model=model,
+        decoding_graph=decoding_graph,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+        beam=beam,
+        max_states=max_states,
+        max_contexts=max_contexts,
+    )
+
+    nbest = Nbest.from_lattice(
+        lattice=lattice,
+        num_paths=num_paths,
+        use_double_scores=use_double_scores,
+        nbest_scale=nbest_scale,
+    )
+
+    # The following code is modified from nbest.intersect()
+    word_fsa = k2.invert(nbest.fsa)
+    if hasattr(lattice, "aux_labels"):
+        # delete token IDs as it is not needed
+        del word_fsa.aux_labels
+    word_fsa.scores.zero_()
+    word_fsa_with_epsilon_loops = k2.linear_fsa_with_self_loops(word_fsa)
+    path_to_utt_map = nbest.shape.row_ids(1)
+
+    if hasattr(lattice, "aux_labels"):
+        # lattice has token IDs as labels and word IDs as aux_labels.
+        # inv_lattice has word IDs as labels and token IDs as aux_labels
+        inv_lattice = k2.invert(lattice)
+        inv_lattice = k2.arc_sort(inv_lattice)
+    else:
+        inv_lattice = k2.arc_sort(lattice)
+
+    if inv_lattice.shape[0] == 1:
+        path_lattice = k2.intersect_device(
+            inv_lattice,
+            word_fsa_with_epsilon_loops,
+            b_to_a_map=torch.zeros_like(path_to_utt_map),
+            sorted_match_a=True,
+        )
+    else:
+        path_lattice = k2.intersect_device(
+            inv_lattice,
+            word_fsa_with_epsilon_loops,
+            b_to_a_map=path_to_utt_map,
+            sorted_match_a=True,
+        )
+
+    # path_lattice has word IDs as labels and token IDs as aux_labels
+    path_lattice = k2.top_sort(k2.connect(path_lattice))
+    tot_scores = path_lattice.get_tot_scores(
+        use_double_scores=use_double_scores,
+        log_semiring=True,  # Note: we always use True
+    )
+    # See https://github.com/k2-fsa/icefall/pull/420 for why
+    # we always use log_semiring=True
+
+    ragged_tot_scores = k2.RaggedTensor(nbest.shape, tot_scores)
+    best_hyp_indexes = ragged_tot_scores.argmax()
+    best_path = k2.index_fsa(nbest.fsa, best_hyp_indexes)
+
+    hyps = get_texts(best_path)
+
+    return hyps
+
+
+def fast_beam_search_nbest(
+    model: Transducer,
+    decoding_graph: k2.Fsa,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+    num_paths: int,
+    nbest_scale: float = 0.5,
+    use_double_scores: bool = True,
+) -> List[List[int]]:
+    """It limits the maximum number of symbols per frame to 1.
+
+    The process to get the results is:
+     - (1) Use fast beam search to get a lattice
+     - (2) Select `num_paths` paths from the lattice using k2.random_paths()
+     - (3) Unique the selected paths
+     - (4) Intersect the selected paths with the lattice and compute the
+           shortest path from the intersection result
+     - (5) The path with the largest score is used as the decoding output.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      decoding_graph:
+        Decoding graph used for decoding, may be a TrivialGraph or a HLG.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      encoder_out_lens:
+        A tensor of shape (N,) containing the number of frames in `encoder_out`
+        before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+      num_paths:
+        Number of paths to extract from the decoded lattice.
+      nbest_scale:
+        It's the scale applied to the lattice.scores. A smaller value
+        yields more unique paths.
+      use_double_scores:
+        True to use double precision for computation. False to use
+        single precision.
+    Returns:
+      Return the decoded result.
+    """
+    lattice = fast_beam_search(
+        model=model,
+        decoding_graph=decoding_graph,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+        beam=beam,
+        max_states=max_states,
+        max_contexts=max_contexts,
+    )
+
+    nbest = Nbest.from_lattice(
+        lattice=lattice,
+        num_paths=num_paths,
+        use_double_scores=use_double_scores,
+        nbest_scale=nbest_scale,
+    )
+
+    # at this point, nbest.fsa.scores are all zeros.
+
+    nbest = nbest.intersect(lattice)
+    # Now nbest.fsa.scores contains acoustic scores
+
+    max_indexes = nbest.tot_scores().argmax()
+
+    best_path = k2.index_fsa(nbest.fsa, max_indexes)
+
+    hyps = get_texts(best_path)
+
+    return hyps
+
+
 def fast_beam_search_nbest_oracle(
     model: Transducer,
     decoding_graph: k2.Fsa,
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless/decode.py
index ea43836bd..f39cc614c 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless/decode.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless/decode.py
@@ -50,20 +50,44 @@ Usage:
     --exp-dir ./pruned_transducer_stateless/exp \
     --max-duration 600 \
     --decoding-method fast_beam_search \
-    --beam 4 \
-    --max-contexts 4 \
-    --max-states 8
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
 
-(5) fast beam search using LG
+(5) fast beam search (nbest)
 ./pruned_transducer_stateless/decode.py \
     --epoch 28 \
     --avg 15 \
     --exp-dir ./pruned_transducer_stateless/exp \
-    --use-LG True \
-    --use-max False \
     --max-duration 600 \
-    --decoding-method fast_beam_search \
-    --beam 8 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
     --max-contexts 8 \
     --max-states 64
 """
@@ -82,6 +106,9 @@ import torch.nn as nn
 from asr_datamodule import LibriSpeechAsrDataModule
 from beam_search import (
     beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
     fast_beam_search_one_best,
     greedy_search,
     greedy_search_batch,
@@ -99,7 +126,6 @@ from icefall.utils import (
     AttributeDict,
     setup_logger,
     store_transcripts,
-    str2bool,
     write_error_stats,
 )
 
@@ -153,7 +179,7 @@ def get_parser():
 
     parser.add_argument(
         "--lang-dir",
-        type=str,
+        type=Path,
         default="data/lang_bpe_500",
         help="The lang dir containing word table and LG graph",
     )
@@ -167,6 +193,11 @@ def get_parser():
           - beam_search
           - modified_beam_search
           - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
         """,
     )
 
@@ -182,30 +213,13 @@ def get_parser():
     parser.add_argument(
         "--beam",
         type=float,
-        default=4,
+        default=20.0,
         help="""A floating point value to calculate the cutoff score during beam
         search (i.e., `cutoff = max-score - beam`), which is the same as the
         `beam` in Kaldi.
-        Used only when --decoding-method is fast_beam_search""",
-    )
-
-    parser.add_argument(
-        "--use-LG",
-        type=str2bool,
-        default=False,
-        help="""Whether to use an LG graph for FSA-based beam search.
-        Used only when --decoding_method is fast_beam_search. If setting true,
-        it assumes there is an LG.pt file in lang_dir.""",
-    )
-
-    parser.add_argument(
-        "--use-max",
-        type=str2bool,
-        default=False,
-        help="""If True, use max-op to select the hypothesis that have the
-        max log_prob in case of duplicate hypotheses.
-        If False, use log_add.
-        Used only for beam_search, modified_beam_search, and fast_beam_search
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
         """,
     )
 
@@ -214,7 +228,7 @@ def get_parser():
         type=float,
         default=0.01,
         help="""
-        Used only when --decoding_method is fast_beam_search.
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
         It specifies the scale for n-gram LM scores.
         """,
     )
@@ -222,9 +236,10 @@ def get_parser():
     parser.add_argument(
         "--max-contexts",
         type=int,
-        default=4,
+        default=8,
         help="""Used only when --decoding-method is
-        fast_beam_search""",
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
     )
 
     parser.add_argument(
@@ -232,7 +247,8 @@ def get_parser():
         type=int,
         default=8,
         help="""Used only when --decoding-method is
-        fast_beam_search""",
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
     )
 
     parser.add_argument(
@@ -250,6 +266,24 @@ def get_parser():
         Used only when --decoding_method is greedy_search""",
     )
 
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
     return parser
 
 
@@ -286,7 +320,8 @@ def decode_one_batch(
         The word symbol table.
       decoding_graph:
         The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
-        only when --decoding_method is fast_beam_search.
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
     Returns:
       Return the decoding result. See above description for the format of
       the returned dict.
@@ -299,6 +334,7 @@ def decode_one_batch(
     # at entry, feature is (N, T, C)
 
     supervisions = batch["supervisions"]
+
     feature_lens = supervisions["num_frames"].to(device)
 
     encoder_out, encoder_out_lens = model.encoder(
@@ -316,12 +352,51 @@ def decode_one_batch(
             max_contexts=params.max_contexts,
             max_states=params.max_states,
         )
-        if params.use_LG:
-            for hyp in hyp_tokens:
-                hyps.append([word_table[i] for i in hyp])
-        else:
-            for hyp in sp.decode(hyp_tokens):
-                hyps.append(hyp.split())
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
     elif (
         params.decoding_method == "greedy_search"
         and params.max_sym_per_frame == 1
@@ -339,7 +414,6 @@ def decode_one_batch(
             encoder_out=encoder_out,
             encoder_out_lens=encoder_out_lens,
             beam=params.beam_size,
-            use_max=params.use_max,
         )
         for hyp in sp.decode(hyp_tokens):
             hyps.append(hyp.split())
@@ -361,7 +435,6 @@ def decode_one_batch(
                     model=model,
                     encoder_out=encoder_out_i,
                     beam=params.beam_size,
-                    use_max=params.use_max,
                 )
             else:
                 raise ValueError(
@@ -371,14 +444,17 @@ def decode_one_batch(
 
     if params.decoding_method == "greedy_search":
         return {"greedy_search": hyps}
-    elif params.decoding_method == "fast_beam_search":
-        return {
-            (
-                f"beam_{params.beam}_"
-                f"max_contexts_{params.max_contexts}_"
-                f"max_states_{params.max_states}"
-            ): hyps
-        }
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
     else:
         return {f"beam_size_{params.beam_size}": hyps}
 
@@ -406,7 +482,8 @@ def decode_dataset(
         The word symbol table.
       decoding_graph:
         The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
-        only when --decoding_method is fast_beam_search.
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
     Returns:
       Return a dict, whose key may be "greedy_search" if greedy search
       is used, or it may be "beam_7" if beam size of 7 is used.
@@ -424,7 +501,7 @@ def decode_dataset(
     if params.decoding_method == "greedy_search":
         log_interval = 50
     else:
-        log_interval = 10
+        log_interval = 20
 
     results = defaultdict(list)
     for batch_idx, batch in enumerate(dl):
@@ -517,6 +594,9 @@ def main():
         "greedy_search",
         "beam_search",
         "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
         "modified_beam_search",
     )
     params.res_dir = params.exp_dir / params.decoding_method
@@ -527,16 +607,18 @@ def main():
         params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
 
     if "fast_beam_search" in params.decoding_method:
-        params.suffix += f"-use-LG-{params.use_LG}"
         params.suffix += f"-beam-{params.beam}"
         params.suffix += f"-max-contexts-{params.max_contexts}"
         params.suffix += f"-max-states-{params.max_states}"
-        params.suffix += f"-use-max-{params.use_max}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
     elif "beam_search" in params.decoding_method:
         params.suffix += (
             f"-{params.decoding_method}-beam-size-{params.beam_size}"
         )
-        params.suffix += f"-use-max-{params.use_max}"
     else:
         params.suffix += f"-context-{params.context_size}"
         params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
@@ -596,12 +678,14 @@ def main():
     model.eval()
     model.device = device
 
-    if params.decoding_method == "fast_beam_search":
-        if params.use_LG:
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
             lexicon = Lexicon(params.lang_dir)
             word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
             decoding_graph = k2.Fsa.from_dict(
-                torch.load(f"{params.lang_dir}/LG.pt", map_location=device)
+                torch.load(lg_filename, map_location=device)
             )
             decoding_graph.scores *= params.ngram_lm_scale
         else:
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py
index 7c936b257..6b6190a09 100644
--- a/egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py
@@ -37,7 +37,7 @@ def fast_beam_search_one_best(
 ) -> List[List[int]]:
     """It limits the maximum number of symbols per frame to 1.
 
-    A lattice is first obtained using modified beam search, and then
+    A lattice is first obtained using fast beam search, and then
     the shortest path within the lattice is used as the final output.
 
     Args:
@@ -74,6 +74,202 @@ def fast_beam_search_one_best(
     return hyps
 
 
+def fast_beam_search_nbest_LG(
+    model: Transducer,
+    decoding_graph: k2.Fsa,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+    num_paths: int,
+    nbest_scale: float = 0.5,
+    use_double_scores: bool = True,
+) -> List[List[int]]:
+    """It limits the maximum number of symbols per frame to 1.
+
+    The process to get the results is:
+     - (1) Use fast beam search to get a lattice
+     - (2) Select `num_paths` paths from the lattice using k2.random_paths()
+     - (3) Unique the selected paths
+     - (4) Intersect the selected paths with the lattice and compute the
+           shortest path from the intersection result
+     - (5) The path with the largest score is used as the decoding output.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      decoding_graph:
+        Decoding graph used for decoding, may be a TrivialGraph or a HLG.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      encoder_out_lens:
+        A tensor of shape (N,) containing the number of frames in `encoder_out`
+        before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+      num_paths:
+        Number of paths to extract from the decoded lattice.
+      nbest_scale:
+        It's the scale applied to the lattice.scores. A smaller value
+        yields more unique paths.
+      use_double_scores:
+        True to use double precision for computation. False to use
+        single precision.
+    Returns:
+      Return the decoded result.
+    """
+    lattice = fast_beam_search(
+        model=model,
+        decoding_graph=decoding_graph,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+        beam=beam,
+        max_states=max_states,
+        max_contexts=max_contexts,
+    )
+
+    nbest = Nbest.from_lattice(
+        lattice=lattice,
+        num_paths=num_paths,
+        use_double_scores=use_double_scores,
+        nbest_scale=nbest_scale,
+    )
+
+    # The following code is modified from nbest.intersect()
+    word_fsa = k2.invert(nbest.fsa)
+    if hasattr(lattice, "aux_labels"):
+        # delete token IDs as it is not needed
+        del word_fsa.aux_labels
+    word_fsa.scores.zero_()
+    word_fsa_with_epsilon_loops = k2.linear_fsa_with_self_loops(word_fsa)
+    path_to_utt_map = nbest.shape.row_ids(1)
+
+    if hasattr(lattice, "aux_labels"):
+        # lattice has token IDs as labels and word IDs as aux_labels.
+        # inv_lattice has word IDs as labels and token IDs as aux_labels
+        inv_lattice = k2.invert(lattice)
+        inv_lattice = k2.arc_sort(inv_lattice)
+    else:
+        inv_lattice = k2.arc_sort(lattice)
+
+    if inv_lattice.shape[0] == 1:
+        path_lattice = k2.intersect_device(
+            inv_lattice,
+            word_fsa_with_epsilon_loops,
+            b_to_a_map=torch.zeros_like(path_to_utt_map),
+            sorted_match_a=True,
+        )
+    else:
+        path_lattice = k2.intersect_device(
+            inv_lattice,
+            word_fsa_with_epsilon_loops,
+            b_to_a_map=path_to_utt_map,
+            sorted_match_a=True,
+        )
+
+    # path_lattice has word IDs as labels and token IDs as aux_labels
+    path_lattice = k2.top_sort(k2.connect(path_lattice))
+    tot_scores = path_lattice.get_tot_scores(
+        use_double_scores=use_double_scores,
+        log_semiring=True,  # Note: we always use True
+    )
+    # See https://github.com/k2-fsa/icefall/pull/420 for why
+    # we always use log_semiring=True
+
+    ragged_tot_scores = k2.RaggedTensor(nbest.shape, tot_scores)
+    best_hyp_indexes = ragged_tot_scores.argmax()
+    best_path = k2.index_fsa(nbest.fsa, best_hyp_indexes)
+
+    hyps = get_texts(best_path)
+
+    return hyps
+
+
+def fast_beam_search_nbest(
+    model: Transducer,
+    decoding_graph: k2.Fsa,
+    encoder_out: torch.Tensor,
+    encoder_out_lens: torch.Tensor,
+    beam: float,
+    max_states: int,
+    max_contexts: int,
+    num_paths: int,
+    nbest_scale: float = 0.5,
+    use_double_scores: bool = True,
+) -> List[List[int]]:
+    """It limits the maximum number of symbols per frame to 1.
+
+    The process to get the results is:
+     - (1) Use fast beam search to get a lattice
+     - (2) Select `num_paths` paths from the lattice using k2.random_paths()
+     - (3) Unique the selected paths
+     - (4) Intersect the selected paths with the lattice and compute the
+           shortest path from the intersection result
+     - (5) The path with the largest score is used as the decoding output.
+
+    Args:
+      model:
+        An instance of `Transducer`.
+      decoding_graph:
+        Decoding graph used for decoding, may be a TrivialGraph or a HLG.
+      encoder_out:
+        A tensor of shape (N, T, C) from the encoder.
+      encoder_out_lens:
+        A tensor of shape (N,) containing the number of frames in `encoder_out`
+        before padding.
+      beam:
+        Beam value, similar to the beam used in Kaldi..
+      max_states:
+        Max states per stream per frame.
+      max_contexts:
+        Max contexts pre stream per frame.
+      num_paths:
+        Number of paths to extract from the decoded lattice.
+      nbest_scale:
+        It's the scale applied to the lattice.scores. A smaller value
+        yields more unique paths.
+      use_double_scores:
+        True to use double precision for computation. False to use
+        single precision.
+    Returns:
+      Return the decoded result.
+    """
+    lattice = fast_beam_search(
+        model=model,
+        decoding_graph=decoding_graph,
+        encoder_out=encoder_out,
+        encoder_out_lens=encoder_out_lens,
+        beam=beam,
+        max_states=max_states,
+        max_contexts=max_contexts,
+    )
+
+    nbest = Nbest.from_lattice(
+        lattice=lattice,
+        num_paths=num_paths,
+        use_double_scores=use_double_scores,
+        nbest_scale=nbest_scale,
+    )
+
+    # at this point, nbest.fsa.scores are all zeros.
+
+    nbest = nbest.intersect(lattice)
+    # Now nbest.fsa.scores contains acoustic scores
+
+    max_indexes = nbest.tot_scores().argmax()
+
+    best_path = k2.index_fsa(nbest.fsa, max_indexes)
+
+    hyps = get_texts(best_path)
+
+    return hyps
+
+
 def fast_beam_search_nbest_oracle(
     model: Transducer,
     decoding_graph: k2.Fsa,
@@ -89,7 +285,7 @@ def fast_beam_search_nbest_oracle(
 ) -> List[List[int]]:
     """It limits the maximum number of symbols per frame to 1.
 
-    A lattice is first obtained using modified beam search, and then
+    A lattice is first obtained using fast beam search, and then
     we select `num_paths` linear paths from the lattice. The path
     that has the minimum edit distance with the given reference transcript
     is used as the output.
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless2/decode.py
index d7d6b1202..ea368fb87 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless2/decode.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless2/decode.py
@@ -43,16 +43,53 @@ Usage:
     --decoding-method modified_beam_search \
     --beam-size 4
 
-(4) fast beam search
+(4) fast beam search (one best)
 ./pruned_transducer_stateless2/decode.py \
     --epoch 28 \
     --avg 15 \
     --exp-dir ./pruned_transducer_stateless2/exp \
     --max-duration 600 \
     --decoding-method fast_beam_search \
-    --beam 4 \
-    --max-contexts 4 \
-    --max-states 8
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless2/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless2/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless2/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless2/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
 """
 
 
@@ -69,6 +106,9 @@ import torch.nn as nn
 from asr_datamodule import LibriSpeechAsrDataModule
 from beam_search import (
     beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
     fast_beam_search_one_best,
     greedy_search,
     greedy_search_batch,
@@ -81,6 +121,7 @@ from icefall.checkpoint import (
     find_checkpoints,
     load_checkpoint,
 )
+from icefall.lexicon import Lexicon
 from icefall.utils import (
     AttributeDict,
     setup_logger,
@@ -136,6 +177,13 @@ def get_parser():
         help="Path to the BPE model",
     )
 
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
     parser.add_argument(
         "--decoding-method",
         type=str,
@@ -145,6 +193,11 @@ def get_parser():
           - beam_search
           - modified_beam_search
           - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
         """,
     )
 
@@ -160,27 +213,42 @@ def get_parser():
     parser.add_argument(
         "--beam",
         type=float,
-        default=4,
+        default=20.0,
         help="""A floating point value to calculate the cutoff score during beam
         search (i.e., `cutoff = max-score - beam`), which is the same as the
         `beam` in Kaldi.
-        Used only when --decoding-method is fast_beam_search""",
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
     )
 
     parser.add_argument(
         "--max-contexts",
         type=int,
-        default=4,
+        default=8,
         help="""Used only when --decoding-method is
-        fast_beam_search""",
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
     )
 
     parser.add_argument(
         "--max-states",
         type=int,
-        default=8,
+        default=64,
         help="""Used only when --decoding-method is
-        fast_beam_search""",
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
     )
 
     parser.add_argument(
@@ -198,6 +266,24 @@ def get_parser():
         Used only when --decoding_method is greedy_search""",
     )
 
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
     return parser
 
 
@@ -206,6 +292,7 @@ def decode_one_batch(
     model: nn.Module,
     sp: spm.SentencePieceProcessor,
     batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
     decoding_graph: Optional[k2.Fsa] = None,
 ) -> Dict[str, List[List[str]]]:
     """Decode one batch and return the result in a dict. The dict has the
@@ -229,9 +316,12 @@ def decode_one_batch(
         It is the return value from iterating
         `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
         for the format of the `batch`.
+      word_table:
+        The word symbol table.
       decoding_graph:
         The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
-        only when --decoding_method is fast_beam_search.
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
     Returns:
       Return the decoding result. See above description for the format of
       the returned dict.
@@ -263,6 +353,49 @@ def decode_one_batch(
         )
         for hyp in sp.decode(hyp_tokens):
             hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
     elif (
         params.decoding_method == "greedy_search"
         and params.max_sym_per_frame == 1
@@ -318,6 +451,17 @@ def decode_one_batch(
                 f"max_states_{params.max_states}"
             ): hyps
         }
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
     else:
         return {f"beam_size_{params.beam_size}": hyps}
 
@@ -327,6 +471,7 @@ def decode_dataset(
     params: AttributeDict,
     model: nn.Module,
     sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
     decoding_graph: Optional[k2.Fsa] = None,
 ) -> Dict[str, List[Tuple[List[str], List[str]]]]:
     """Decode dataset.
@@ -340,9 +485,12 @@ def decode_dataset(
         The neural model.
       sp:
         The BPE model.
+      word_table:
+        The word symbol table.
       decoding_graph:
         The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
-        only when --decoding_method is fast_beam_search.
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
     Returns:
       Return a dict, whose key may be "greedy_search" if greedy search
       is used, or it may be "beam_7" if beam size of 7 is used.
@@ -360,7 +508,7 @@ def decode_dataset(
     if params.decoding_method == "greedy_search":
         log_interval = 50
     else:
-        log_interval = 10
+        log_interval = 20
 
     results = defaultdict(list)
     for batch_idx, batch in enumerate(dl):
@@ -370,6 +518,7 @@ def decode_dataset(
             params=params,
             model=model,
             sp=sp,
+            word_table=word_table,
             decoding_graph=decoding_graph,
             batch=batch,
         )
@@ -452,6 +601,9 @@ def main():
         "greedy_search",
         "beam_search",
         "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
         "modified_beam_search",
     )
     params.res_dir = params.exp_dir / params.decoding_method
@@ -465,6 +617,11 @@ def main():
         params.suffix += f"-beam-{params.beam}"
         params.suffix += f"-max-contexts-{params.max_contexts}"
         params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
     elif "beam_search" in params.decoding_method:
         params.suffix += (
             f"-{params.decoding_method}-beam-size-{params.beam_size}"
@@ -528,10 +685,24 @@ def main():
     model.eval()
     model.device = device
 
-    if params.decoding_method == "fast_beam_search":
-        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(
+                params.vocab_size - 1, device=device
+            )
     else:
         decoding_graph = None
+        word_table = None
 
     num_param = sum([p.numel() for p in model.parameters()])
     logging.info(f"Number of model parameters: {num_param}")
@@ -553,6 +724,7 @@ def main():
             params=params,
             model=model,
             sp=sp,
+            word_table=word_table,
             decoding_graph=decoding_graph,
         )
 
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless3/decode.py
index 5b3dce853..8b1ddc930 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless3/decode.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless3/decode.py
@@ -19,40 +19,77 @@
 Usage:
 (1) greedy search
 ./pruned_transducer_stateless3/decode.py \
-        --epoch 28 \
-        --avg 15 \
-        --exp-dir ./pruned_transducer_stateless3/exp \
-        --max-duration 600 \
-        --decoding-method greedy_search
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
 
 (2) beam search (not recommended)
 ./pruned_transducer_stateless3/decode.py \
-        --epoch 28 \
-        --avg 15 \
-        --exp-dir ./pruned_transducer_stateless3/exp \
-        --max-duration 600 \
-        --decoding-method beam_search \
-        --beam-size 4
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
 
 (3) modified beam search
 ./pruned_transducer_stateless3/decode.py \
-        --epoch 28 \
-        --avg 15 \
-        --exp-dir ./pruned_transducer_stateless3/exp \
-        --max-duration 600 \
-        --decoding-method modified_beam_search \
-        --beam-size 4
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
 
-(4) fast beam search
+(4) fast beam search (one best)
 ./pruned_transducer_stateless3/decode.py \
-        --epoch 28 \
-        --avg 15 \
-        --exp-dir ./pruned_transducer_stateless3/exp \
-        --max-duration 600 \
-        --decoding-method fast_beam_search \
-        --beam 4 \
-        --max-contexts 4 \
-        --max-states 8
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless3/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless3/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless3/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
 """
 
 
@@ -69,6 +106,8 @@ import torch.nn as nn
 from asr_datamodule import AsrDataModule
 from beam_search import (
     beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
     fast_beam_search_nbest_oracle,
     fast_beam_search_one_best,
     greedy_search,
@@ -83,6 +122,7 @@ from icefall.checkpoint import (
     find_checkpoints,
     load_checkpoint,
 )
+from icefall.lexicon import Lexicon
 from icefall.utils import (
     AttributeDict,
     setup_logger,
@@ -138,6 +178,13 @@ def get_parser():
         help="Path to the BPE model",
     )
 
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
     parser.add_argument(
         "--decoding-method",
         type=str,
@@ -147,7 +194,11 @@ def get_parser():
           - beam_search
           - modified_beam_search
           - fast_beam_search
+          - fast_beam_search_nbest
           - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
         """,
     )
 
@@ -163,28 +214,42 @@ def get_parser():
     parser.add_argument(
         "--beam",
         type=float,
-        default=4,
+        default=20.0,
         help="""A floating point value to calculate the cutoff score during beam
         search (i.e., `cutoff = max-score - beam`), which is the same as the
         `beam` in Kaldi.
-        Used only when --decoding-method is
-        fast_beam_search or fast_beam_search_nbest_oracle""",
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
     )
 
     parser.add_argument(
         "--max-contexts",
         type=int,
-        default=4,
+        default=8,
         help="""Used only when --decoding-method is
-        fast_beam_search or fast_beam_search_nbest_oracle""",
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
     )
 
     parser.add_argument(
         "--max-states",
         type=int,
-        default=8,
+        default=64,
         help="""Used only when --decoding-method is
-        fast_beam_search or fast_beam_search_nbest_oracle""",
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
     )
 
     parser.add_argument(
@@ -205,10 +270,10 @@ def get_parser():
     parser.add_argument(
         "--num-paths",
         type=int,
-        default=100,
-        help="""Number of paths for computed nbest oracle WER
-        when the decoding method is fast_beam_search_nbest_oracle.
-        """,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
     )
 
     parser.add_argument(
@@ -216,9 +281,10 @@ def get_parser():
         type=float,
         default=0.5,
         help="""Scale applied to lattice scores when computing nbest paths.
-        Used only when the decoding_method is fast_beam_search_nbest_oracle.
-        """,
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
     )
+
     return parser
 
 
@@ -227,6 +293,7 @@ def decode_one_batch(
     model: nn.Module,
     sp: spm.SentencePieceProcessor,
     batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
     decoding_graph: Optional[k2.Fsa] = None,
 ) -> Dict[str, List[List[str]]]:
     """Decode one batch and return the result in a dict. The dict has the
@@ -250,10 +317,12 @@ def decode_one_batch(
         It is the return value from iterating
         `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
         for the format of the `batch`.
+      word_table:
+        The word symbol table.
       decoding_graph:
         The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
-        only when --decoding_method is
-        fast_beam_search or fast_beam_search_nbest_oracle.
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
     Returns:
       Return the decoding result. See above description for the format of
       the returned dict.
@@ -285,6 +354,34 @@ def decode_one_batch(
         )
         for hyp in sp.decode(hyp_tokens):
             hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
     elif params.decoding_method == "fast_beam_search_nbest_oracle":
         hyp_tokens = fast_beam_search_nbest_oracle(
             model=model,
@@ -355,16 +452,25 @@ def decode_one_batch(
                 f"max_states_{params.max_states}"
             ): hyps
         }
-    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+    elif params.decoding_method == "fast_beam_search":
         return {
             (
                 f"beam_{params.beam}_"
                 f"max_contexts_{params.max_contexts}_"
-                f"max_states_{params.max_states}_"
-                f"num_paths_{params.num_paths}_"
-                f"nbest_scale_{params.nbest_scale}"
+                f"max_states_{params.max_states}"
             ): hyps
         }
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
     else:
         return {f"beam_size_{params.beam_size}": hyps}
 
@@ -374,6 +480,7 @@ def decode_dataset(
     params: AttributeDict,
     model: nn.Module,
     sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
     decoding_graph: Optional[k2.Fsa] = None,
 ) -> Dict[str, List[Tuple[List[str], List[str]]]]:
     """Decode dataset.
@@ -387,9 +494,12 @@ def decode_dataset(
         The neural model.
       sp:
         The BPE model.
+      word_table:
+        The word symbol table.
       decoding_graph:
         The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
-        only when --decoding_method is fast_beam_search.
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
     Returns:
       Return a dict, whose key may be "greedy_search" if greedy search
       is used, or it may be "beam_7" if beam size of 7 is used.
@@ -407,7 +517,7 @@ def decode_dataset(
     if params.decoding_method == "greedy_search":
         log_interval = 50
     else:
-        log_interval = 10
+        log_interval = 20
 
     results = defaultdict(list)
     for batch_idx, batch in enumerate(dl):
@@ -417,6 +527,7 @@ def decode_dataset(
             params=params,
             model=model,
             sp=sp,
+            word_table=word_table,
             decoding_graph=decoding_graph,
             batch=batch,
         )
@@ -499,6 +610,8 @@ def main():
         "greedy_search",
         "beam_search",
         "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
         "fast_beam_search_nbest_oracle",
         "modified_beam_search",
     )
@@ -509,16 +622,15 @@ def main():
     else:
         params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
 
-    if params.decoding_method == "fast_beam_search":
+    if "fast_beam_search" in params.decoding_method:
         params.suffix += f"-beam-{params.beam}"
         params.suffix += f"-max-contexts-{params.max_contexts}"
         params.suffix += f"-max-states-{params.max_states}"
-    elif params.decoding_method == "fast_beam_search_nbest_oracle":
-        params.suffix += f"-beam-{params.beam}"
-        params.suffix += f"-max-contexts-{params.max_contexts}"
-        params.suffix += f"-max-states-{params.max_states}"
-        params.suffix += f"-num-paths-{params.num_paths}"
-        params.suffix += f"-nbest-scale-{params.nbest_scale}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
     elif "beam_search" in params.decoding_method:
         params.suffix += (
             f"-{params.decoding_method}-beam-size-{params.beam_size}"
@@ -539,9 +651,9 @@ def main():
     sp = spm.SentencePieceProcessor()
     sp.load(params.bpe_model)
 
-    # <blk> and <unk> is defined in local/train_bpe_model.py
+    # <blk> and <unk> are defined in local/train_bpe_model.py
     params.blank_id = sp.piece_to_id("<blk>")
-    params.unk_id = sp.unk_id()
+    params.unk_id = sp.piece_to_id("<unk>")
     params.vocab_size = sp.get_piece_size()
 
     logging.info(params)
@@ -583,13 +695,24 @@ def main():
     model.device = device
     model.unk_id = params.unk_id
 
-    if params.decoding_method in (
-        "fast_beam_search",
-        "fast_beam_search_nbest_oracle",
-    ):
-        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(
+                params.vocab_size - 1, device=device
+            )
     else:
         decoding_graph = None
+        word_table = None
 
     num_param = sum([p.numel() for p in model.parameters()])
     logging.info(f"Number of model parameters: {num_param}")
@@ -612,6 +735,7 @@ def main():
             params=params,
             model=model,
             sp=sp,
+            word_table=word_table,
             decoding_graph=decoding_graph,
         )
 
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless4/decode.py
index 70afc3ea3..a8d730ad6 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless4/decode.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless4/decode.py
@@ -44,16 +44,53 @@ Usage:
     --decoding-method modified_beam_search \
     --beam-size 4
 
-(4) fast beam search
+(4) fast beam search (one best)
 ./pruned_transducer_stateless4/decode.py \
     --epoch 30 \
     --avg 15 \
     --exp-dir ./pruned_transducer_stateless4/exp \
     --max-duration 600 \
     --decoding-method fast_beam_search \
-    --beam 4 \
-    --max-contexts 4 \
-    --max-states 8
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless4/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless3/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless4/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless4/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless4/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless4/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
 """
 
 
@@ -70,6 +107,9 @@ import torch.nn as nn
 from asr_datamodule import LibriSpeechAsrDataModule
 from beam_search import (
     beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
     fast_beam_search_one_best,
     greedy_search,
     greedy_search_batch,
@@ -83,6 +123,7 @@ from icefall.checkpoint import (
     find_checkpoints,
     load_checkpoint,
 )
+from icefall.lexicon import Lexicon
 from icefall.utils import (
     AttributeDict,
     setup_logger,
@@ -150,6 +191,13 @@ def get_parser():
         help="Path to the BPE model",
     )
 
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
     parser.add_argument(
         "--decoding-method",
         type=str,
@@ -159,6 +207,11 @@ def get_parser():
           - beam_search
           - modified_beam_search
           - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
         """,
     )
 
@@ -174,27 +227,42 @@ def get_parser():
     parser.add_argument(
         "--beam",
         type=float,
-        default=4,
+        default=20.0,
         help="""A floating point value to calculate the cutoff score during beam
         search (i.e., `cutoff = max-score - beam`), which is the same as the
         `beam` in Kaldi.
-        Used only when --decoding-method is fast_beam_search""",
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
     )
 
     parser.add_argument(
         "--max-contexts",
         type=int,
-        default=4,
+        default=8,
         help="""Used only when --decoding-method is
-        fast_beam_search""",
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
     )
 
     parser.add_argument(
         "--max-states",
         type=int,
-        default=8,
+        default=64,
         help="""Used only when --decoding-method is
-        fast_beam_search""",
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
     )
 
     parser.add_argument(
@@ -212,6 +280,24 @@ def get_parser():
         Used only when --decoding_method is greedy_search""",
     )
 
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
     return parser
 
 
@@ -220,6 +306,7 @@ def decode_one_batch(
     model: nn.Module,
     sp: spm.SentencePieceProcessor,
     batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
     decoding_graph: Optional[k2.Fsa] = None,
 ) -> Dict[str, List[List[str]]]:
     """Decode one batch and return the result in a dict. The dict has the
@@ -243,9 +330,12 @@ def decode_one_batch(
         It is the return value from iterating
         `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
         for the format of the `batch`.
+      word_table:
+        The word symbol table.
       decoding_graph:
         The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
-        only when --decoding_method is fast_beam_search.
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
     Returns:
       Return the decoding result. See above description for the format of
       the returned dict.
@@ -277,6 +367,49 @@ def decode_one_batch(
         )
         for hyp in sp.decode(hyp_tokens):
             hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
     elif (
         params.decoding_method == "greedy_search"
         and params.max_sym_per_frame == 1
@@ -324,14 +457,17 @@ def decode_one_batch(
 
     if params.decoding_method == "greedy_search":
         return {"greedy_search": hyps}
-    elif params.decoding_method == "fast_beam_search":
-        return {
-            (
-                f"beam_{params.beam}_"
-                f"max_contexts_{params.max_contexts}_"
-                f"max_states_{params.max_states}"
-            ): hyps
-        }
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
     else:
         return {f"beam_size_{params.beam_size}": hyps}
 
@@ -341,6 +477,7 @@ def decode_dataset(
     params: AttributeDict,
     model: nn.Module,
     sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
     decoding_graph: Optional[k2.Fsa] = None,
 ) -> Dict[str, List[Tuple[List[str], List[str]]]]:
     """Decode dataset.
@@ -354,9 +491,12 @@ def decode_dataset(
         The neural model.
       sp:
         The BPE model.
+      word_table:
+        The word symbol table.
       decoding_graph:
         The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
-        only when --decoding_method is fast_beam_search.
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
     Returns:
       Return a dict, whose key may be "greedy_search" if greedy search
       is used, or it may be "beam_7" if beam size of 7 is used.
@@ -374,7 +514,7 @@ def decode_dataset(
     if params.decoding_method == "greedy_search":
         log_interval = 50
     else:
-        log_interval = 10
+        log_interval = 20
 
     results = defaultdict(list)
     for batch_idx, batch in enumerate(dl):
@@ -385,6 +525,7 @@ def decode_dataset(
             model=model,
             sp=sp,
             decoding_graph=decoding_graph,
+            word_table=word_table,
             batch=batch,
         )
 
@@ -466,6 +607,9 @@ def main():
         "greedy_search",
         "beam_search",
         "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
         "modified_beam_search",
     )
     params.res_dir = params.exp_dir / params.decoding_method
@@ -479,6 +623,11 @@ def main():
         params.suffix += f"-beam-{params.beam}"
         params.suffix += f"-max-contexts-{params.max_contexts}"
         params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
     elif "beam_search" in params.decoding_method:
         params.suffix += (
             f"-{params.decoding_method}-beam-size-{params.beam_size}"
@@ -592,10 +741,24 @@ def main():
     model.to(device)
     model.eval()
 
-    if params.decoding_method == "fast_beam_search":
-        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(
+                params.vocab_size - 1, device=device
+            )
     else:
         decoding_graph = None
+        word_table = None
 
     num_param = sum([p.numel() for p in model.parameters()])
     logging.info(f"Number of model parameters: {num_param}")
@@ -617,6 +780,7 @@ def main():
             params=params,
             model=model,
             sp=sp,
+            word_table=word_table,
             decoding_graph=decoding_graph,
         )
 
diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless5/decode.py
index c2ca07480..f87d23cc9 100755
--- a/egs/librispeech/ASR/pruned_transducer_stateless5/decode.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/decode.py
@@ -44,16 +44,53 @@ Usage:
     --decoding-method modified_beam_search \
     --beam-size 4
 
-(4) fast beam search
+(4) fast beam search (one best)
 ./pruned_transducer_stateless5/decode.py \
     --epoch 28 \
     --avg 15 \
     --exp-dir ./pruned_transducer_stateless5/exp \
     --max-duration 600 \
     --decoding-method fast_beam_search \
-    --beam 4 \
-    --max-contexts 4 \
-    --max-states 8
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
+
+(5) fast beam search (nbest)
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(6) fast beam search (nbest oracle WER)
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_oracle \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64 \
+    --num-paths 200 \
+    --nbest-scale 0.5
+
+(7) fast beam search (with LG)
+./pruned_transducer_stateless5/decode.py \
+    --epoch 28 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless5/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search_nbest_LG \
+    --beam 20.0 \
+    --max-contexts 8 \
+    --max-states 64
 """
 
 
@@ -70,6 +107,9 @@ import torch.nn as nn
 from asr_datamodule import LibriSpeechAsrDataModule
 from beam_search import (
     beam_search,
+    fast_beam_search_nbest,
+    fast_beam_search_nbest_LG,
+    fast_beam_search_nbest_oracle,
     fast_beam_search_one_best,
     greedy_search,
     greedy_search_batch,
@@ -83,6 +123,7 @@ from icefall.checkpoint import (
     find_checkpoints,
     load_checkpoint,
 )
+from icefall.lexicon import Lexicon
 from icefall.utils import (
     AttributeDict,
     setup_logger,
@@ -128,7 +169,7 @@ def get_parser():
     parser.add_argument(
         "--use-averaged-model",
         type=str2bool,
-        default=False,
+        default=True,
         help="Whether to load averaged model. Currently it only supports "
         "using --epoch. If True, it would decode with the averaged model "
         "over the epoch range from `epoch-avg` (excluded) to `epoch`."
@@ -150,6 +191,13 @@ def get_parser():
         help="Path to the BPE model",
     )
 
+    parser.add_argument(
+        "--lang-dir",
+        type=Path,
+        default="data/lang_bpe_500",
+        help="The lang dir containing word table and LG graph",
+    )
+
     parser.add_argument(
         "--decoding-method",
         type=str,
@@ -159,6 +207,11 @@ def get_parser():
           - beam_search
           - modified_beam_search
           - fast_beam_search
+          - fast_beam_search_nbest
+          - fast_beam_search_nbest_oracle
+          - fast_beam_search_nbest_LG
+        If you use fast_beam_search_nbest_LG, you have to specify
+        `--lang-dir`, which should contain `LG.pt`.
         """,
     )
 
@@ -174,27 +227,42 @@ def get_parser():
     parser.add_argument(
         "--beam",
         type=float,
-        default=4,
+        default=20.0,
         help="""A floating point value to calculate the cutoff score during beam
         search (i.e., `cutoff = max-score - beam`), which is the same as the
         `beam` in Kaldi.
-        Used only when --decoding-method is fast_beam_search""",
+        Used only when --decoding-method is fast_beam_search,
+        fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle
+        """,
+    )
+
+    parser.add_argument(
+        "--ngram-lm-scale",
+        type=float,
+        default=0.01,
+        help="""
+        Used only when --decoding_method is fast_beam_search_nbest_LG.
+        It specifies the scale for n-gram LM scores.
+        """,
     )
 
     parser.add_argument(
         "--max-contexts",
         type=int,
-        default=4,
+        default=8,
         help="""Used only when --decoding-method is
-        fast_beam_search""",
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
     )
 
     parser.add_argument(
         "--max-states",
         type=int,
-        default=8,
+        default=64,
         help="""Used only when --decoding-method is
-        fast_beam_search""",
+        fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
+        and fast_beam_search_nbest_oracle""",
     )
 
     parser.add_argument(
@@ -212,6 +280,24 @@ def get_parser():
         Used only when --decoding_method is greedy_search""",
     )
 
+    parser.add_argument(
+        "--num-paths",
+        type=int,
+        default=200,
+        help="""Number of paths for nbest decoding.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
+    parser.add_argument(
+        "--nbest-scale",
+        type=float,
+        default=0.5,
+        help="""Scale applied to lattice scores when computing nbest paths.
+        Used only when the decoding method is fast_beam_search_nbest,
+        fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
+    )
+
     add_model_arguments(parser)
 
     return parser
@@ -222,6 +308,7 @@ def decode_one_batch(
     model: nn.Module,
     sp: spm.SentencePieceProcessor,
     batch: dict,
+    word_table: Optional[k2.SymbolTable] = None,
     decoding_graph: Optional[k2.Fsa] = None,
 ) -> Dict[str, List[List[str]]]:
     """Decode one batch and return the result in a dict. The dict has the
@@ -245,9 +332,12 @@ def decode_one_batch(
         It is the return value from iterating
         `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
         for the format of the `batch`.
+      word_table:
+        The word symbol table.
       decoding_graph:
         The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
-        only when --decoding_method is fast_beam_search.
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
     Returns:
       Return the decoding result. See above description for the format of
       the returned dict.
@@ -279,6 +369,49 @@ def decode_one_batch(
         )
         for hyp in sp.decode(hyp_tokens):
             hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_LG":
+        hyp_tokens = fast_beam_search_nbest_LG(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in hyp_tokens:
+            hyps.append([word_table[i] for i in hyp])
+    elif params.decoding_method == "fast_beam_search_nbest":
+        hyp_tokens = fast_beam_search_nbest(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    elif params.decoding_method == "fast_beam_search_nbest_oracle":
+        hyp_tokens = fast_beam_search_nbest_oracle(
+            model=model,
+            decoding_graph=decoding_graph,
+            encoder_out=encoder_out,
+            encoder_out_lens=encoder_out_lens,
+            beam=params.beam,
+            max_contexts=params.max_contexts,
+            max_states=params.max_states,
+            num_paths=params.num_paths,
+            ref_texts=sp.encode(supervisions["text"]),
+            nbest_scale=params.nbest_scale,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
     elif (
         params.decoding_method == "greedy_search"
         and params.max_sym_per_frame == 1
@@ -326,14 +459,17 @@ def decode_one_batch(
 
     if params.decoding_method == "greedy_search":
         return {"greedy_search": hyps}
-    elif params.decoding_method == "fast_beam_search":
-        return {
-            (
-                f"beam_{params.beam}_"
-                f"max_contexts_{params.max_contexts}_"
-                f"max_states_{params.max_states}"
-            ): hyps
-        }
+    elif "fast_beam_search" in params.decoding_method:
+        key = f"beam_{params.beam}_"
+        key += f"max_contexts_{params.max_contexts}_"
+        key += f"max_states_{params.max_states}"
+        if "nbest" in params.decoding_method:
+            key += f"_num_paths_{params.num_paths}_"
+            key += f"nbest_scale_{params.nbest_scale}"
+            if "LG" in params.decoding_method:
+                key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
+
+        return {key: hyps}
     else:
         return {f"beam_size_{params.beam_size}": hyps}
 
@@ -343,6 +479,7 @@ def decode_dataset(
     params: AttributeDict,
     model: nn.Module,
     sp: spm.SentencePieceProcessor,
+    word_table: Optional[k2.SymbolTable] = None,
     decoding_graph: Optional[k2.Fsa] = None,
 ) -> Dict[str, List[Tuple[List[str], List[str]]]]:
     """Decode dataset.
@@ -356,9 +493,12 @@ def decode_dataset(
         The neural model.
       sp:
         The BPE model.
+      word_table:
+        The word symbol table.
       decoding_graph:
         The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
-        only when --decoding_method is fast_beam_search.
+        only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
+        fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
     Returns:
       Return a dict, whose key may be "greedy_search" if greedy search
       is used, or it may be "beam_7" if beam size of 7 is used.
@@ -387,6 +527,7 @@ def decode_dataset(
             model=model,
             sp=sp,
             decoding_graph=decoding_graph,
+            word_table=word_table,
             batch=batch,
         )
 
@@ -468,6 +609,9 @@ def main():
         "greedy_search",
         "beam_search",
         "fast_beam_search",
+        "fast_beam_search_nbest",
+        "fast_beam_search_nbest_LG",
+        "fast_beam_search_nbest_oracle",
         "modified_beam_search",
     )
     params.res_dir = params.exp_dir / params.decoding_method
@@ -481,6 +625,11 @@ def main():
         params.suffix += f"-beam-{params.beam}"
         params.suffix += f"-max-contexts-{params.max_contexts}"
         params.suffix += f"-max-states-{params.max_states}"
+        if "nbest" in params.decoding_method:
+            params.suffix += f"-nbest-scale-{params.nbest_scale}"
+            params.suffix += f"-num-paths-{params.num_paths}"
+            if "LG" in params.decoding_method:
+                params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
     elif "beam_search" in params.decoding_method:
         params.suffix += (
             f"-{params.decoding_method}-beam-size-{params.beam_size}"
@@ -594,10 +743,24 @@ def main():
     model.to(device)
     model.eval()
 
-    if params.decoding_method == "fast_beam_search":
-        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    if "fast_beam_search" in params.decoding_method:
+        if params.decoding_method == "fast_beam_search_nbest_LG":
+            lexicon = Lexicon(params.lang_dir)
+            word_table = lexicon.word_table
+            lg_filename = params.lang_dir / "LG.pt"
+            logging.info(f"Loading {lg_filename}")
+            decoding_graph = k2.Fsa.from_dict(
+                torch.load(lg_filename, map_location=device)
+            )
+            decoding_graph.scores *= params.ngram_lm_scale
+        else:
+            word_table = None
+            decoding_graph = k2.trivial_graph(
+                params.vocab_size - 1, device=device
+            )
     else:
         decoding_graph = None
+        word_table = None
 
     num_param = sum([p.numel() for p in model.parameters()])
     logging.info(f"Number of model parameters: {num_param}")
@@ -619,6 +782,7 @@ def main():
             params=params,
             model=model,
             sp=sp,
+            word_table=word_table,
             decoding_graph=decoding_graph,
         )
 
diff --git a/icefall/decode.py b/icefall/decode.py
index 94f3e88ba..3ba899b4e 100644
--- a/icefall/decode.py
+++ b/icefall/decode.py
@@ -308,9 +308,7 @@ class Nbest(object):
             del word_fsa.aux_labels
 
         word_fsa.scores.zero_()
-        word_fsa_with_epsilon_loops = k2.remove_epsilon_and_add_self_loops(
-            word_fsa
-        )
+        word_fsa_with_epsilon_loops = k2.linear_fsa_with_self_loops(word_fsa)
 
         path_to_utt_map = self.shape.row_ids(1)
 
@@ -609,7 +607,7 @@ def rescore_with_n_best_list(
       num_paths:
         Size of nbest list.
       lm_scale_list:
-        A list of float representing LM score scales.
+        A list of floats representing LM score scales.
       nbest_scale:
         Scale to be applied to ``lattice.score`` when sampling paths
         using ``k2.random_paths``.

From 0475d75d15d4fca2c6f7a6ab15c4674a429f460b Mon Sep 17 00:00:00 2001
From: ezerhouni <61225408+ezerhouni@users.noreply.github.com>
Date: Thu, 23 Jun 2022 13:37:03 +0200
Subject: [PATCH 7/9] [Ready to be merged] Add RNN-LM to Conformer-CTC decoding
 (#439)

---
 egs/librispeech/ASR/RESULTS.md                |  45 +-
 egs/librispeech/ASR/conformer_ctc/decode.py   | 152 ++++-
 egs/librispeech/ASR/local/download_lm.py      |   2 +
 .../ASR/local/prepare_lm_training_data.py     | 172 +++++
 .../ASR/local/sort_lm_training_data.py        |   1 +
 egs/librispeech/ASR/local/train_bpe_model.py  |   1 -
 egs/librispeech/ASR/prepare.sh                | 103 ++-
 egs/ptb/LM/README.md                          |  18 +
 egs/ptb/LM/local/prepare_lm_training_data.py  |   1 +
 egs/ptb/LM/local/sort_lm_training_data.py     | 143 ++++
 .../LM/local/test_prepare_lm_training_data.py |  62 ++
 egs/ptb/LM/local/train_bpe_model.py           |   1 +
 egs/ptb/LM/prepare.sh                         | 115 ++++
 egs/ptb/LM/shared                             |   1 +
 icefall/decode.py                             | 187 +++++-
 icefall/dist.py                               |  46 +-
 icefall/rnn_lm/compute_perplexity.py          | 237 +++++++
 icefall/rnn_lm/dataset.py                     | 218 +++++++
 icefall/rnn_lm/export.py                      | 167 +++++
 icefall/rnn_lm/model.py                       | 120 ++++
 icefall/rnn_lm/test_dataset.py                |  71 ++
 icefall/rnn_lm/test_dataset_ddp.py            | 103 +++
 icefall/rnn_lm/test_model.py                  |  69 ++
 icefall/rnn_lm/train.py                       | 617 ++++++++++++++++++
 icefall/utils.py                              |  49 +-
 25 files changed, 2659 insertions(+), 42 deletions(-)
 create mode 100755 egs/librispeech/ASR/local/prepare_lm_training_data.py
 create mode 120000 egs/librispeech/ASR/local/sort_lm_training_data.py
 create mode 100644 egs/ptb/LM/README.md
 create mode 120000 egs/ptb/LM/local/prepare_lm_training_data.py
 create mode 100755 egs/ptb/LM/local/sort_lm_training_data.py
 create mode 100755 egs/ptb/LM/local/test_prepare_lm_training_data.py
 create mode 120000 egs/ptb/LM/local/train_bpe_model.py
 create mode 100755 egs/ptb/LM/prepare.sh
 create mode 120000 egs/ptb/LM/shared
 create mode 100755 icefall/rnn_lm/compute_perplexity.py
 create mode 100644 icefall/rnn_lm/dataset.py
 create mode 100644 icefall/rnn_lm/export.py
 create mode 100644 icefall/rnn_lm/model.py
 create mode 100755 icefall/rnn_lm/test_dataset.py
 create mode 100755 icefall/rnn_lm/test_dataset_ddp.py
 create mode 100755 icefall/rnn_lm/test_model.py
 create mode 100755 icefall/rnn_lm/train.py

diff --git a/egs/librispeech/ASR/RESULTS.md b/egs/librispeech/ASR/RESULTS.md
index 5eb07fae5..3c5027c77 100644
--- a/egs/librispeech/ASR/RESULTS.md
+++ b/egs/librispeech/ASR/RESULTS.md
@@ -1299,17 +1299,18 @@ You can find the tensorboard log at: <https://tensorboard.dev/experiment/D7NQc3x
 
 #### 2021-11-09
 
-The best WER, as of 2021-11-09, for the librispeech test dataset is below
-(using HLG decoding + n-gram LM rescoring + attention decoder rescoring):
+The best WER, as of 2022-06-20, for the librispeech test dataset is below
+(using HLG decoding + n-gram LM rescoring + attention decoder rescoring + rnn lm rescoring):
 
 |     | test-clean | test-other |
 |-----|------------|------------|
-| WER | 2.42       | 5.73       |
+| WER | 2.32       | 5.39       |
 
 Scale values used in n-gram LM rescoring and attention rescoring for the best WERs are:
-| ngram_lm_scale | attention_scale |
-|----------------|-----------------|
-| 2.0            | 2.0             |
+
+| ngram_lm_scale | attention_scale | rnn_lm_scale |
+|----------------|-----------------|--------------|
+| 0.3            | 2.1             | 2.2          |
 
 
 To reproduce the above result, use the following commands for training:
@@ -1330,11 +1331,27 @@ export CUDA_VISIBLE_DEVICES="0,1,2,3"
   --start-epoch 0 \
   --num-epochs 90
 # Note: It trains for 90 epochs, but the best WER is at epoch-77.pt
+
+# Train the RNN-LM
+cd icefall
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+./rnn_lm/train.py \
+  --exp-dir rnn_lm/exp_2048_3_tied \
+  --start-epoch 0 \
+  --world-size 4 \
+  --num-epochs 30 \
+  --use-fp16 1 \
+  --embedding-dim 2048 \
+  --hidden-dim 2048 \
+  --num-layers 3 \
+  --batch-size 500 \
+  --tie-weights true
 ```
 
 and the following command for decoding
 
 ```
+rnn_dir=$(git rev-parse --show-toplevel)/icefall/rnn_lm
 ./conformer_ctc/decode.py \
   --exp-dir conformer_ctc/exp_500_att0.8 \
   --lang-dir data/lang_bpe_500 \
@@ -1344,13 +1361,23 @@ and the following command for decoding
   --num-paths 1000 \
   --epoch 77 \
   --avg 55 \
-  --method attention-decoder \
-  --nbest-scale 0.5
+  --nbest-scale 0.5 \
+  --rnn-lm-exp-dir ${rnn_dir}/exp_2048_3_tied \
+  --rnn-lm-epoch 29 \
+  --rnn-lm-avg 3 \
+  --rnn-lm-embedding-dim 2048 \
+  --rnn-lm-hidden-dim 2048 \
+  --rnn-lm-num-layers 3 \
+  --rnn-lm-tie-weights true \
+  --method rnn-lm
 ```
 
-You can find the pre-trained model by visiting
+You can find the Conformer-CTC pre-trained model by visiting
 <https://huggingface.co/csukuangfj/icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09>
 
+and the RNN-LM pre-trained model:
+<https://huggingface.co/ezerhouni/icefall-librispeech-rnn-lm/tree/main>
+
 The tensorboard log for training is available at
 <https://tensorboard.dev/experiment/hZDWrZfaSqOMqtW0NEfXKg/#scalars>
 
diff --git a/egs/librispeech/ASR/conformer_ctc/decode.py b/egs/librispeech/ASR/conformer_ctc/decode.py
index 177e33a6e..0e8247b8d 100755
--- a/egs/librispeech/ASR/conformer_ctc/decode.py
+++ b/egs/librispeech/ASR/conformer_ctc/decode.py
@@ -30,7 +30,7 @@ from asr_datamodule import LibriSpeechAsrDataModule
 from conformer import Conformer
 
 from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler
-from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.checkpoint import load_checkpoint
 from icefall.decode import (
     get_lattice,
     nbest_decoding,
@@ -38,15 +38,19 @@ from icefall.decode import (
     one_best_decoding,
     rescore_with_attention_decoder,
     rescore_with_n_best_list,
+    rescore_with_rnn_lm,
     rescore_with_whole_lattice,
 )
 from icefall.env import get_env_info
 from icefall.lexicon import Lexicon
+from icefall.rnn_lm.model import RnnLmModel
 from icefall.utils import (
     AttributeDict,
     get_texts,
+    load_averaged_model,
     setup_logger,
     store_transcripts,
+    str2bool,
     write_error_stats,
 )
 
@@ -93,7 +97,9 @@ def get_parser():
               is the decoding result.
             - (5) attention-decoder. Extract n paths from the LM rescored
               lattice, the path with the highest score is the decoding result.
-            - (6) nbest-oracle. Its WER is the lower bound of any n-best
+            - (6) rnn-lm. Rescoring with attention-decoder and RNN LM. We assume
+              you have trained an RNN LM using ./rnn_lm/train.py
+            - (7) nbest-oracle. Its WER is the lower bound of any n-best
               rescoring method can achieve. Useful for debugging n-best
               rescoring method.
         """,
@@ -105,7 +111,7 @@ def get_parser():
         default=100,
         help="""Number of paths for n-best based decoding method.
         Used only when "method" is one of the following values:
-        nbest, nbest-rescoring, attention-decoder, and nbest-oracle
+        nbest, nbest-rescoring, attention-decoder, rnn-lm, and nbest-oracle
         """,
     )
 
@@ -116,7 +122,7 @@ def get_parser():
         help="""The scale to be applied to `lattice.scores`.
         It's needed if you use any kinds of n-best based rescoring.
         Used only when "method" is one of the following values:
-        nbest, nbest-rescoring, attention-decoder, and nbest-oracle
+        nbest, nbest-rescoring, attention-decoder, rnn-lm, and nbest-oracle
         A smaller value results in more unique paths.
         """,
     )
@@ -139,11 +145,67 @@ def get_parser():
         "--lm-dir",
         type=str,
         default="data/lm",
-        help="""The LM dir.
+        help="""The n-gram LM dir.
         It should contain either G_4_gram.pt or G_4_gram.fst.txt
         """,
     )
 
+    parser.add_argument(
+        "--rnn-lm-exp-dir",
+        type=str,
+        default="rnn_lm/exp",
+        help="""Used only when --method is rnn-lm.
+        It specifies the path to RNN LM exp dir.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-epoch",
+        type=int,
+        default=7,
+        help="""Used only when --method is rnn-lm.
+        It specifies the checkpoint to use.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-avg",
+        type=int,
+        default=2,
+        help="""Used only when --method is rnn-lm.
+        It specifies the number of checkpoints to average.
+        """,
+    )
+
+    parser.add_argument(
+        "--rnn-lm-embedding-dim",
+        type=int,
+        default=2048,
+        help="Embedding dim of the model",
+    )
+
+    parser.add_argument(
+        "--rnn-lm-hidden-dim",
+        type=int,
+        default=2048,
+        help="Hidden dim of the model",
+    )
+
+    parser.add_argument(
+        "--rnn-lm-num-layers",
+        type=int,
+        default=4,
+        help="Number of RNN layers the model",
+    )
+    parser.add_argument(
+        "--rnn-lm-tie-weights",
+        type=str2bool,
+        default=False,
+        help="""True to share the weights between the input embedding layer and the
+        last output linear layer
+        """,
+    )
+
     return parser
 
 
@@ -173,6 +235,7 @@ def get_params() -> AttributeDict:
 def decode_one_batch(
     params: AttributeDict,
     model: nn.Module,
+    rnn_lm_model: Optional[nn.Module],
     HLG: Optional[k2.Fsa],
     H: Optional[k2.Fsa],
     bpe_model: Optional[spm.SentencePieceProcessor],
@@ -205,6 +268,8 @@ def decode_one_batch(
 
       model:
         The neural model.
+      rnn_lm_model:
+        The neural model for RNN LM.
       HLG:
         The decoding graph. Used only when params.method is NOT ctc-decoding.
       H:
@@ -330,6 +395,7 @@ def decode_one_batch(
         "nbest-rescoring",
         "whole-lattice-rescoring",
         "attention-decoder",
+        "rnn-lm",
     ]
 
     lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
@@ -357,8 +423,6 @@ def decode_one_batch(
             G_with_epsilon_loops=G,
             lm_scale_list=None,
         )
-        # TODO: pass `lattice` instead of `rescored_lattice` to
-        # `rescore_with_attention_decoder`
 
         best_path_dict = rescore_with_attention_decoder(
             lattice=rescored_lattice,
@@ -370,6 +434,26 @@ def decode_one_batch(
             eos_id=eos_id,
             nbest_scale=params.nbest_scale,
         )
+    elif params.method == "rnn-lm":
+        # lattice uses a 3-gram Lm. We rescore it with a 4-gram LM.
+        rescored_lattice = rescore_with_whole_lattice(
+            lattice=lattice,
+            G_with_epsilon_loops=G,
+            lm_scale_list=None,
+        )
+
+        best_path_dict = rescore_with_rnn_lm(
+            lattice=rescored_lattice,
+            num_paths=params.num_paths,
+            rnn_lm_model=rnn_lm_model,
+            model=model,
+            memory=memory,
+            memory_key_padding_mask=memory_key_padding_mask,
+            sos_id=sos_id,
+            eos_id=eos_id,
+            blank_id=0,
+            nbest_scale=params.nbest_scale,
+        )
     else:
         assert False, f"Unsupported decoding method: {params.method}"
 
@@ -388,6 +472,7 @@ def decode_dataset(
     dl: torch.utils.data.DataLoader,
     params: AttributeDict,
     model: nn.Module,
+    rnn_lm_model: Optional[nn.Module],
     HLG: Optional[k2.Fsa],
     H: Optional[k2.Fsa],
     bpe_model: Optional[spm.SentencePieceProcessor],
@@ -405,6 +490,8 @@ def decode_dataset(
         It is returned by :func:`get_params`.
       model:
         The neural model.
+      rnn_lm_model:
+        The neural model for RNN LM.
       HLG:
         The decoding graph. Used only when params.method is NOT ctc-decoding.
       H:
@@ -442,6 +529,7 @@ def decode_dataset(
         hyps_dict = decode_one_batch(
             params=params,
             model=model,
+            rnn_lm_model=rnn_lm_model,
             HLG=HLG,
             H=H,
             bpe_model=bpe_model,
@@ -490,7 +578,7 @@ def save_results(
     test_set_name: str,
     results_dict: Dict[str, List[Tuple[List[int], List[int]]]],
 ):
-    if params.method == "attention-decoder":
+    if params.method in ("attention-decoder", "rnn-lm"):
         # Set it to False since there are too many logs.
         enable_log = False
     else:
@@ -566,6 +654,10 @@ def main():
     sos_id = graph_compiler.sos_id
     eos_id = graph_compiler.eos_id
 
+    params.num_classes = num_classes
+    params.sos_id = sos_id
+    params.eos_id = eos_id
+
     if params.method == "ctc-decoding":
         HLG = None
         H = k2.ctc_topo(
@@ -590,6 +682,7 @@ def main():
         "nbest-rescoring",
         "whole-lattice-rescoring",
         "attention-decoder",
+        "rnn-lm",
     ):
         if not (params.lm_dir / "G_4_gram.pt").is_file():
             logging.info("Loading G_4_gram.fst.txt")
@@ -621,7 +714,11 @@ def main():
             d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device)
             G = k2.Fsa.from_dict(d)
 
-        if params.method in ["whole-lattice-rescoring", "attention-decoder"]:
+        if params.method in [
+            "whole-lattice-rescoring",
+            "attention-decoder",
+            "rnn-lm",
+        ]:
             # Add epsilon self-loops to G as we will compose
             # it with the whole lattice later
             G = k2.add_epsilon_self_loops(G)
@@ -648,20 +745,40 @@ def main():
     if params.avg == 1:
         load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
     else:
-        start = params.epoch - params.avg + 1
-        filenames = []
-        for i in range(start, params.epoch + 1):
-            if start >= 0:
-                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
-        logging.info(f"averaging {filenames}")
-        model.to(device)
-        model.load_state_dict(average_checkpoints(filenames, device=device))
+        model = load_averaged_model(
+            params.exp_dir, model, params.epoch, params.avg, device
+        )
 
     model.to(device)
     model.eval()
     num_param = sum([p.numel() for p in model.parameters()])
     logging.info(f"Number of model parameters: {num_param}")
 
+    rnn_lm_model = None
+    if params.method == "rnn-lm":
+        rnn_lm_model = RnnLmModel(
+            vocab_size=params.num_classes,
+            embedding_dim=params.rnn_lm_embedding_dim,
+            hidden_dim=params.rnn_lm_hidden_dim,
+            num_layers=params.rnn_lm_num_layers,
+            tie_weights=params.rnn_lm_tie_weights,
+        )
+        if params.rnn_lm_avg == 1:
+            load_checkpoint(
+                f"{params.rnn_lm_exp_dir}/epoch-{params.rnn_lm_epoch}.pt",
+                rnn_lm_model,
+            )
+            rnn_lm_model.to(device)
+        else:
+            rnn_lm_model = load_averaged_model(
+                params.rnn_lm_exp_dir,
+                rnn_lm_model,
+                params.rnn_lm_epoch,
+                params.rnn_lm_avg,
+                device,
+            )
+        rnn_lm_model.eval()
+
     librispeech = LibriSpeechAsrDataModule(args)
 
     test_clean_cuts = librispeech.test_clean_cuts()
@@ -678,6 +795,7 @@ def main():
             dl=test_dl,
             params=params,
             model=model,
+            rnn_lm_model=rnn_lm_model,
             HLG=HLG,
             H=H,
             bpe_model=bpe_model,
diff --git a/egs/librispeech/ASR/local/download_lm.py b/egs/librispeech/ASR/local/download_lm.py
index 94d23afed..030122aa7 100755
--- a/egs/librispeech/ASR/local/download_lm.py
+++ b/egs/librispeech/ASR/local/download_lm.py
@@ -23,6 +23,7 @@ This file downloads the following LibriSpeech LM files:
     - 4-gram.arpa.gz
     - librispeech-vocab.txt
     - librispeech-lexicon.txt
+    - librispeech-lm-norm.txt.gz
 
 from http://www.openslr.org/resources/11
 and save them in the user provided directory.
@@ -61,6 +62,7 @@ def main(out_dir: str):
         "4-gram.arpa.gz",
         "librispeech-vocab.txt",
         "librispeech-lexicon.txt",
+        "librispeech-lm-norm.txt.gz",
     )
 
     for f in tqdm(files_to_download, desc="Downloading LibriSpeech LM files"):
diff --git a/egs/librispeech/ASR/local/prepare_lm_training_data.py b/egs/librispeech/ASR/local/prepare_lm_training_data.py
new file mode 100755
index 000000000..5070341f1
--- /dev/null
+++ b/egs/librispeech/ASR/local/prepare_lm_training_data.py
@@ -0,0 +1,172 @@
+#!/usr/bin/env python3
+
+# Copyright (c)  2021  Xiaomi Corporation (authors: Daniel Povey
+#                                                   Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This script takes a `bpe.model` and a text file such as
+./download/lm/librispeech-lm-norm.txt
+and outputs the LM training data to a supplied directory such
+as data/lm_training_bpe_500.  The format is as follows:
+
+It creates a PyTorch archive (.pt file), say data/lm_training.pt, which is a
+representation of a dict with the following format:
+
+  'words' -> a k2.RaggedTensor of two axes [word][token] with dtype torch.int32
+             containing the BPE representations of each word, indexed by
+             integer word ID. (These integer word IDS are present in
+             'lm_data').  The sentencepiece object can be used to turn the
+             words and BPE units into string form.
+  'sentences' -> a k2.RaggedTensor of two axes [sentence][word] with dtype
+            torch.int32 containing all the sentences, as word-ids (we don't
+            output the string form of this directly but it can be worked out
+            together with 'words' and the bpe.model).
+  'sentence_lengths' -> a 1-D torch.Tensor of dtype torch.int32, containing
+            number of BPE tokens of each sentence.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import sentencepiece as spm
+import torch
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        help="Input BPE model, e.g. data/bpe_500/bpe.model",
+    )
+    parser.add_argument(
+        "--lm-data",
+        type=str,
+        help="""Input LM training data as text, e.g.
+        download/pb.train.txt""",
+    )
+    parser.add_argument(
+        "--lm-archive",
+        type=str,
+        help="""Path to output archive, e.g. data/bpe_500/lm_data.pt;
+        look at the source of this script to see the format.""",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+
+    if Path(args.lm_archive).exists():
+        logging.warning(f"{args.lm_archive} exists - skipping")
+        return
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(args.bpe_model)
+
+    # word2index is a dictionary from words to integer ids.  No need to reserve
+    # space for epsilon, etc.; the words are just used as a convenient way to
+    # compress the sequences of BPE pieces.
+    word2index = dict()
+
+    word2bpe = []  # Will be a list-of-list-of-int, representing BPE pieces.
+    sentences = []  # Will be a list-of-list-of-int, representing word-ids.
+
+    if "librispeech-lm-norm" in args.lm_data:
+        num_lines_in_total = 40418261.0
+        step = 5000000
+    elif "valid" in args.lm_data:
+        num_lines_in_total = 5567.0
+        step = 3000
+    elif "test" in args.lm_data:
+        num_lines_in_total = 5559.0
+        step = 3000
+    else:
+        num_lines_in_total = None
+        step = None
+
+    processed = 0
+
+    with open(args.lm_data) as f:
+        while True:
+            line = f.readline()
+            if line == "":
+                break
+
+            if step and processed % step == 0:
+                logging.info(
+                    f"Processed number of lines: {processed} "
+                    f"({processed/num_lines_in_total*100: .3f}%)"
+                )
+            processed += 1
+
+            line_words = line.split()
+            for w in line_words:
+                if w not in word2index:
+                    w_bpe = sp.encode(w)
+                    word2index[w] = len(word2bpe)
+                    word2bpe.append(w_bpe)
+            sentences.append([word2index[w] for w in line_words])
+
+    logging.info("Constructing ragged tensors")
+    words = k2.ragged.RaggedTensor(word2bpe)
+    sentences = k2.ragged.RaggedTensor(sentences)
+
+    output = dict(words=words, sentences=sentences)
+
+    num_sentences = sentences.dim0
+    logging.info(f"Computing sentence lengths, num_sentences: {num_sentences}")
+    sentence_lengths = [0] * num_sentences
+    for i in range(num_sentences):
+        if step and i % step == 0:
+            logging.info(
+                f"Processed number of lines: {i} "
+                f"({i/num_sentences*100: .3f}%)"
+            )
+
+        word_ids = sentences[i]
+
+        # NOTE: If word_ids is a tensor with only 1 entry,
+        # token_ids is a torch.Tensor
+        token_ids = words[word_ids]
+        if isinstance(token_ids, k2.RaggedTensor):
+            token_ids = token_ids.values
+
+        # token_ids is a 1-D tensor containing the BPE tokens
+        # of the current sentence
+
+        sentence_lengths[i] = token_ids.numel()
+
+    output["sentence_lengths"] = torch.tensor(
+        sentence_lengths, dtype=torch.int32
+    )
+
+    torch.save(output, args.lm_archive)
+    logging.info(f"Saved to {args.lm_archive}")
+
+
+if __name__ == "__main__":
+    formatter = (
+        "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    )
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    main()
diff --git a/egs/librispeech/ASR/local/sort_lm_training_data.py b/egs/librispeech/ASR/local/sort_lm_training_data.py
new file mode 120000
index 000000000..bb26b5f5c
--- /dev/null
+++ b/egs/librispeech/ASR/local/sort_lm_training_data.py
@@ -0,0 +1 @@
+../../../ptb/LM/local/sort_lm_training_data.py
\ No newline at end of file
diff --git a/egs/librispeech/ASR/local/train_bpe_model.py b/egs/librispeech/ASR/local/train_bpe_model.py
index bc5812810..42aba9572 100755
--- a/egs/librispeech/ASR/local/train_bpe_model.py
+++ b/egs/librispeech/ASR/local/train_bpe_model.py
@@ -38,7 +38,6 @@ def get_args():
         "--lang-dir",
         type=str,
         help="""Input and output directory.
-        It should contain the training corpus: transcript_words.txt.
         The generated bpe.model is saved to this directory.
         """,
     )
diff --git a/egs/librispeech/ASR/prepare.sh b/egs/librispeech/ASR/prepare.sh
index 17a638502..94e003036 100755
--- a/egs/librispeech/ASR/prepare.sh
+++ b/egs/librispeech/ASR/prepare.sh
@@ -24,6 +24,7 @@ stop_stage=100
 #        - 4-gram.arpa
 #        - librispeech-vocab.txt
 #        - librispeech-lexicon.txt
+#        - librispeech-lm-norm.txt.gz
 #
 #  - $dl_dir/musan
 #      This directory contains the following directories downloaded from
@@ -40,9 +41,9 @@ dl_dir=$PWD/download
 # It will generate data/lang_bpe_xxx,
 # data/lang_bpe_yyy if the array contains xxx, yyy
 vocab_sizes=(
-  # 5000
-  # 2000
-  # 1000
+  5000
+  2000
+  1000
   500
 )
 
@@ -278,3 +279,99 @@ if [ $stage -le 10 ] && [ $stop_stage -ge 10 ]; then
     ./local/compile_lg.py --lang-dir $lang_dir
   done
 fi
+
+if [ $stage -le 11 ] && [ $stop_stage -ge 11 ]; then
+  log "Stage 11: Generate LM training data"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    log "Processing vocab_size == ${vocab_size}"
+    lang_dir=data/lang_bpe_${vocab_size}
+    out_dir=data/lm_training_bpe_${vocab_size}
+    mkdir -p $out_dir
+
+    ./local/prepare_lm_training_data.py \
+      --bpe-model $lang_dir/bpe.model \
+      --lm-data $dl_dir/lm/librispeech-lm-norm.txt \
+      --lm-archive $out_dir/lm_data.pt
+  done
+fi
+
+if [ $stage -le 12 ] && [ $stop_stage -ge 12 ]; then
+  log "Stage 12: Generate LM validation data"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    log "Processing vocab_size == ${vocab_size}"
+    out_dir=data/lm_training_bpe_${vocab_size}
+    mkdir -p $out_dir
+
+    if [ ! -f $out_dir/valid.txt ]; then
+      files=$(
+        find "$dl_dir/LibriSpeech/dev-clean" -name "*.trans.txt"
+        find "$dl_dir/LibriSpeech/dev-other" -name "*.trans.txt"
+      )
+      for f in ${files[@]}; do
+        cat $f | cut -d " " -f 2-
+      done > $out_dir/valid.txt
+    fi
+
+    lang_dir=data/lang_bpe_${vocab_size}
+    ./local/prepare_lm_training_data.py \
+      --bpe-model $lang_dir/bpe.model \
+      --lm-data $out_dir/valid.txt \
+      --lm-archive $out_dir/lm_data-valid.pt
+  done
+fi
+
+if [ $stage -le 13 ] && [ $stop_stage -ge 13 ]; then
+  log "Stage 13: Generate LM test data"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    log "Processing vocab_size == ${vocab_size}"
+    out_dir=data/lm_training_bpe_${vocab_size}
+    mkdir -p $out_dir
+
+    if [ ! -f $out_dir/test.txt ]; then
+      files=$(
+        find "$dl_dir/LibriSpeech/test-clean" -name "*.trans.txt"
+        find "$dl_dir/LibriSpeech/test-other" -name "*.trans.txt"
+      )
+      for f in ${files[@]}; do
+        cat $f | cut -d " " -f 2-
+      done > $out_dir/test.txt
+    fi
+
+    lang_dir=data/lang_bpe_${vocab_size}
+    ./local/prepare_lm_training_data.py \
+      --bpe-model $lang_dir/bpe.model \
+      --lm-data $out_dir/test.txt \
+      --lm-archive $out_dir/lm_data-test.pt
+  done
+fi
+
+if [ $stage -le 14 ] && [ $stop_stage -ge 14 ]; then
+  log "Stage 14: Sort LM training data"
+  # Sort LM training data by sentence length in descending order
+  # for ease of training.
+  #
+  # Sentence length equals to the number of BPE tokens
+  # in a sentence.
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    out_dir=data/lm_training_bpe_${vocab_size}
+    mkdir -p $out_dir
+    ./local/sort_lm_training_data.py \
+      --in-lm-data $out_dir/lm_data.pt \
+      --out-lm-data $out_dir/sorted_lm_data.pt \
+      --out-statistics $out_dir/statistics.txt
+
+    ./local/sort_lm_training_data.py \
+      --in-lm-data $out_dir/lm_data-valid.pt \
+      --out-lm-data $out_dir/sorted_lm_data-valid.pt \
+      --out-statistics $out_dir/statistics-valid.txt
+
+    ./local/sort_lm_training_data.py \
+      --in-lm-data $out_dir/lm_data-test.pt \
+      --out-lm-data $out_dir/sorted_lm_data-test.pt \
+      --out-statistics $out_dir/statistics-test.txt
+  done
+fi
diff --git a/egs/ptb/LM/README.md b/egs/ptb/LM/README.md
new file mode 100644
index 000000000..7629a950d
--- /dev/null
+++ b/egs/ptb/LM/README.md
@@ -0,0 +1,18 @@
+## Description
+
+(Note: the experiments here are only about language modeling)
+
+ptb is short for Penn Treebank.
+
+
+About the Penn Treebank corpus:
+  - This corpus is free for research purposes
+  - ptb.train.txt: train set
+  - ptb.valid.txt: development set (should be used just for tuning hyper-parameters, but not for training)
+  - ptb.test.txt: test set for reporting perplexity
+
+You can download the dataset from one of the following URLs:
+
+- https://github.com/townie/PTB-dataset-from-Tomas-Mikolov-s-webpage
+- http://www.fit.vutbr.cz/~imikolov/rnnlm/simple-examples.tgz
+- https://deepai.org/dataset/penn-treebank
diff --git a/egs/ptb/LM/local/prepare_lm_training_data.py b/egs/ptb/LM/local/prepare_lm_training_data.py
new file mode 120000
index 000000000..eebce957e
--- /dev/null
+++ b/egs/ptb/LM/local/prepare_lm_training_data.py
@@ -0,0 +1 @@
+../../librispeech/ASR/local/prepare_lm_training_data.py
\ No newline at end of file
diff --git a/egs/ptb/LM/local/sort_lm_training_data.py b/egs/ptb/LM/local/sort_lm_training_data.py
new file mode 100755
index 000000000..af54dbd07
--- /dev/null
+++ b/egs/ptb/LM/local/sort_lm_training_data.py
@@ -0,0 +1,143 @@
+#!/usr/bin/env python3
+
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+This file takes as input the filename of LM training data
+generated by ./local/prepare_lm_training_data.py and sorts
+it by sentence length.
+
+Sentence length equals to the number of BPE tokens in a sentence.
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import k2
+import numpy as np
+import torch
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--in-lm-data",
+        type=str,
+        help="Input LM training data, e.g., data/bpe_500/lm_data.pt",
+    )
+
+    parser.add_argument(
+        "--out-lm-data",
+        type=str,
+        help="Input LM training data, e.g., data/bpe_500/sorted_lm_data.pt",
+    )
+
+    parser.add_argument(
+        "--out-statistics",
+        type=str,
+        help="Statistics about LM training data., data/bpe_500/statistics.txt",
+    )
+
+    return parser.parse_args()
+
+
+def main():
+    args = get_args()
+    in_lm_data = Path(args.in_lm_data)
+    out_lm_data = Path(args.out_lm_data)
+    assert in_lm_data.is_file(), f"{in_lm_data}"
+    if out_lm_data.is_file():
+        logging.warning(f"{out_lm_data} exists - skipping")
+        return
+    data = torch.load(in_lm_data)
+    words2bpe = data["words"]
+    sentences = data["sentences"]
+    sentence_lengths = data["sentence_lengths"]
+
+    num_sentences = sentences.dim0
+    assert num_sentences == sentence_lengths.numel(), (
+        num_sentences,
+        sentence_lengths.numel(),
+    )
+
+    indices = torch.argsort(sentence_lengths, descending=True)
+
+    sorted_sentences = sentences[indices.to(torch.int32)]
+    sorted_sentence_lengths = sentence_lengths[indices]
+
+    # Check that sentences are ordered by length
+    assert num_sentences == sorted_sentences.dim0, (
+        num_sentences,
+        sorted_sentences.dim0,
+    )
+
+    cur = None
+    for i in range(num_sentences):
+        word_ids = sorted_sentences[i]
+        token_ids = words2bpe[word_ids]
+        if isinstance(token_ids, k2.RaggedTensor):
+            token_ids = token_ids.values
+        if cur is not None:
+            assert cur >= token_ids.numel(), (cur, token_ids.numel())
+
+        cur = token_ids.numel()
+        assert cur == sorted_sentence_lengths[i]
+
+    data["sentences"] = sorted_sentences
+    data["sentence_lengths"] = sorted_sentence_lengths
+    torch.save(data, args.out_lm_data)
+    logging.info(f"Saved to {args.out_lm_data}")
+
+    statistics = Path(args.out_statistics)
+
+    # Write statistics
+    num_words = sorted_sentences.numel()
+    num_tokens = sentence_lengths.sum().item()
+    max_sentence_length = sentence_lengths[indices[0]]
+    min_sentence_length = sentence_lengths[indices[-1]]
+
+    step = 10
+    hist, bins = np.histogram(
+        sentence_lengths.numpy(),
+        bins=np.arange(1, max_sentence_length + step, step),
+    )
+
+    histogram = np.stack((bins[:-1], hist)).transpose()
+
+    with open(statistics, "w") as f:
+        f.write(f"num_sentences: {num_sentences}\n")
+        f.write(f"num_words: {num_words}\n")
+        f.write(f"num_tokens: {num_tokens}\n")
+        f.write(f"max_sentence_length: {max_sentence_length}\n")
+        f.write(f"min_sentence_length: {min_sentence_length}\n")
+        f.write("histogram:\n")
+        f.write("  bin  count  percent\n")
+        for row in histogram:
+            f.write(
+                f"{int(row[0]):>5} {int(row[1]):>5}   "
+                f"{100.*row[1]/num_sentences:.3f}%\n"
+            )
+
+
+if __name__ == "__main__":
+    formatter = (
+        "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    )
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/ptb/LM/local/test_prepare_lm_training_data.py b/egs/ptb/LM/local/test_prepare_lm_training_data.py
new file mode 100755
index 000000000..877720e7b
--- /dev/null
+++ b/egs/ptb/LM/local/test_prepare_lm_training_data.py
@@ -0,0 +1,62 @@
+#!/usr/bin/env python3
+
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+
+
+def main():
+    lm_training_data = Path("./data/bpe_500/lm_data.pt")
+    bpe_model = Path("./data/bpe_500/bpe.model")
+    if not lm_training_data.exists():
+        logging.warning(f"{lm_training_data} does not exist - skipping")
+        return
+
+    if not bpe_model.exists():
+        logging.warning(f"{bpe_model} does not exist - skipping")
+        return
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(str(bpe_model))
+
+    data = torch.load(lm_training_data)
+    words2bpe = data["words"]
+    sentences = data["sentences"]
+
+    ss = []
+    unk = sp.decode(sp.unk_id()).strip()
+    for i in range(10):
+        s = sp.decode(words2bpe[sentences[i]].values.tolist())
+        s = s.replace(unk, "<unk>")
+        ss.append(s)
+
+    for s in ss:
+        print(s)
+    # You can compare the output with the first 10 lines of ptb.train.txt
+
+
+if __name__ == "__main__":
+    formatter = (
+        "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    )
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/egs/ptb/LM/local/train_bpe_model.py b/egs/ptb/LM/local/train_bpe_model.py
new file mode 120000
index 000000000..6f018a0e2
--- /dev/null
+++ b/egs/ptb/LM/local/train_bpe_model.py
@@ -0,0 +1 @@
+../../librispeech/ASR/local/train_bpe_model.py
\ No newline at end of file
diff --git a/egs/ptb/LM/prepare.sh b/egs/ptb/LM/prepare.sh
new file mode 100755
index 000000000..70586785d
--- /dev/null
+++ b/egs/ptb/LM/prepare.sh
@@ -0,0 +1,115 @@
+#!/usr/bin/env bash
+
+set -eou pipefail
+
+nj=15
+stage=-1
+stop_stage=100
+
+dl_dir=$PWD/download
+# The following files will be downloaded to $dl_dir
+#  - ptb.train.txt
+#  - ptb.valid.txt
+#  - ptb.test.txt
+
+. shared/parse_options.sh || exit 1
+
+# vocab size for sentence piece models.
+# It will generate data/bpe_xxx, data/bpe_yyy
+# if the array contains xxx, yyy
+vocab_sizes=(
+  500
+  1000
+  2000
+  5000
+)
+
+# All files generated by this script are saved in "data".
+# You can safely remove "data" and rerun this script to regenerate it.
+mkdir -p data
+mkdir -p $dl_dir
+
+log() {
+  # This function is from espnet
+  local fname=${BASH_SOURCE[1]##*/}
+  echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
+}
+
+log "dl_dir: $dl_dir"
+
+if [ $stage -le -1 ] && [ $stop_stage -ge -1 ]; then
+  log "Stage -1: Download data"
+  if [ ! -f $dl_dir/.complete ]; then
+    url=https://raw.githubusercontent.com/townie/PTB-dataset-from-Tomas-Mikolov-s-webpage/master/data/
+    wget --no-verbose --directory-prefix $dl_dir $url/ptb.train.txt
+    wget --no-verbose --directory-prefix $dl_dir $url/ptb.valid.txt
+    wget --no-verbose --directory-prefix $dl_dir $url/ptb.test.txt
+    touch $dl_dir/.complete
+  fi
+fi
+
+if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Train BPE model"
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    out_dir=data/bpe_${vocab_size}
+    mkdir -p $out_dir
+    ./local/train_bpe_model.py \
+      --out-dir $out_dir \
+      --vocab-size $vocab_size \
+      --transcript $dl_dir/ptb.train.txt
+  done
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Generate LM training data"
+  # Note: ptb.train.txt has already been normalized
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    out_dir=data/bpe_${vocab_size}
+    mkdir -p $out_dir
+    ./local/prepare_lm_training_data.py \
+      --bpe-model $out_dir/bpe.model \
+      --lm-data $dl_dir/ptb.train.txt \
+      --lm-archive $out_dir/lm_data.pt
+
+    ./local/prepare_lm_training_data.py \
+      --bpe-model $out_dir/bpe.model \
+      --lm-data $dl_dir/ptb.valid.txt \
+      --lm-archive $out_dir/lm_data-valid.pt
+
+    ./local/prepare_lm_training_data.py \
+      --bpe-model $out_dir/bpe.model \
+      --lm-data $dl_dir/ptb.test.txt \
+      --lm-archive $out_dir/lm_data-test.pt
+  done
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Sort LM training data"
+  # Sort LM training data generated in stage 1
+  # by sentence length in descending order
+  # for ease of training.
+  #
+  # Sentence length equals to the number of BPE tokens
+  # in a sentence.
+
+  for vocab_size in ${vocab_sizes[@]}; do
+    out_dir=data/bpe_${vocab_size}
+    mkdir -p $out_dir
+    ./local/sort_lm_training_data.py \
+      --in-lm-data $out_dir/lm_data.pt \
+      --out-lm-data $out_dir/sorted_lm_data.pt \
+      --out-statistics $out_dir/statistics.txt
+
+    ./local/sort_lm_training_data.py \
+      --in-lm-data $out_dir/lm_data-valid.pt \
+      --out-lm-data $out_dir/sorted_lm_data-valid.pt \
+      --out-statistics $out_dir/statistics-valid.txt
+
+    ./local/sort_lm_training_data.py \
+      --in-lm-data $out_dir/lm_data-test.pt \
+      --out-lm-data $out_dir/sorted_lm_data-test.pt \
+      --out-statistics $out_dir/statistics-test.txt
+  done
+fi
diff --git a/egs/ptb/LM/shared b/egs/ptb/LM/shared
new file mode 120000
index 000000000..4c5e91438
--- /dev/null
+++ b/egs/ptb/LM/shared
@@ -0,0 +1 @@
+../../../icefall/shared/
\ No newline at end of file
diff --git a/icefall/decode.py b/icefall/decode.py
index 3ba899b4e..680e29619 100644
--- a/icefall/decode.py
+++ b/icefall/decode.py
@@ -20,7 +20,34 @@ from typing import Dict, List, Optional, Union
 import k2
 import torch
 
-from icefall.utils import get_texts
+from icefall.utils import add_eos, add_sos, get_texts
+
+DEFAULT_LM_SCALE = [
+    0.01,
+    0.05,
+    0.08,
+    0.1,
+    0.3,
+    0.5,
+    0.6,
+    0.7,
+    0.9,
+    1.0,
+    1.1,
+    1.2,
+    1.3,
+    1.5,
+    1.7,
+    1.9,
+    2.0,
+    2.1,
+    2.2,
+    2.3,
+    2.5,
+    3.0,
+    4.0,
+    5.0,
+]
 
 
 def _intersect_device(
@@ -952,3 +979,161 @@ def rescore_with_attention_decoder(
             key = f"ngram_lm_scale_{n_scale}_attention_scale_{a_scale}"
             ans[key] = best_path
     return ans
+
+
+def rescore_with_rnn_lm(
+    lattice: k2.Fsa,
+    num_paths: int,
+    rnn_lm_model: torch.nn.Module,
+    model: torch.nn.Module,
+    memory: torch.Tensor,
+    memory_key_padding_mask: Optional[torch.Tensor],
+    sos_id: int,
+    eos_id: int,
+    blank_id: int,
+    nbest_scale: float = 1.0,
+    ngram_lm_scale: Optional[float] = None,
+    attention_scale: Optional[float] = None,
+    rnn_lm_scale: Optional[float] = None,
+    use_double_scores: bool = True,
+) -> Dict[str, k2.Fsa]:
+    """This function extracts `num_paths` paths from the given lattice and uses
+    an attention decoder to rescore them. The path with the highest score is
+    the decoding output.
+
+    Args:
+      lattice:
+        An FsaVec with axes [utt][state][arc].
+      num_paths:
+        Number of paths to extract from the given lattice for rescoring.
+      model:
+        A transformer model. See the class "Transformer" in
+        conformer_ctc/transformer.py for its interface.
+      memory:
+        The encoder memory of the given model. It is the output of
+        the last torch.nn.TransformerEncoder layer in the given model.
+        Its shape is `(T, N, C)`.
+      memory_key_padding_mask:
+        The padding mask for memory with shape `(N, T)`.
+      sos_id:
+        The token ID for SOS.
+      eos_id:
+        The token ID for EOS.
+      nbest_scale:
+        It's the scale applied to `lattice.scores`. A smaller value
+        leads to more unique paths at the risk of missing the correct path.
+      ngram_lm_scale:
+        Optional. It specifies the scale for n-gram LM scores.
+      attention_scale:
+        Optional. It specifies the scale for attention decoder scores.
+      rnn_lm_scale:
+        Optional. It specifies the scale for RNN LM scores.
+    Returns:
+      A dict of FsaVec, whose key contains a string
+      ngram_lm_scale_attention_scale and the value is the
+      best decoding path for each utterance in the lattice.
+    """
+    nbest = Nbest.from_lattice(
+        lattice=lattice,
+        num_paths=num_paths,
+        use_double_scores=use_double_scores,
+        nbest_scale=nbest_scale,
+    )
+    # nbest.fsa.scores are all 0s at this point
+
+    nbest = nbest.intersect(lattice)
+    # Now nbest.fsa has its scores set.
+    # Also, nbest.fsa inherits the attributes from `lattice`.
+    assert hasattr(nbest.fsa, "lm_scores")
+
+    am_scores = nbest.compute_am_scores()
+    ngram_lm_scores = nbest.compute_lm_scores()
+
+    # The `tokens` attribute is set inside `compile_hlg.py`
+    assert hasattr(nbest.fsa, "tokens")
+    assert isinstance(nbest.fsa.tokens, torch.Tensor)
+
+    path_to_utt_map = nbest.shape.row_ids(1).to(torch.long)
+    # the shape of memory is (T, N, C), so we use axis=1 here
+    expanded_memory = memory.index_select(1, path_to_utt_map)
+
+    if memory_key_padding_mask is not None:
+        # The shape of memory_key_padding_mask is (N, T), so we
+        # use axis=0 here.
+        expanded_memory_key_padding_mask = memory_key_padding_mask.index_select(
+            0, path_to_utt_map
+        )
+    else:
+        expanded_memory_key_padding_mask = None
+
+    # remove axis corresponding to states.
+    tokens_shape = nbest.fsa.arcs.shape().remove_axis(1)
+    tokens = k2.RaggedTensor(tokens_shape, nbest.fsa.tokens)
+    tokens = tokens.remove_values_leq(0)
+    token_ids = tokens.tolist()
+
+    if len(token_ids) == 0:
+        print("Warning: rescore_with_attention_decoder(): empty token-ids")
+        return None
+
+    nll = model.decoder_nll(
+        memory=expanded_memory,
+        memory_key_padding_mask=expanded_memory_key_padding_mask,
+        token_ids=token_ids,
+        sos_id=sos_id,
+        eos_id=eos_id,
+    )
+    assert nll.ndim == 2
+    assert nll.shape[0] == len(token_ids)
+
+    attention_scores = -nll.sum(dim=1)
+
+    # Now for RNN LM
+    sos_tokens = add_sos(tokens, sos_id)
+    tokens_eos = add_eos(tokens, eos_id)
+    sos_tokens_row_splits = sos_tokens.shape.row_splits(1)
+    sentence_lengths = sos_tokens_row_splits[1:] - sos_tokens_row_splits[:-1]
+
+    x_tokens = sos_tokens.pad(mode="constant", padding_value=blank_id)
+    y_tokens = tokens_eos.pad(mode="constant", padding_value=blank_id)
+
+    x_tokens = x_tokens.to(torch.int64)
+    y_tokens = y_tokens.to(torch.int64)
+    sentence_lengths = sentence_lengths.to(torch.int64)
+
+    rnn_lm_nll = rnn_lm_model(x=x_tokens, y=y_tokens, lengths=sentence_lengths)
+    assert rnn_lm_nll.ndim == 2
+    assert rnn_lm_nll.shape[0] == len(token_ids)
+
+    rnn_lm_scores = -1 * rnn_lm_nll.sum(dim=1)
+
+    ngram_lm_scale_list = DEFAULT_LM_SCALE
+    attention_scale_list = DEFAULT_LM_SCALE
+    rnn_lm_scale_list = DEFAULT_LM_SCALE
+
+    if ngram_lm_scale:
+        ngram_lm_scale_list = [ngram_lm_scale]
+
+    if attention_scale:
+        attention_scale_list = [attention_scale]
+
+    if rnn_lm_scale:
+        rnn_lm_scale_list = [rnn_lm_scale]
+
+    ans = dict()
+    for n_scale in ngram_lm_scale_list:
+        for a_scale in attention_scale_list:
+            for r_scale in rnn_lm_scale_list:
+                tot_scores = (
+                    am_scores.values
+                    + n_scale * ngram_lm_scores.values
+                    + a_scale * attention_scores
+                    + r_scale * rnn_lm_scores
+                )
+                ragged_tot_scores = k2.RaggedTensor(nbest.shape, tot_scores)
+                max_indexes = ragged_tot_scores.argmax()
+                best_path = k2.index_fsa(nbest.fsa, max_indexes)
+
+                key = f"ngram_lm_scale_{n_scale}_attention_scale_{a_scale}_rnn_lm_scale_{r_scale}"  # noqa
+                ans[key] = best_path
+    return ans
diff --git a/icefall/dist.py b/icefall/dist.py
index 203c7c563..6334f9c13 100644
--- a/icefall/dist.py
+++ b/icefall/dist.py
@@ -21,14 +21,46 @@ import torch
 from torch import distributed as dist
 
 
-def setup_dist(rank, world_size, master_port=None):
-    os.environ["MASTER_ADDR"] = "localhost"
-    os.environ["MASTER_PORT"] = (
-        "12354" if master_port is None else str(master_port)
-    )
-    dist.init_process_group("nccl", rank=rank, world_size=world_size)
-    torch.cuda.set_device(rank)
+def setup_dist(rank, world_size, master_port=None, use_ddp_launch=False):
+    """
+    rank and world_size are used only if use_ddp_launch is False.
+    """
+    if "MASTER_ADDR" not in os.environ:
+        os.environ["MASTER_ADDR"] = "localhost"
+
+    if "MASTER_PORT" not in os.environ:
+        os.environ["MASTER_PORT"] = (
+            "12354" if master_port is None else str(master_port)
+        )
+
+    if use_ddp_launch is False:
+        dist.init_process_group("nccl", rank=rank, world_size=world_size)
+        torch.cuda.set_device(rank)
+    else:
+        dist.init_process_group("nccl")
 
 
 def cleanup_dist():
     dist.destroy_process_group()
+
+
+def get_world_size():
+    if "WORLD_SIZE" in os.environ:
+        return int(os.environ["WORLD_SIZE"])
+    if dist.is_available() and dist.is_initialized():
+        return dist.get_world_size()
+    else:
+        return 1
+
+
+def get_rank():
+    if "RANK" in os.environ:
+        return int(os.environ["RANK"])
+    elif dist.is_available() and dist.is_initialized():
+        return dist.rank()
+    else:
+        return 1
+
+
+def get_local_rank():
+    return int(os.environ.get("LOCAL_RANK", 0))
diff --git a/icefall/rnn_lm/compute_perplexity.py b/icefall/rnn_lm/compute_perplexity.py
new file mode 100755
index 000000000..550801a8f
--- /dev/null
+++ b/icefall/rnn_lm/compute_perplexity.py
@@ -0,0 +1,237 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+  ./rnn_lm/compute_perplexity.py \
+    --epoch 4 \
+    --avg 2 \
+    --lm-data ./data/bpe_500/sorted_lm_data-test.pt
+
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+
+import torch
+from dataset import get_dataloader
+from model import RnnLmModel
+
+from icefall.checkpoint import average_checkpoints, load_checkpoint
+from icefall.utils import AttributeDict, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=49,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=20,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="rnn_lm/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--lm-data",
+        type=str,
+        help="Path to the LM test data for computing perplexity",
+    )
+
+    parser.add_argument(
+        "--vocab-size",
+        type=int,
+        default=500,
+        help="Vocabulary size of the model",
+    )
+
+    parser.add_argument(
+        "--embedding-dim",
+        type=int,
+        default=2048,
+        help="Embedding dim of the model",
+    )
+
+    parser.add_argument(
+        "--hidden-dim",
+        type=int,
+        default=2048,
+        help="Hidden dim of the model",
+    )
+
+    parser.add_argument(
+        "--num-layers",
+        type=int,
+        default=3,
+        help="Number of RNN layers the model",
+    )
+
+    parser.add_argument(
+        "--tie-weights",
+        type=str2bool,
+        default=False,
+        help="""True to share the weights between the input embedding layer and the
+        last output linear layer
+        """,
+    )
+
+    parser.add_argument(
+        "--batch-size",
+        type=int,
+        default=50,
+        help="Number of RNN layers the model",
+    )
+
+    parser.add_argument(
+        "--max-sent-len",
+        type=int,
+        default=100,
+        help="Number of RNN layers the model",
+    )
+
+    parser.add_argument(
+        "--sos-id",
+        type=int,
+        default=1,
+        help="SOS ID",
+    )
+
+    parser.add_argument(
+        "--eos-id",
+        type=int,
+        default=1,
+        help="EOS ID",
+    )
+
+    parser.add_argument(
+        "--blank-id",
+        type=int,
+        default=0,
+        help="Blank ID",
+    )
+    return parser
+
+
+@torch.no_grad()
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+    args.lm_data = Path(args.lm_data)
+
+    params = AttributeDict(vars(args))
+
+    setup_logger(f"{params.exp_dir}/log-ppl/")
+    logging.info("Computing perplexity started")
+    logging.info(params)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"Device: {device}")
+
+    logging.info("About to create model")
+    model = RnnLmModel(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.embedding_dim,
+        hidden_dim=params.hidden_dim,
+        num_layers=params.num_layers,
+        tie_weights=params.tie_weights,
+    )
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+        model.to(device)
+    else:
+        start = params.epoch - params.avg + 1
+        filenames = []
+        for i in range(start, params.epoch + 1):
+            if start >= 0:
+                filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    model.eval()
+    num_param = sum([p.numel() for p in model.parameters()])
+    num_param_requires_grad = sum(
+        [p.numel() for p in model.parameters() if p.requires_grad]
+    )
+
+    logging.info(f"Number of model parameters: {num_param}")
+    logging.info(
+        f"Number of model parameters (requires_grad): "
+        f"{num_param_requires_grad} "
+        f"({num_param_requires_grad/num_param_requires_grad*100}%)"
+    )
+
+    logging.info(f"Loading LM test data from {params.lm_data}")
+    test_dl = get_dataloader(
+        filename=params.lm_data,
+        is_distributed=False,
+        params=params,
+    )
+
+    tot_loss = 0.0
+    num_tokens = 0
+    num_sentences = 0
+    for batch_idx, batch in enumerate(test_dl):
+        x, y, sentence_lengths = batch
+        x = x.to(device)
+        y = y.to(device)
+        sentence_lengths = sentence_lengths.to(device)
+
+        nll = model(x, y, sentence_lengths)
+        loss = nll.sum().cpu().item()
+
+        tot_loss += loss
+        num_tokens += sentence_lengths.sum().cpu().item()
+        num_sentences += x.size(0)
+
+    ppl = math.exp(tot_loss / num_tokens)
+    logging.info(
+        f"total nll: {tot_loss}, num tokens: {num_tokens}, "
+        f"num sentences: {num_sentences}, ppl: {ppl:.3f}"
+    )
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/icefall/rnn_lm/dataset.py b/icefall/rnn_lm/dataset.py
new file mode 100644
index 000000000..598e329c4
--- /dev/null
+++ b/icefall/rnn_lm/dataset.py
@@ -0,0 +1,218 @@
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import List, Tuple
+
+import k2
+import torch
+from torch.utils.data import DataLoader
+from torch.utils.data.distributed import DistributedSampler
+
+from icefall.utils import AttributeDict, add_eos, add_sos
+
+
+class LmDataset(torch.utils.data.Dataset):
+    def __init__(
+        self,
+        sentences: k2.RaggedTensor,
+        words: k2.RaggedTensor,
+        sentence_lengths: torch.Tensor,
+        max_sent_len: int,
+        batch_size: int,
+    ):
+        """
+        Args:
+          sentences:
+            A ragged tensor of dtype torch.int32 with 2 axes [sentence][word].
+          words:
+            A ragged tensor of dtype torch.int32 with 2 axes [word][token].
+          sentence_lengths:
+            A 1-D tensor of dtype torch.int32 containing number of tokens
+            of each sentence.
+          max_sent_len:
+            Maximum sentence length. It is used to change the batch size
+            dynamically. In general, we try to keep the product of
+            "max_sent_len in a batch" and "num_of_sent in a batch" being
+            a constant.
+          batch_size:
+            The expected batch size. It is changed dynamically according
+            to the "max_sent_len".
+
+        See `../local/prepare_lm_training_data.py` for how `sentences` and
+        `words` are generated. We assume that `sentences` are sorted by length.
+        See `../local/sort_lm_training_data.py`.
+        """
+        super().__init__()
+        self.sentences = sentences
+        self.words = words
+
+        sentence_lengths = sentence_lengths.tolist()
+
+        assert batch_size > 0, batch_size
+        assert max_sent_len > 1, max_sent_len
+        batch_indexes = []
+        num_sentences = sentences.dim0
+        cur = 0
+        while cur < num_sentences:
+            sz = sentence_lengths[cur] // max_sent_len + 1
+            # Assume the current sentence has 3 * max_sent_len tokens,
+            # in the worst case, the subsequent sentences also have
+            # this number of tokens, we should reduce the batch size
+            # so that this batch will not contain too many tokens
+            actual_batch_size = batch_size // sz + 1
+            actual_batch_size = min(actual_batch_size, batch_size)
+            end = cur + actual_batch_size
+            end = min(end, num_sentences)
+            this_batch_indexes = torch.arange(cur, end).tolist()
+            batch_indexes.append(this_batch_indexes)
+            cur = end
+        assert batch_indexes[-1][-1] == num_sentences - 1
+
+        self.batch_indexes = k2.RaggedTensor(batch_indexes)
+
+    def __len__(self) -> int:
+        """Return number of batches in this dataset"""
+        return self.batch_indexes.dim0
+
+    def __getitem__(self, i: int) -> k2.RaggedTensor:
+        """Get the i'th batch in this dataset
+        Return a ragged tensor with 2 axes [sentence][token].
+        """
+        assert 0 <= i < len(self), i
+
+        # indexes is a 1-D tensor containing sentence indexes
+        indexes = self.batch_indexes[i]
+
+        # sentence_words is a ragged tensor with 2 axes
+        # [sentence][word]
+        sentence_words = self.sentences[indexes]
+
+        # in case indexes contains only 1 entry, the returned
+        # sentence_words is a 1-D tensor, we have to convert
+        # it to a ragged tensor
+        if isinstance(sentence_words, torch.Tensor):
+            sentence_words = k2.RaggedTensor(sentence_words.unsqueeze(0))
+
+        # sentence_word_tokens is a ragged tensor with 3 axes
+        # [sentence][word][token]
+        sentence_word_tokens = self.words.index(sentence_words)
+        assert sentence_word_tokens.num_axes == 3
+
+        sentence_tokens = sentence_word_tokens.remove_axis(1)
+        return sentence_tokens
+
+
+class LmDatasetCollate:
+    def __init__(self, sos_id: int, eos_id: int, blank_id: int):
+        """
+        Args:
+          sos_id:
+            Token ID of the SOS symbol.
+          eos_id:
+            Token ID of the EOS symbol.
+          blank_id:
+            Token ID of the blank symbol.
+        """
+        self.sos_id = sos_id
+        self.eos_id = eos_id
+        self.blank_id = blank_id
+
+    def __call__(
+        self, batch: List[k2.RaggedTensor]
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Return a tuple containing 3 tensors:
+
+        - x, a 2-D tensor of dtype torch.int32; each row contains tokens
+             for a sentence starting with `self.sos_id`. It is padded to
+             the max sentence length with `self.blank_id`.
+
+        - y, a 2-D tensor of dtype torch.int32; each row contains tokens
+             for a sentence ending with `self.eos_id` before padding.
+             Then it is padded to the max sentence length with
+             `self.blank_id`.
+
+        - lengths, a 2-D tensor of dtype torch.int32, containing the number of
+                   tokens of each sentence before padding.
+        """
+        # The batching stuff has already been done in LmDataset
+        assert len(batch) == 1
+        sentence_tokens = batch[0]
+        row_splits = sentence_tokens.shape.row_splits(1)
+        sentence_token_lengths = row_splits[1:] - row_splits[:-1]
+        sentence_tokens_with_sos = add_sos(sentence_tokens, self.sos_id)
+        sentence_tokens_with_eos = add_eos(sentence_tokens, self.eos_id)
+
+        x = sentence_tokens_with_sos.pad(
+            mode="constant", padding_value=self.blank_id
+        )
+        y = sentence_tokens_with_eos.pad(
+            mode="constant", padding_value=self.blank_id
+        )
+        sentence_token_lengths += 1  # plus 1 since we added a SOS
+
+        return x.to(torch.int64), y.to(torch.int64), sentence_token_lengths
+
+
+def get_dataloader(
+    filename: str,
+    is_distributed: bool,
+    params: AttributeDict,
+) -> torch.utils.data.DataLoader:
+    """Get dataloader for LM training.
+
+    Args:
+      filename:
+        Path to the file containing LM data. The file is assumed to
+        be generated by `../local/sort_lm_training_data.py`.
+      is_distributed:
+        True if using DDP training. False otherwise.
+      params:
+        Set `get_params()` from `rnn_lm/train.py`
+    Returns:
+      Return a dataloader containing the LM data.
+    """
+    lm_data = torch.load(filename)
+
+    words = lm_data["words"]
+    sentences = lm_data["sentences"]
+    sentence_lengths = lm_data["sentence_lengths"]
+
+    dataset = LmDataset(
+        sentences=sentences,
+        words=words,
+        sentence_lengths=sentence_lengths,
+        max_sent_len=params.max_sent_len,
+        batch_size=params.batch_size,
+    )
+    if is_distributed:
+        sampler = DistributedSampler(dataset, shuffle=True, drop_last=False)
+    else:
+        sampler = None
+
+    collate_fn = LmDatasetCollate(
+        sos_id=params.sos_id,
+        eos_id=params.eos_id,
+        blank_id=params.blank_id,
+    )
+
+    dataloader = DataLoader(
+        dataset,
+        batch_size=1,
+        collate_fn=collate_fn,
+        sampler=sampler,
+        shuffle=sampler is None,
+    )
+    return dataloader
diff --git a/icefall/rnn_lm/export.py b/icefall/rnn_lm/export.py
new file mode 100644
index 000000000..094035fce
--- /dev/null
+++ b/icefall/rnn_lm/export.py
@@ -0,0 +1,167 @@
+#!/usr/bin/env python3
+#
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This script converts several saved checkpoints
+# to a single one using model averaging.
+
+import argparse
+import logging
+from pathlib import Path
+
+import torch
+from model import RnnLmModel
+
+from icefall.checkpoint import load_checkpoint
+from icefall.utils import AttributeDict, load_averaged_model, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=29,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=5,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--vocab-size",
+        type=int,
+        default=500,
+        help="Vocabulary size of the model",
+    )
+
+    parser.add_argument(
+        "--embedding-dim",
+        type=int,
+        default=2048,
+        help="Embedding dim of the model",
+    )
+
+    parser.add_argument(
+        "--hidden-dim",
+        type=int,
+        default=2048,
+        help="Hidden dim of the model",
+    )
+
+    parser.add_argument(
+        "--num-layers",
+        type=int,
+        default=3,
+        help="Number of RNN layers the model",
+    )
+
+    parser.add_argument(
+        "--tie-weights",
+        type=str2bool,
+        default=True,
+        help="""True to share the weights between the input embedding layer and the
+        last output linear layer
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="rnn_lm/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=True,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    params = AttributeDict({})
+    params.update(vars(args))
+
+    logging.info(params)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    model = RnnLmModel(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.embedding_dim,
+        hidden_dim=params.hidden_dim,
+        num_layers=params.num_layers,
+        tie_weights=params.tie_weights,
+    )
+
+    model.to(device)
+
+    if params.avg == 1:
+        load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
+    else:
+        model = load_averaged_model(
+            params.exp_dir, model, params.epoch, params.avg, device
+        )
+
+    model.to("cpu")
+    model.eval()
+
+    if params.jit:
+        logging.info("Using torch.jit.script")
+        model = torch.jit.script(model)
+        filename = params.exp_dir / "cpu_jit.pt"
+        model.save(str(filename))
+        logging.info(f"Saved to {filename}")
+    else:
+        logging.info("Not using torch.jit.script")
+        # Save it using a format so that it can be loaded
+        # by :func:`load_checkpoint`
+        filename = params.exp_dir / "pretrained.pt"
+        torch.save({"model": model.state_dict()}, str(filename))
+        logging.info(f"Saved to {filename}")
+
+
+if __name__ == "__main__":
+    formatter = (
+        "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    )
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+    main()
diff --git a/icefall/rnn_lm/model.py b/icefall/rnn_lm/model.py
new file mode 100644
index 000000000..88b2cc41f
--- /dev/null
+++ b/icefall/rnn_lm/model.py
@@ -0,0 +1,120 @@
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import logging
+
+import torch
+import torch.nn.functional as F
+
+from icefall.utils import make_pad_mask
+
+
+class RnnLmModel(torch.nn.Module):
+    def __init__(
+        self,
+        vocab_size: int,
+        embedding_dim: int,
+        hidden_dim: int,
+        num_layers: int,
+        tie_weights: bool = False,
+    ):
+        """
+        Args:
+          vocab_size:
+            Vocabulary size of BPE model.
+          embedding_dim:
+            Input embedding dimension.
+          hidden_dim:
+            Hidden dimension of RNN layers.
+          num_layers:
+            Number of RNN layers.
+          tie_weights:
+            True to share the weights between the input embedding layer and the
+            last output linear layer. See https://arxiv.org/abs/1608.05859
+            and https://arxiv.org/abs/1611.01462
+        """
+        super().__init__()
+
+        self.input_embedding = torch.nn.Embedding(
+            num_embeddings=vocab_size,
+            embedding_dim=embedding_dim,
+        )
+
+        self.rnn = torch.nn.LSTM(
+            input_size=embedding_dim,
+            hidden_size=hidden_dim,
+            num_layers=num_layers,
+            batch_first=True,
+        )
+
+        self.output_linear = torch.nn.Linear(
+            in_features=hidden_dim, out_features=vocab_size
+        )
+
+        self.vocab_size = vocab_size
+        if tie_weights:
+            logging.info("Tying weights")
+            assert embedding_dim == hidden_dim, (embedding_dim, hidden_dim)
+            self.output_linear.weight = self.input_embedding.weight
+        else:
+            logging.info("Not tying weights")
+
+    def forward(
+        self, x: torch.Tensor, y: torch.Tensor, lengths: torch.Tensor
+    ) -> torch.Tensor:
+        """
+        Args:
+          x:
+            A 2-D tensor with shape (N, L). Each row
+            contains token IDs for a sentence and starts with the SOS token.
+          y:
+            A shifted version of `x` and with EOS appended.
+          lengths:
+            A 1-D tensor of shape (N,). It contains the sentence lengths
+            before padding.
+        Returns:
+          Return a 2-D tensor of shape (N, L) containing negative log-likelihood
+          loss values. Note: Loss values for padding positions are set to 0.
+        """
+        assert x.ndim == y.ndim == 2, (x.ndim, y.ndim)
+        assert lengths.ndim == 1, lengths.ndim
+        assert x.shape == y.shape, (x.shape, y.shape)
+
+        batch_size = x.size(0)
+        assert lengths.size(0) == batch_size, (lengths.size(0), batch_size)
+
+        # embedding is of shape (N, L, embedding_dim)
+        embedding = self.input_embedding(x)
+
+        # Note: We use batch_first==True
+        rnn_out, _ = self.rnn(embedding)
+        logits = self.output_linear(rnn_out)
+
+        # Note: No need to use `log_softmax()` here
+        # since F.cross_entropy() expects unnormalized probabilities
+
+        # nll_loss is of shape (N*L,)
+        # nll -> negative log-likelihood
+        nll_loss = F.cross_entropy(
+            logits.reshape(-1, self.vocab_size), y.reshape(-1), reduction="none"
+        )
+        # Set loss values for padding positions to 0
+        mask = make_pad_mask(lengths).reshape(-1)
+        nll_loss.masked_fill_(mask, 0)
+
+        nll_loss = nll_loss.reshape(batch_size, -1)
+
+        return nll_loss
diff --git a/icefall/rnn_lm/test_dataset.py b/icefall/rnn_lm/test_dataset.py
new file mode 100755
index 000000000..bf961f54b
--- /dev/null
+++ b/icefall/rnn_lm/test_dataset.py
@@ -0,0 +1,71 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import k2
+import torch
+from rnn_lm.dataset import LmDataset, LmDatasetCollate
+
+
+def main():
+    sentences = k2.RaggedTensor(
+        [[0, 1, 2], [1, 0, 1], [0, 1], [1, 3, 0, 2, 0], [3], [0, 2, 1]]
+    )
+    words = k2.RaggedTensor([[3, 6], [2, 8, 9, 3], [5], [5, 6, 7, 8, 9]])
+
+    num_sentences = sentences.dim0
+
+    sentence_lengths = [0] * num_sentences
+    for i in range(num_sentences):
+        word_ids = sentences[i]
+
+        # NOTE: If word_ids is a tensor with only 1 entry,
+        # token_ids is a torch.Tensor
+        token_ids = words[word_ids]
+        if isinstance(token_ids, k2.RaggedTensor):
+            token_ids = token_ids.values
+
+        # token_ids is a 1-D tensor containing the BPE tokens
+        # of the current sentence
+
+        sentence_lengths[i] = token_ids.numel()
+
+    sentence_lengths = torch.tensor(sentence_lengths, dtype=torch.int32)
+
+    indices = torch.argsort(sentence_lengths, descending=True)
+    sentences = sentences[indices.to(torch.int32)]
+    sentence_lengths = sentence_lengths[indices]
+
+    dataset = LmDataset(
+        sentences=sentences,
+        words=words,
+        sentence_lengths=sentence_lengths,
+        max_sent_len=3,
+        batch_size=4,
+    )
+
+    collate_fn = LmDatasetCollate(sos_id=1, eos_id=-1, blank_id=0)
+    dataloader = torch.utils.data.DataLoader(
+        dataset, batch_size=1, collate_fn=collate_fn
+    )
+
+    for i in dataloader:
+        print(i)
+    # I've checked the output manually; the output is as expected.
+
+
+if __name__ == "__main__":
+    main()
diff --git a/icefall/rnn_lm/test_dataset_ddp.py b/icefall/rnn_lm/test_dataset_ddp.py
new file mode 100755
index 000000000..48fbb19cb
--- /dev/null
+++ b/icefall/rnn_lm/test_dataset_ddp.py
@@ -0,0 +1,103 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+
+import k2
+import torch
+import torch.multiprocessing as mp
+from rnn_lm.dataset import LmDataset, LmDatasetCollate
+from torch import distributed as dist
+
+
+def generate_data():
+    sentences = k2.RaggedTensor(
+        [[0, 1, 2], [1, 0, 1], [0, 1], [1, 3, 0, 2, 0], [3], [0, 2, 1]]
+    )
+    words = k2.RaggedTensor([[3, 6], [2, 8, 9, 3], [5], [5, 6, 7, 8, 9]])
+
+    num_sentences = sentences.dim0
+
+    sentence_lengths = [0] * num_sentences
+    for i in range(num_sentences):
+        word_ids = sentences[i]
+
+        # NOTE: If word_ids is a tensor with only 1 entry,
+        # token_ids is a torch.Tensor
+        token_ids = words[word_ids]
+        if isinstance(token_ids, k2.RaggedTensor):
+            token_ids = token_ids.values
+
+        # token_ids is a 1-D tensor containing the BPE tokens
+        # of the current sentence
+
+        sentence_lengths[i] = token_ids.numel()
+
+    sentence_lengths = torch.tensor(sentence_lengths, dtype=torch.int32)
+
+    indices = torch.argsort(sentence_lengths, descending=True)
+    sentences = sentences[indices.to(torch.int32)]
+    sentence_lengths = sentence_lengths[indices]
+
+    return sentences, words, sentence_lengths
+
+
+def run(rank, world_size):
+    os.environ["MASTER_ADDR"] = "localhost"
+    os.environ["MASTER_PORT"] = "12352"
+
+    dist.init_process_group("nccl", rank=rank, world_size=world_size)
+    torch.cuda.set_device(rank)
+
+    sentences, words, sentence_lengths = generate_data()
+
+    dataset = LmDataset(
+        sentences=sentences,
+        words=words,
+        sentence_lengths=sentence_lengths,
+        max_sent_len=3,
+        batch_size=4,
+    )
+    sampler = torch.utils.data.distributed.DistributedSampler(
+        dataset, shuffle=True, drop_last=False
+    )
+
+    collate_fn = LmDatasetCollate(sos_id=1, eos_id=-1, blank_id=0)
+    dataloader = torch.utils.data.DataLoader(
+        dataset,
+        batch_size=1,
+        collate_fn=collate_fn,
+        sampler=sampler,
+        shuffle=False,
+    )
+
+    for i in dataloader:
+        print(f"rank: {rank}", i)
+
+    dist.destroy_process_group()
+
+
+def main():
+    world_size = 2
+    mp.spawn(run, args=(world_size,), nprocs=world_size, join=True)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/icefall/rnn_lm/test_model.py b/icefall/rnn_lm/test_model.py
new file mode 100755
index 000000000..5a216a3fb
--- /dev/null
+++ b/icefall/rnn_lm/test_model.py
@@ -0,0 +1,69 @@
+#!/usr/bin/env python3
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+from rnn_lm.model import RnnLmModel
+
+
+def test_rnn_lm_model():
+    vocab_size = 4
+    model = RnnLmModel(
+        vocab_size=vocab_size, embedding_dim=10, hidden_dim=10, num_layers=2
+    )
+    x = torch.tensor(
+        [
+            [1, 3, 2, 2],
+            [1, 2, 2, 0],
+            [1, 2, 0, 0],
+        ]
+    )
+    y = torch.tensor(
+        [
+            [3, 2, 2, 1],
+            [2, 2, 1, 0],
+            [2, 1, 0, 0],
+        ]
+    )
+    lengths = torch.tensor([4, 3, 2])
+    nll_loss = model(x, y, lengths)
+    print(nll_loss)
+    """
+    tensor([[1.1180, 1.3059, 1.2426, 1.7773],
+            [1.4231, 1.2783, 1.7321, 0.0000],
+            [1.4231, 1.6752, 0.0000, 0.0000]], grad_fn=<ViewBackward>)
+    """
+
+
+def test_rnn_lm_model_tie_weights():
+    model = RnnLmModel(
+        vocab_size=10,
+        embedding_dim=10,
+        hidden_dim=10,
+        num_layers=2,
+        tie_weights=True,
+    )
+    assert model.input_embedding.weight is model.output_linear.weight
+
+
+def main():
+    test_rnn_lm_model()
+    test_rnn_lm_model_tie_weights()
+
+
+if __name__ == "__main__":
+    torch.manual_seed(20211122)
+    main()
diff --git a/icefall/rnn_lm/train.py b/icefall/rnn_lm/train.py
new file mode 100755
index 000000000..bb5f03fb9
--- /dev/null
+++ b/icefall/rnn_lm/train.py
@@ -0,0 +1,617 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+#
+# See ../../../../LICENSE for clarification regarding multiple authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage:
+    ./rnn_lm/train.py \
+        --start-epoch 0 \
+        --world-size 2 \
+        --num-epochs 1 \
+        --use-fp16 0 \
+        --embedding-dim 800 \
+        --hidden-dim 200 \
+        --num-layers 2\
+        --batch-size 400
+
+"""
+
+import argparse
+import logging
+import math
+from pathlib import Path
+from shutil import copyfile
+from typing import Optional, Tuple
+
+import torch
+import torch.multiprocessing as mp
+import torch.nn as nn
+import torch.optim as optim
+from dataset import get_dataloader
+from lhotse.utils import fix_random_seed
+from model import RnnLmModel
+from torch.nn.parallel import DistributedDataParallel as DDP
+from torch.nn.utils import clip_grad_norm_
+from torch.utils.tensorboard import SummaryWriter
+
+from icefall.checkpoint import load_checkpoint
+from icefall.checkpoint import save_checkpoint as save_checkpoint_impl
+from icefall.dist import cleanup_dist, setup_dist
+from icefall.env import get_env_info
+from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--world-size",
+        type=int,
+        default=1,
+        help="Number of GPUs for DDP training.",
+    )
+
+    parser.add_argument(
+        "--master-port",
+        type=int,
+        default=12354,
+        help="Master port to use for DDP training.",
+    )
+
+    parser.add_argument(
+        "--tensorboard",
+        type=str2bool,
+        default=True,
+        help="Should various information be logged in tensorboard.",
+    )
+
+    parser.add_argument(
+        "--num-epochs",
+        type=int,
+        default=10,
+        help="Number of epochs to train.",
+    )
+
+    parser.add_argument(
+        "--start-epoch",
+        type=int,
+        default=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        exp_dir/epoch-{start_epoch-1}.pt
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="rnn_lm/exp",
+        help="""The experiment dir.
+        It specifies the directory where all training related
+        files, e.g., checkpoints, logs, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    parser.add_argument(
+        "--batch-size",
+        type=int,
+        default=50,
+    )
+
+    parser.add_argument(
+        "--lm-data",
+        type=str,
+        default="data/lm_training_bpe_500/sorted_lm_data.pt",
+        help="LM training data",
+    )
+
+    parser.add_argument(
+        "--lm-data-valid",
+        type=str,
+        default="data/lm_training_bpe_500/sorted_lm_data-valid.pt",
+        help="LM validation data",
+    )
+
+    parser.add_argument(
+        "--vocab-size",
+        type=int,
+        default=500,
+        help="Vocabulary size of the model",
+    )
+
+    parser.add_argument(
+        "--embedding-dim",
+        type=int,
+        default=2048,
+        help="Embedding dim of the model",
+    )
+
+    parser.add_argument(
+        "--hidden-dim",
+        type=int,
+        default=2048,
+        help="Hidden dim of the model",
+    )
+
+    parser.add_argument(
+        "--num-layers",
+        type=int,
+        default=3,
+        help="Number of RNN layers the model",
+    )
+
+    parser.add_argument(
+        "--tie-weights",
+        type=str2bool,
+        default=False,
+        help="""True to share the weights between the input embedding layer and the
+        last output linear layer
+        """,
+    )
+
+    parser.add_argument(
+        "--seed",
+        type=int,
+        default=42,
+        help="The seed for random generators intended for reproducibility",
+    )
+
+    return parser
+
+
+def get_params() -> AttributeDict:
+    """Return a dict containing training parameters."""
+
+    params = AttributeDict(
+        {
+            "max_sent_len": 200,
+            "sos_id": 1,
+            "eos_id": 1,
+            "blank_id": 0,
+            "lr": 1e-3,
+            "weight_decay": 1e-6,
+            "best_train_loss": float("inf"),
+            "best_valid_loss": float("inf"),
+            "best_train_epoch": -1,
+            "best_valid_epoch": -1,
+            "batch_idx_train": 0,
+            "log_interval": 200,
+            "reset_interval": 2000,
+            "valid_interval": 5000,
+            "env_info": get_env_info(),
+        }
+    )
+    return params
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+) -> None:
+    """Load checkpoint from file.
+
+    If params.start_epoch is positive, it will load the checkpoint from
+    `params.start_epoch - 1`. Otherwise, this function does nothing.
+
+    Apart from loading state dict for `model`, `optimizer` and `scheduler`,
+    it also updates `best_train_epoch`, `best_train_loss`, `best_valid_epoch`,
+    and `best_valid_loss` in `params`.
+
+    Args:
+      params:
+        The return value of :func:`get_params`.
+      model:
+        The training model.
+      optimizer:
+        The optimizer that we are using.
+      scheduler:
+        The learning rate scheduler we are using.
+    Returns:
+      Return None.
+    """
+    if params.start_epoch <= 0:
+        return
+
+    filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt"
+    logging.info(f"Loading checkpoint: {filename}")
+    saved_params = load_checkpoint(
+        filename,
+        model=model,
+        optimizer=optimizer,
+        scheduler=scheduler,
+    )
+
+    keys = [
+        "best_train_epoch",
+        "best_valid_epoch",
+        "batch_idx_train",
+        "best_train_loss",
+        "best_valid_loss",
+    ]
+    for k in keys:
+        params[k] = saved_params[k]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: Optional[torch.optim.Optimizer] = None,
+    scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None,
+    rank: int = 0,
+) -> None:
+    """Save model, optimizer, scheduler and training stats to file.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The training model.
+    """
+    if rank != 0:
+        return
+    filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt"
+    save_checkpoint_impl(
+        filename=filename,
+        model=model,
+        params=params,
+        optimizer=optimizer,
+        scheduler=scheduler,
+        rank=rank,
+    )
+
+    if params.best_train_epoch == params.cur_epoch:
+        best_train_filename = params.exp_dir / "best-train-loss.pt"
+        copyfile(src=filename, dst=best_train_filename)
+
+    if params.best_valid_epoch == params.cur_epoch:
+        best_valid_filename = params.exp_dir / "best-valid-loss.pt"
+        copyfile(src=filename, dst=best_valid_filename)
+
+
+def compute_loss(
+    model: nn.Module,
+    x: torch.Tensor,
+    y: torch.Tensor,
+    sentence_lengths: torch.Tensor,
+    is_training: bool,
+) -> Tuple[torch.Tensor, MetricsTracker]:
+    """Compute the negative log-likelihood loss given a model and its input.
+    Args:
+      model:
+        The NN model, e.g., RnnLmModel.
+      x:
+        A 2-D tensor. Each row contains BPE token IDs for a sentence. Also,
+        each row starts with SOS ID.
+      y:
+        A 2-D tensor. Each row is a shifted version of the corresponding row
+        in `x` but ends with an EOS ID (before padding).
+     sentence_lengths:
+       A 1-D tensor containing number of tokens of each sentence
+       before padding.
+     is_training:
+       True for training. False for validation.
+    """
+    with torch.set_grad_enabled(is_training):
+        device = model.device
+        x = x.to(device)
+        y = y.to(device)
+        sentence_lengths = sentence_lengths.to(device)
+
+        nll = model(x, y, sentence_lengths)
+        loss = nll.sum()
+
+        num_tokens = sentence_lengths.sum().item()
+
+        loss_info = MetricsTracker()
+        # Note: Due to how MetricsTracker() is designed,
+        # we use "frames" instead of "num_tokens" as a key here
+        loss_info["frames"] = num_tokens
+        loss_info["loss"] = loss.detach().item()
+    return loss, loss_info
+
+
+def compute_validation_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    valid_dl: torch.utils.data.DataLoader,
+    world_size: int = 1,
+) -> MetricsTracker:
+    """Run the validation process. The validation loss
+    is saved in `params.valid_loss`.
+    """
+    model.eval()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(valid_dl):
+        x, y, sentence_lengths = batch
+        with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            loss, loss_info = compute_loss(
+                model=model,
+                x=x,
+                y=y,
+                sentence_lengths=sentence_lengths,
+                is_training=False,
+            )
+
+        assert loss.requires_grad is False
+        tot_loss = tot_loss + loss_info
+
+    if world_size > 1:
+        tot_loss.reduce(loss.device)
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    if loss_value < params.best_valid_loss:
+        params.best_valid_epoch = params.cur_epoch
+        params.best_valid_loss = loss_value
+
+    return tot_loss
+
+
+def train_one_epoch(
+    params: AttributeDict,
+    model: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    tb_writer: Optional[SummaryWriter] = None,
+    world_size: int = 1,
+) -> None:
+    """Train the model for one epoch.
+
+    The training loss from the mean of all sentences is saved in
+    `params.train_loss`. It runs the validation process every
+    `params.valid_interval` batches.
+
+    Args:
+      params:
+        It is returned by :func:`get_params`.
+      model:
+        The model for training.
+      optimizer:
+        The optimizer we are using.
+      train_dl:
+        Dataloader for the training dataset.
+      valid_dl:
+        Dataloader for the validation dataset.
+      tb_writer:
+        Writer to write log messages to tensorboard.
+      world_size:
+        Number of nodes in DDP training. If it is 1, DDP is disabled.
+    """
+    model.train()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+        params.batch_idx_train += 1
+        x, y, sentence_lengths = batch
+        batch_size = x.size(0)
+        with torch.cuda.amp.autocast(enabled=params.use_fp16):
+            loss, loss_info = compute_loss(
+                model=model,
+                x=x,
+                y=y,
+                sentence_lengths=sentence_lengths,
+                is_training=True,
+            )
+
+        # summary stats
+        tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
+
+        optimizer.zero_grad()
+        loss.backward()
+        clip_grad_norm_(model.parameters(), 5.0, 2.0)
+        optimizer.step()
+
+        if batch_idx % params.log_interval == 0:
+            # Note: "frames" here means "num_tokens"
+            this_batch_ppl = math.exp(loss_info["loss"] / loss_info["frames"])
+            tot_ppl = math.exp(tot_loss["loss"] / tot_loss["frames"])
+
+            logging.info(
+                f"Epoch {params.cur_epoch}, "
+                f"batch {batch_idx}, loss[{loss_info}, ppl: {this_batch_ppl}] "
+                f"tot_loss[{tot_loss}, ppl: {tot_ppl}], "
+                f"batch size: {batch_size}"
+            )
+
+            if tb_writer is not None:
+                loss_info.write_summary(
+                    tb_writer, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(
+                    tb_writer, "train/tot_", params.batch_idx_train
+                )
+
+                tb_writer.add_scalar(
+                    "train/current_ppl", this_batch_ppl, params.batch_idx_train
+                )
+
+                tb_writer.add_scalar(
+                    "train/tot_ppl", tot_ppl, params.batch_idx_train
+                )
+
+        if batch_idx > 0 and batch_idx % params.valid_interval == 0:
+            logging.info("Computing validation loss")
+
+            valid_info = compute_validation_loss(
+                params=params,
+                model=model,
+                valid_dl=valid_dl,
+                world_size=world_size,
+            )
+            model.train()
+
+            valid_ppl = math.exp(valid_info["loss"] / valid_info["frames"])
+            logging.info(
+                f"Epoch {params.cur_epoch}, validation: {valid_info}, "
+                f"ppl: {valid_ppl}"
+            )
+
+            if tb_writer is not None:
+                valid_info.write_summary(
+                    tb_writer, "train/valid_", params.batch_idx_train
+                )
+
+                tb_writer.add_scalar(
+                    "train/valid_ppl", valid_ppl, params.batch_idx_train
+                )
+
+    loss_value = tot_loss["loss"] / tot_loss["frames"]
+    params.train_loss = loss_value
+    if params.train_loss < params.best_train_loss:
+        params.best_train_epoch = params.cur_epoch
+        params.best_train_loss = params.train_loss
+
+
+def run(rank, world_size, args):
+    """
+    Args:
+      rank:
+        It is a value between 0 and `world_size-1`, which is
+        passed automatically by `mp.spawn()` in :func:`main`.
+        The node with rank 0 is responsible for saving checkpoint.
+      world_size:
+        Number of GPUs for DDP training.
+      args:
+        The return value of get_parser().parse_args()
+    """
+    params = get_params()
+    params.update(vars(args))
+    is_distributed = world_size > 1
+
+    fix_random_seed(params.seed)
+    if is_distributed:
+        setup_dist(rank, world_size, params.master_port)
+
+    setup_logger(f"{params.exp_dir}/log/log-train")
+    logging.info("Training started")
+    logging.info(params)
+
+    if args.tensorboard and rank == 0:
+        tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard")
+    else:
+        tb_writer = None
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", rank)
+
+    logging.info(f"Device: {device}")
+
+    logging.info("About to create model")
+    model = RnnLmModel(
+        vocab_size=params.vocab_size,
+        embedding_dim=params.embedding_dim,
+        hidden_dim=params.hidden_dim,
+        num_layers=params.num_layers,
+        tie_weights=params.tie_weights,
+    )
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if is_distributed:
+        model = DDP(model, device_ids=[rank])
+
+    model.device = device
+
+    optimizer = optim.Adam(
+        model.parameters(),
+        lr=params.lr,
+        weight_decay=params.weight_decay,
+    )
+    if checkpoints:
+        logging.info("Load optimizer state_dict from checkpoint")
+        optimizer.load_state_dict(checkpoints["optimizer"])
+
+    logging.info(f"Loading LM training data from {params.lm_data}")
+    train_dl = get_dataloader(
+        filename=params.lm_data,
+        is_distributed=is_distributed,
+        params=params,
+    )
+
+    logging.info(f"Loading LM validation data from {params.lm_data_valid}")
+    valid_dl = get_dataloader(
+        filename=params.lm_data_valid,
+        is_distributed=is_distributed,
+        params=params,
+    )
+
+    # Note: No learning rate scheduler is used here
+    for epoch in range(params.start_epoch, params.num_epochs):
+        if is_distributed:
+            train_dl.sampler.set_epoch(epoch)
+
+        params.cur_epoch = epoch
+
+        train_one_epoch(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            tb_writer=tb_writer,
+            world_size=world_size,
+        )
+
+        save_checkpoint(
+            params=params,
+            model=model,
+            optimizer=optimizer,
+            rank=rank,
+        )
+
+    logging.info("Done!")
+
+    if is_distributed:
+        torch.distributed.barrier()
+        cleanup_dist()
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    world_size = args.world_size
+    assert world_size >= 1
+    if world_size > 1:
+        mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True)
+    else:
+        run(rank=0, world_size=1, args=args)
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()
diff --git a/icefall/utils.py b/icefall/utils.py
index b38574f0c..10a2e6301 100644
--- a/icefall/utils.py
+++ b/icefall/utils.py
@@ -35,6 +35,8 @@ import torch.distributed as dist
 import torch.nn as nn
 from torch.utils.tensorboard import SummaryWriter
 
+from icefall.checkpoint import average_checkpoints
+
 Pathlike = Union[str, Path]
 
 
@@ -90,7 +92,11 @@ def str2bool(v):
 
 
 def setup_logger(
-    log_filename: Pathlike, log_level: str = "info", use_console: bool = True
+    log_filename: Pathlike,
+    log_level: str = "info",
+    rank: int = 0,
+    world_size: int = 1,
+    use_console: bool = True,
 ) -> None:
     """Setup log level.
 
@@ -100,12 +106,16 @@ def setup_logger(
       log_level:
         The log level to use, e.g., "debug", "info", "warning", "error",
         "critical"
+      rank:
+        Rank of this node in DDP training.
+      world_size:
+        Number of nodes in DDP training.
+      use_console:
+        True to also print logs to console.
     """
     now = datetime.now()
     date_time = now.strftime("%Y-%m-%d-%H-%M-%S")
-    if dist.is_available() and dist.is_initialized():
-        world_size = dist.get_world_size()
-        rank = dist.get_rank()
+    if world_size > 1:
         formatter = f"%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] ({rank}/{world_size}) %(message)s"  # noqa
         log_filename = f"{log_filename}-{date_time}-{rank}"
     else:
@@ -799,3 +809,34 @@ def optim_step_and_measure_param_change(
             delta = l2_norm(p_orig - p_new) / l2_norm(p_orig)
             relative_change[n] = delta.item()
     return relative_change
+
+
+def load_averaged_model(
+    model_dir: str,
+    model: torch.nn.Module,
+    epoch: int,
+    avg: int,
+    device: torch.device,
+):
+    """
+    Load a model which is the average of all checkpoints
+
+    :param model_dir: a str of the experiment directory
+    :param model: a torch.nn.Module instance
+
+    :param epoch: the last epoch to load from
+    :param avg: how many models to average from
+    :param device: move model to this device
+
+    :return: A model averaged
+    """
+
+    # start cannot be negative
+    start = max(epoch - avg + 1, 0)
+    filenames = [f"{model_dir}/epoch-{i}.pt" for i in range(start, epoch + 1)]
+
+    logging.info(f"averaging {filenames}")
+    model.to(device)
+    model.load_state_dict(average_checkpoints(filenames, device=device))
+
+    return model

From c391bfd1000e0b2d7ff906be529d4c5fd1bcb7b0 Mon Sep 17 00:00:00 2001
From: Mingshuang Luo <37799481+luomingshuang@users.noreply.github.com>
Date: Fri, 24 Jun 2022 10:40:46 +0800
Subject: [PATCH 8/9] fix errors for soft connection (#443)

---
 egs/ptb/LM/local/prepare_lm_training_data.py | 2 +-
 egs/ptb/LM/local/train_bpe_model.py          | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/egs/ptb/LM/local/prepare_lm_training_data.py b/egs/ptb/LM/local/prepare_lm_training_data.py
index eebce957e..abc00d421 120000
--- a/egs/ptb/LM/local/prepare_lm_training_data.py
+++ b/egs/ptb/LM/local/prepare_lm_training_data.py
@@ -1 +1 @@
-../../librispeech/ASR/local/prepare_lm_training_data.py
\ No newline at end of file
+../../../librispeech/ASR/local/prepare_lm_training_data.py
\ No newline at end of file
diff --git a/egs/ptb/LM/local/train_bpe_model.py b/egs/ptb/LM/local/train_bpe_model.py
index 6f018a0e2..6fad36421 120000
--- a/egs/ptb/LM/local/train_bpe_model.py
+++ b/egs/ptb/LM/local/train_bpe_model.py
@@ -1 +1 @@
-../../librispeech/ASR/local/train_bpe_model.py
\ No newline at end of file
+../../../librispeech/ASR/local/train_bpe_model.py
\ No newline at end of file

From c0ea33473868d964bb1242207ac9011b6bde3d08 Mon Sep 17 00:00:00 2001
From: Tiance Wang <wangtiance@gmail.com>
Date: Fri, 24 Jun 2022 19:31:09 +0800
Subject: [PATCH 9/9] fix bug of concatenating list to tuple (#444)

---
 egs/librispeech/ASR/pruned_transducer_stateless5/conformer.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/conformer.py b/egs/librispeech/ASR/pruned_transducer_stateless5/conformer.py
index bf3917df0..0fa4b6907 100644
--- a/egs/librispeech/ASR/pruned_transducer_stateless5/conformer.py
+++ b/egs/librispeech/ASR/pruned_transducer_stateless5/conformer.py
@@ -1195,7 +1195,7 @@ class RandomCombine(nn.Module):
         ans = torch.matmul(stacked_inputs, weights)
         # ans: (*, num_channels)
 
-        ans = ans.reshape(inputs[0].shape[:-1] + [num_channels])
+        ans = ans.reshape(inputs[0].shape[:-1] + (num_channels,))
 
         # The following if causes errors for torch script in torch 1.6.0
         #  if __name__ == "__main__":