Merge remote-tracking branch 'k2-fsa/master' into pruned4_result_new

.github/workflows/run-pretrained-conformer-ctc.yml

@@ -49,6 +49,8 @@ jobs:
       - 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

.github/workflows/run-pretrained-transducer-stateless-modified-2-aishell.yml

@@ -49,6 +49,8 @@ jobs:
       - 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

.github/workflows/run-pretrained-transducer-stateless-modified-aishell.yml

@@ -49,6 +49,8 @@ jobs:
       - 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

.github/workflows/run-pretrained-transducer.yml

@@ -49,6 +49,8 @@ jobs:
       - 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

.github/workflows/run-yesno-recipe.yml

@@ -62,6 +62,8 @@ jobs:
       - 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: Run yesno recipe
         shell: bash

.github/workflows/test.yml

@@ -76,6 +76,9 @@ jobs:
           pip install k2==${{ matrix.k2-version }}+cpu.torch${{ matrix.torch }} -f https://k2-fsa.org/nightly/
           pip install git+https://github.com/lhotse-speech/lhotse
           # icefall requirements
+          pip uninstall -y protobuf
+          pip install --no-binary protobuf protobuf
+
           pip install -r requirements.txt
 
       - name: Install graphviz

README.md

@@ -12,7 +12,7 @@ for installation.
 Please refer to <https://icefall.readthedocs.io/en/latest/recipes/index.html>
 for more information.
 
-We provide 6 recipes at present:
+We provide the following recipes:
 
   - [yesno][yesno]
   - [LibriSpeech][librispeech]
@@ -22,6 +22,7 @@ We provide 6 recipes at present:
   - [GigaSpeech][gigaspeech]
   - [Aidatatang_200zh][aidatatang_200zh]
   - [WenetSpeech][wenetspeech]
+  - [Alimeeting][alimeeting]
 
 ### yesno
 
@@ -126,7 +127,7 @@ The best CER we currently have is:
 | CER | 4.26 |
 
 
-We provide a Colab notebook to run a pre-trained conformer CTC model: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1WnG17io5HEZ0Gn_cnh_VzK5QYOoiiklC?usp=sharing)
+We provide a Colab notebook to run a pre-trained conformer CTC model: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg](https://colab.research.google.com/drive/1WnG17io5HEZ0Gn_cnh_VzK5QYOoiiklC?usp=sharing)
 
 #### Transducer Stateless Model
 
@@ -247,6 +248,20 @@ We provide one model for this recipe: [Pruned stateless RNN-T: Conformer encoder
 
 We provide a Colab notebook to run a pre-trained Pruned Transducer Stateless model: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1EV4e1CHa1GZgEF-bZgizqI9RyFFehIiN?usp=sharing)
 
+### Alimeeting
+
+We provide one model for this recipe: [Pruned stateless RNN-T: Conformer encoder + Embedding decoder + k2 pruned RNN-T loss][Alimeeting_pruned_transducer_stateless2].
+
+#### Pruned stateless RNN-T: Conformer encoder + Embedding decoder + k2 pruned RNN-T loss (trained with far subset)
+
+|                      |  Eval  | Test-Net |
+|----------------------|--------|----------|
+|    greedy search     | 31.77  |  34.66   |
+|   fast beam search   | 31.39  |  33.02   |
+| modified beam search | 30.38  |  34.25   |
+
+We provide a Colab notebook to run a pre-trained Pruned Transducer Stateless model: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1tKr3f0mL17uO_ljdHGKtR7HOmthYHwJG?usp=sharing)
+
 ## Deployment with C++
 
 Once you have trained a model in icefall, you may want to deploy it with C++,
@@ -274,6 +289,7 @@ Please see: [![Open In Colab](https://colab.research.google.com/assets/colab-bad
 [GigaSpeech_pruned_transducer_stateless2]: egs/gigaspeech/ASR/pruned_transducer_stateless2
 [Aidatatang_200zh_pruned_transducer_stateless2]: egs/aidatatang_200zh/ASR/pruned_transducer_stateless2
 [WenetSpeech_pruned_transducer_stateless2]: egs/wenetspeech/ASR/pruned_transducer_stateless2
+[Alimeeting_pruned_transducer_stateless2]: egs/alimeeting/ASR/pruned_transducer_stateless2
 [yesno]: egs/yesno/ASR
 [librispeech]: egs/librispeech/ASR
 [aishell]: egs/aishell/ASR
@@ -282,4 +298,6 @@ Please see: [![Open In Colab](https://colab.research.google.com/assets/colab-bad
 [gigaspeech]: egs/gigaspeech/ASR
 [aidatatang_200zh]: egs/aidatatang_200zh/ASR
 [wenetspeech]: egs/wenetspeech/ASR
+[alimeeting]: egs/alimeeting/ASR
 [k2]: https://github.com/k2-fsa/k2
+)

egs/aidatatang_200zh/ASR/local/compute_fbank_aidatatang_200zh.py

@@ -29,7 +29,7 @@ import os
 from pathlib import Path
 
 import torch
-from lhotse import CutSet, Fbank, FbankConfig, LilcomHdf5Writer
+from lhotse import ChunkedLilcomHdf5Writer, CutSet, Fbank, FbankConfig
 from lhotse.recipes.utils import read_manifests_if_cached
 
 from icefall.utils import get_executor
@@ -81,7 +81,7 @@ def compute_fbank_aidatatang_200zh(num_mel_bins: int = 80):
                 # when an executor is specified, make more partitions
                 num_jobs=num_jobs if ex is None else 80,
                 executor=ex,
-                storage_type=LilcomHdf5Writer,
+                storage_type=ChunkedLilcomHdf5Writer,
             )
             cut_set.to_json(output_dir / f"cuts_{partition}.json.gz")
 

egs/alimeeting/ASR/README.md

@@ -0,0 +1,19 @@
+
+# Introduction
+
+This recipe includes some different ASR models trained with Alimeeting (far).
+
+[./RESULTS.md](./RESULTS.md) contains the latest results.
+
+# Transducers
+
+There are various folders containing the name `transducer` in this folder.
+The following table lists the differences among them.
+
+|                                       | Encoder             | Decoder            | Comment                     |
+|---------------------------------------|---------------------|--------------------|-----------------------------|
+| `pruned_transducer_stateless2`        | Conformer(modified) | Embedding + Conv1d | Using k2 pruned RNN-T loss  |                      |
+
+The decoder in `transducer_stateless` is modified from the paper
+[Rnn-Transducer with Stateless Prediction Network](https://ieeexplore.ieee.org/document/9054419/).
+We place an additional Conv1d layer right after the input embedding layer.

egs/alimeeting/ASR/RESULTS.md

@@ -0,0 +1,71 @@
+## Results
+
+### Alimeeting Char training results (Pruned Transducer Stateless2)
+
+#### 2022-06-01
+
+Using the codes from this PR https://github.com/k2-fsa/icefall/pull/378.
+
+The WERs are
+|                                    |     eval   |    test    | comment                                  |
+|------------------------------------|------------|------------|------------------------------------------|
+|          greedy search             | 31.77      | 34.66      | --epoch 29, --avg 18, --max-duration 100 |
+| modified beam search (beam size 4) | 30.38      | 33.02      | --epoch 29, --avg 18, --max-duration 100 |
+| fast beam search (set as default)  | 31.39      | 34.25      | --epoch 29, --avg 18, --max-duration 1500|
+
+The training command for reproducing is given below:
+
+```
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless2/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --lang-dir data/lang_char \
+  --max-duration 220 \
+  --save-every-n 1000
+
+```
+
+The tensorboard training log can be found at
+https://tensorboard.dev/experiment/AoqgSvZKTZCJhJbOuG3W6g/#scalars
+
+The decoding command is:
+```
+epoch=29
+avg=18
+
+## greedy search
+./pruned_transducer_stateless2/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --lang-dir ./data/lang_char \
+  --max-duration 100
+
+## modified beam search
+./pruned_transducer_stateless2/decode.py \
+  --epoch $epoch \
+  --avg $avg \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --lang-dir ./data/lang_char \
+  --max-duration 100 \
+  --decoding-method modified_beam_search \
+  --beam-size 4
+
+## fast beam search
+./pruned_transducer_stateless2/decode.py \
+        --epoch $epoch \
+        --avg $avg \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir ./data/lang_char \
+        --max-duration 1500 \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8
+```
+
+A pre-trained model and decoding logs can be found at <https://huggingface.co/luomingshuang/icefall_asr_alimeeting_pruned_transducer_stateless2>

egs/alimeeting/ASR/local/compute_fbank_alimeeting.py

@@ -0,0 +1,120 @@
+#!/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.
+
+
+"""
+This file computes fbank features of the aishell dataset.
+It looks for manifests in the directory data/manifests.
+
+The generated fbank features are saved in data/fbank.
+"""
+
+import argparse
+import logging
+import os
+from pathlib import Path
+
+import torch
+from lhotse import ChunkedLilcomHdf5Writer, CutSet, Fbank, FbankConfig
+from lhotse.recipes.utils import read_manifests_if_cached
+
+from icefall.utils import get_executor
+
+# Torch's multithreaded behavior needs to be disabled or
+# it wastes a lot of CPU and slow things down.
+# Do this outside of main() in case it needs to take effect
+# even when we are not invoking the main (e.g. when spawning subprocesses).
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+
+def compute_fbank_alimeeting(num_mel_bins: int = 80):
+    src_dir = Path("data/manifests/alimeeting")
+    output_dir = Path("data/fbank")
+    num_jobs = min(15, os.cpu_count())
+
+    dataset_parts = (
+        "train",
+        "eval",
+        "test",
+    )
+    manifests = read_manifests_if_cached(
+        dataset_parts=dataset_parts,
+        output_dir=src_dir,
+        suffix="jsonl.gz",
+    )
+    assert manifests is not None
+
+    extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins))
+
+    with get_executor() as ex:  # Initialize the executor only once.
+        for partition, m in manifests.items():
+            if (output_dir / f"cuts_{partition}.json.gz").is_file():
+                logging.info(f"{partition} already exists - skipping.")
+                continue
+            logging.info(f"Processing {partition}")
+            cut_set = CutSet.from_manifests(
+                recordings=m["recordings"],
+                supervisions=m["supervisions"],
+            )
+            if "train" in partition:
+                cut_set = (
+                    cut_set
+                    + cut_set.perturb_speed(0.9)
+                    + cut_set.perturb_speed(1.1)
+                )
+            cur_num_jobs = num_jobs if ex is None else 80
+            cur_num_jobs = min(cur_num_jobs, len(cut_set))
+
+            cut_set = cut_set.compute_and_store_features(
+                extractor=extractor,
+                storage_path=f"{output_dir}/feats_{partition}",
+                # when an executor is specified, make more partitions
+                num_jobs=cur_num_jobs,
+                executor=ex,
+                storage_type=ChunkedLilcomHdf5Writer,
+            )
+
+            logging.info("About splitting cuts into smaller chunks")
+            cut_set = cut_set.trim_to_supervisions(
+                keep_overlapping=False,
+                min_duration=None,
+            )
+            cut_set.to_json(output_dir / f"cuts_{partition}.json.gz")
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--num-mel-bins",
+        type=int,
+        default=80,
+        help="""The number of mel bins for Fbank""",
+    )
+
+    return parser.parse_args()
+
+
+if __name__ == "__main__":
+    formatter = (
+        "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
+    )
+
+    logging.basicConfig(format=formatter, level=logging.INFO)
+
+    args = get_args()
+    compute_fbank_alimeeting(num_mel_bins=args.num_mel_bins)

egs/alimeeting/ASR/local/compute_fbank_musan.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/local/compute_fbank_musan.py

egs/alimeeting/ASR/local/display_manifest_statistics.py

@@ -0,0 +1,96 @@
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang
+# 						                           Mingshuang 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.
+
+"""
+This file displays duration statistics of utterances in a manifest.
+You can use the displayed value to choose minimum/maximum duration
+to remove short and long utterances during the training.
+See the function `remove_short_and_long_utt()`
+in ../../../librispeech/ASR/transducer/train.py
+for usage.
+"""
+
+
+from lhotse import load_manifest
+
+
+def main():
+    paths = [
+        "./data/fbank/cuts_train.json.gz",
+        "./data/fbank/cuts_eval.json.gz",
+        "./data/fbank/cuts_test.json.gz",
+    ]
+
+    for path in paths:
+        print(f"Starting display the statistics for {path}")
+        cuts = load_manifest(path)
+        cuts.describe()
+
+
+if __name__ == "__main__":
+    main()
+
+"""
+Starting display the statistics for ./data/fbank/cuts_train.json.gz
+Cuts count: 559092
+Total duration (hours): 424.6
+Speech duration (hours): 424.6 (100.0%)
+***
+Duration statistics (seconds):
+mean    2.7
+std     3.0
+min     0.0
+25%     0.7
+50%     1.7
+75%     3.6
+99%     13.6
+99.5%   14.7
+99.9%   16.2
+max     284.3
+Starting display the statistics for ./data/fbank/cuts_eval.json.gz
+Cuts count: 6457
+Total duration (hours): 4.9
+Speech duration (hours): 4.9 (100.0%)
+***
+Duration statistics (seconds):
+mean    2.7
+std     3.1
+min     0.1
+25%     0.6
+50%     1.6
+75%     3.5
+99%     13.6
+99.5%   14.1
+99.9%   14.7
+max     15.8
+Starting display the statistics for ./data/fbank/cuts_test.json.gz
+Cuts count: 16358
+Total duration (hours): 12.5
+Speech duration (hours): 12.5 (100.0%)
+***
+Duration statistics (seconds):
+mean    2.7
+std     2.9
+min     0.1
+25%     0.7
+50%     1.7
+75%     3.5
+99%     13.7
+99.5%   14.2
+99.9%   14.8
+max     15.7
+"""

egs/alimeeting/ASR/local/prepare_char.py

@@ -0,0 +1,248 @@
+#!/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.
+
+
+"""
+
+This script takes as input `lang_dir`, which should contain::
+
+    - lang_dir/text,
+    - lang_dir/words.txt
+
+and generates the following files in the directory `lang_dir`:
+
+    - lexicon.txt
+    - lexicon_disambig.txt
+    - L.pt
+    - L_disambig.pt
+    - tokens.txt
+"""
+
+import re
+from pathlib import Path
+from typing import Dict, List
+
+import k2
+import torch
+from prepare_lang import (
+    Lexicon,
+    add_disambig_symbols,
+    add_self_loops,
+    write_lexicon,
+    write_mapping,
+)
+
+
+def lexicon_to_fst_no_sil(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format).
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    loop_state = 0  # words enter and leave from here
+    next_state = 1  # the next un-allocated state, will be incremented as we go
+
+    arcs = []
+
+    # The blank symbol <blk> is defined in local/train_bpe_model.py
+    assert token2id["<blk>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    for word, pieces in lexicon:
+        assert len(pieces) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        pieces = [
+            token2id[i] if i in token2id else token2id["<unk>"] for i in pieces
+        ]
+
+        for i in range(len(pieces) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, pieces[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last piece of this word
+        i = len(pieces) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, pieces[i], w, 0])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def contain_oov(token_sym_table: Dict[str, int], tokens: List[str]) -> bool:
+    """Check if all the given tokens are in token symbol table.
+
+    Args:
+      token_sym_table:
+        Token symbol table that contains all the valid tokens.
+      tokens:
+        A list of tokens.
+    Returns:
+      Return True if there is any token not in the token_sym_table,
+      otherwise False.
+    """
+    for tok in tokens:
+        if tok not in token_sym_table:
+            return True
+    return False
+
+
+def generate_lexicon(
+    token_sym_table: Dict[str, int], words: List[str]
+) -> Lexicon:
+    """Generate a lexicon from a word list and token_sym_table.
+
+    Args:
+      token_sym_table:
+        Token symbol table that mapping token to token ids.
+      words:
+        A list of strings representing words.
+    Returns:
+      Return a dict whose keys are words and values are the corresponding
+          tokens.
+    """
+    lexicon = []
+    for word in words:
+        chars = list(word.strip(" \t"))
+        if contain_oov(token_sym_table, chars):
+            continue
+        lexicon.append((word, chars))
+
+    # The OOV word is <UNK>
+    lexicon.append(("<UNK>", ["<unk>"]))
+    return lexicon
+
+
+def generate_tokens(text_file: str) -> Dict[str, int]:
+    """Generate tokens from the given text file.
+
+    Args:
+      text_file:
+        A file that contains text lines to generate tokens.
+    Returns:
+      Return a dict whose keys are tokens and values are token ids ranged
+      from 0 to len(keys) - 1.
+    """
+    tokens: Dict[str, int] = dict()
+    tokens["<blk>"] = 0
+    tokens["<sos/eos>"] = 1
+    tokens["<unk>"] = 2
+    whitespace = re.compile(r"([ \t\r\n]+)")
+    with open(text_file, "r", encoding="utf-8") as f:
+        for line in f:
+            line = re.sub(whitespace, "", line)
+            chars = list(line)
+            for char in chars:
+                if char not in tokens:
+                    tokens[char] = len(tokens)
+    return tokens
+
+
+def main():
+    lang_dir = Path("data/lang_char")
+    text_file = lang_dir / "text"
+
+    word_sym_table = k2.SymbolTable.from_file(lang_dir / "words.txt")
+
+    words = word_sym_table.symbols
+
+    excluded = ["<eps>", "!SIL", "<SPOKEN_NOISE>", "<UNK>", "#0", "<s>", "</s>"]
+    for w in excluded:
+        if w in words:
+            words.remove(w)
+
+    token_sym_table = generate_tokens(text_file)
+
+    lexicon = generate_lexicon(token_sym_table, words)
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    next_token_id = max(token_sym_table.values()) + 1
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in token_sym_table
+        token_sym_table[disambig] = next_token_id
+        next_token_id += 1
+
+    word_sym_table.add("#0")
+    word_sym_table.add("<s>")
+    word_sym_table.add("</s>")
+
+    write_mapping(lang_dir / "tokens.txt", token_sym_table)
+
+    write_lexicon(lang_dir / "lexicon.txt", lexicon)
+    write_lexicon(lang_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst_no_sil(
+        lexicon,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+    )
+
+    L_disambig = lexicon_to_fst_no_sil(
+        lexicon_disambig,
+        token2id=token_sym_table,
+        word2id=word_sym_table,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), lang_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), lang_dir / "L_disambig.pt")
+
+
+if __name__ == "__main__":
+    main()

egs/alimeeting/ASR/local/prepare_lang.py

@@ -0,0 +1,390 @@
+#!/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.
+
+
+"""
+This script takes as input a lexicon file "data/lang_phone/lexicon.txt"
+consisting of words and tokens (i.e., phones) and does the following:
+
+1. Add disambiguation symbols to the lexicon and generate lexicon_disambig.txt
+
+2. Generate tokens.txt, the token table mapping a token to a unique integer.
+
+3. Generate words.txt, the word table mapping a word to a unique integer.
+
+4. Generate L.pt, in k2 format. It can be loaded by
+
+        d = torch.load("L.pt")
+        lexicon = k2.Fsa.from_dict(d)
+
+5. Generate L_disambig.pt, in k2 format.
+"""
+import argparse
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Any, Dict, List, Tuple
+
+import k2
+import torch
+
+from icefall.lexicon import read_lexicon, write_lexicon
+
+Lexicon = List[Tuple[str, List[str]]]
+
+
+def write_mapping(filename: str, sym2id: Dict[str, int]) -> None:
+    """Write a symbol to ID mapping to a file.
+
+    Note:
+      No need to implement `read_mapping` as it can be done
+      through :func:`k2.SymbolTable.from_file`.
+
+    Args:
+      filename:
+        Filename to save the mapping.
+      sym2id:
+        A dict mapping symbols to IDs.
+    Returns:
+      Return None.
+    """
+    with open(filename, "w", encoding="utf-8") as f:
+        for sym, i in sym2id.items():
+            f.write(f"{sym} {i}\n")
+
+
+def get_tokens(lexicon: Lexicon) -> List[str]:
+    """Get tokens from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique tokens.
+    """
+    ans = set()
+    for _, tokens in lexicon:
+        ans.update(tokens)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def get_words(lexicon: Lexicon) -> List[str]:
+    """Get words from a lexicon.
+
+    Args:
+      lexicon:
+        It is the return value of :func:`read_lexicon`.
+    Returns:
+      Return a list of unique words.
+    """
+    ans = set()
+    for word, _ in lexicon:
+        ans.add(word)
+    sorted_ans = sorted(list(ans))
+    return sorted_ans
+
+
+def add_disambig_symbols(lexicon: Lexicon) -> Tuple[Lexicon, int]:
+    """It adds pseudo-token disambiguation symbols #1, #2 and so on
+    at the ends of tokens to ensure that all pronunciations are different,
+    and that none is a prefix of another.
+
+    See also add_lex_disambig.pl from kaldi.
+
+    Args:
+      lexicon:
+        It is returned by :func:`read_lexicon`.
+    Returns:
+      Return a tuple with two elements:
+
+        - The output lexicon with disambiguation symbols
+        - The ID of the max disambiguation symbol that appears
+          in the lexicon
+    """
+
+    # (1) Work out the count of each token-sequence in the
+    # lexicon.
+    count = defaultdict(int)
+    for _, tokens in lexicon:
+        count[" ".join(tokens)] += 1
+
+    # (2) For each left sub-sequence of each token-sequence, note down
+    # that it exists (for identifying prefixes of longer strings).
+    issubseq = defaultdict(int)
+    for _, tokens in lexicon:
+        tokens = tokens.copy()
+        tokens.pop()
+        while tokens:
+            issubseq[" ".join(tokens)] = 1
+            tokens.pop()
+
+    # (3) For each entry in the lexicon:
+    # if the token sequence is unique and is not a
+    # prefix of another word, no disambig symbol.
+    # Else output #1, or #2, #3, ... if the same token-seq
+    # has already been assigned a disambig symbol.
+    ans = []
+
+    # We start with #1 since #0 has its own purpose
+    first_allowed_disambig = 1
+    max_disambig = first_allowed_disambig - 1
+    last_used_disambig_symbol_of = defaultdict(int)
+
+    for word, tokens in lexicon:
+        tokenseq = " ".join(tokens)
+        assert tokenseq != ""
+        if issubseq[tokenseq] == 0 and count[tokenseq] == 1:
+            ans.append((word, tokens))
+            continue
+
+        cur_disambig = last_used_disambig_symbol_of[tokenseq]
+        if cur_disambig == 0:
+            cur_disambig = first_allowed_disambig
+        else:
+            cur_disambig += 1
+
+        if cur_disambig > max_disambig:
+            max_disambig = cur_disambig
+        last_used_disambig_symbol_of[tokenseq] = cur_disambig
+        tokenseq += f" #{cur_disambig}"
+        ans.append((word, tokenseq.split()))
+    return ans, max_disambig
+
+
+def generate_id_map(symbols: List[str]) -> Dict[str, int]:
+    """Generate ID maps, i.e., map a symbol to a unique ID.
+
+    Args:
+      symbols:
+        A list of unique symbols.
+    Returns:
+      A dict containing the mapping between symbols and IDs.
+    """
+    return {sym: i for i, sym in enumerate(symbols)}
+
+
+def add_self_loops(
+    arcs: List[List[Any]], disambig_token: int, disambig_word: int
+) -> List[List[Any]]:
+    """Adds self-loops to states of an FST to propagate disambiguation symbols
+    through it. They are added on each state with non-epsilon output symbols
+    on at least one arc out of the state.
+
+    See also fstaddselfloops.pl from Kaldi. One difference is that
+    Kaldi uses OpenFst style FSTs and it has multiple final states.
+    This function uses k2 style FSTs and it does not need to add self-loops
+    to the final state.
+
+    The input label of a self-loop is `disambig_token`, while the output
+    label is `disambig_word`.
+
+    Args:
+      arcs:
+        A list-of-list. The sublist contains
+        `[src_state, dest_state, label, aux_label, score]`
+      disambig_token:
+        It is the token ID of the symbol `#0`.
+      disambig_word:
+        It is the word ID of the symbol `#0`.
+
+    Return:
+      Return new `arcs` containing self-loops.
+    """
+    states_needs_self_loops = set()
+    for arc in arcs:
+        src, dst, ilabel, olabel, score = arc
+        if olabel != 0:
+            states_needs_self_loops.add(src)
+
+    ans = []
+    for s in states_needs_self_loops:
+        ans.append([s, s, disambig_token, disambig_word, 0])
+
+    return arcs + ans
+
+
+def lexicon_to_fst(
+    lexicon: Lexicon,
+    token2id: Dict[str, int],
+    word2id: Dict[str, int],
+    sil_token: str = "SIL",
+    sil_prob: float = 0.5,
+    need_self_loops: bool = False,
+) -> k2.Fsa:
+    """Convert a lexicon to an FST (in k2 format) with optional silence at
+    the beginning and end of each word.
+
+    Args:
+      lexicon:
+        The input lexicon. See also :func:`read_lexicon`
+      token2id:
+        A dict mapping tokens to IDs.
+      word2id:
+        A dict mapping words to IDs.
+      sil_token:
+        The silence token.
+      sil_prob:
+        The probability for adding a silence at the beginning and end
+        of the word.
+      need_self_loops:
+        If True, add self-loop to states with non-epsilon output symbols
+        on at least one arc out of the state. The input label for this
+        self loop is `token2id["#0"]` and the output label is `word2id["#0"]`.
+    Returns:
+      Return an instance of `k2.Fsa` representing the given lexicon.
+    """
+    assert sil_prob > 0.0 and sil_prob < 1.0
+    # CAUTION: we use score, i.e, negative cost.
+    sil_score = math.log(sil_prob)
+    no_sil_score = math.log(1.0 - sil_prob)
+
+    start_state = 0
+    loop_state = 1  # words enter and leave from here
+    sil_state = 2  # words terminate here when followed by silence; this state
+    # has a silence transition to loop_state.
+    next_state = 3  # the next un-allocated state, will be incremented as we go.
+    arcs = []
+
+    assert token2id["<eps>"] == 0
+    assert word2id["<eps>"] == 0
+
+    eps = 0
+
+    sil_token = token2id[sil_token]
+
+    arcs.append([start_state, loop_state, eps, eps, no_sil_score])
+    arcs.append([start_state, sil_state, eps, eps, sil_score])
+    arcs.append([sil_state, loop_state, sil_token, eps, 0])
+
+    for word, tokens in lexicon:
+        assert len(tokens) > 0, f"{word} has no pronunciations"
+        cur_state = loop_state
+
+        word = word2id[word]
+        tokens = [token2id[i] for i in tokens]
+
+        for i in range(len(tokens) - 1):
+            w = word if i == 0 else eps
+            arcs.append([cur_state, next_state, tokens[i], w, 0])
+
+            cur_state = next_state
+            next_state += 1
+
+        # now for the last token of this word
+        # It has two out-going arcs, one to the loop state,
+        # the other one to the sil_state.
+        i = len(tokens) - 1
+        w = word if i == 0 else eps
+        arcs.append([cur_state, loop_state, tokens[i], w, no_sil_score])
+        arcs.append([cur_state, sil_state, tokens[i], w, sil_score])
+
+    if need_self_loops:
+        disambig_token = token2id["#0"]
+        disambig_word = word2id["#0"]
+        arcs = add_self_loops(
+            arcs,
+            disambig_token=disambig_token,
+            disambig_word=disambig_word,
+        )
+
+    final_state = next_state
+    arcs.append([loop_state, final_state, -1, -1, 0])
+    arcs.append([final_state])
+
+    arcs = sorted(arcs, key=lambda arc: arc[0])
+    arcs = [[str(i) for i in arc] for arc in arcs]
+    arcs = [" ".join(arc) for arc in arcs]
+    arcs = "\n".join(arcs)
+
+    fsa = k2.Fsa.from_str(arcs, acceptor=False)
+    return fsa
+
+
+def get_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--lang-dir", type=str, help="The lang dir, data/lang_phone"
+    )
+    return parser.parse_args()
+
+
+def main():
+    out_dir = Path(get_args().lang_dir)
+    lexicon_filename = out_dir / "lexicon.txt"
+    sil_token = "SIL"
+    sil_prob = 0.5
+
+    lexicon = read_lexicon(lexicon_filename)
+    tokens = get_tokens(lexicon)
+    words = get_words(lexicon)
+
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+
+    for i in range(max_disambig + 1):
+        disambig = f"#{i}"
+        assert disambig not in tokens
+        tokens.append(f"#{i}")
+
+    assert "<eps>" not in tokens
+    tokens = ["<eps>"] + tokens
+
+    assert "<eps>" not in words
+    assert "#0" not in words
+    assert "<s>" not in words
+    assert "</s>" not in words
+
+    words = ["<eps>"] + words + ["#0", "<s>", "</s>"]
+
+    token2id = generate_id_map(tokens)
+    word2id = generate_id_map(words)
+
+    write_mapping(out_dir / "tokens.txt", token2id)
+    write_mapping(out_dir / "words.txt", word2id)
+    write_lexicon(out_dir / "lexicon_disambig.txt", lexicon_disambig)
+
+    L = lexicon_to_fst(
+        lexicon,
+        token2id=token2id,
+        word2id=word2id,
+        sil_token=sil_token,
+        sil_prob=sil_prob,
+    )
+
+    L_disambig = lexicon_to_fst(
+        lexicon_disambig,
+        token2id=token2id,
+        word2id=word2id,
+        sil_token=sil_token,
+        sil_prob=sil_prob,
+        need_self_loops=True,
+    )
+    torch.save(L.as_dict(), out_dir / "L.pt")
+    torch.save(L_disambig.as_dict(), out_dir / "L_disambig.pt")
+
+    if False:
+        # Just for debugging, will remove it
+        L.labels_sym = k2.SymbolTable.from_file(out_dir / "tokens.txt")
+        L.aux_labels_sym = k2.SymbolTable.from_file(out_dir / "words.txt")
+        L_disambig.labels_sym = L.labels_sym
+        L_disambig.aux_labels_sym = L.aux_labels_sym
+        L.draw(out_dir / "L.png", title="L")
+        L_disambig.draw(out_dir / "L_disambig.png", title="L_disambig")
+
+
+if __name__ == "__main__":
+    main()

egs/alimeeting/ASR/local/prepare_words.py

@@ -0,0 +1,84 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright    2021  Xiaomi Corp.        (authors: Mingshuang 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.
+
+
+"""
+This script takes as input words.txt without ids:
+    - words_no_ids.txt
+and generates the new words.txt with related ids.
+    - words.txt
+"""
+
+
+import argparse
+import logging
+
+from tqdm import tqdm
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        description="Prepare words.txt",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument(
+        "--input-file",
+        default="data/lang_char/words_no_ids.txt",
+        type=str,
+        help="the words file without ids for WenetSpeech",
+    )
+    parser.add_argument(
+        "--output-file",
+        default="data/lang_char/words.txt",
+        type=str,
+        help="the words file with ids for WenetSpeech",
+    )
+
+    return parser
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    input_file = args.input_file
+    output_file = args.output_file
+
+    f = open(input_file, "r", encoding="utf-8")
+    lines = f.readlines()
+    new_lines = []
+    add_words = ["<eps> 0", "!SIL 1", "<SPOKEN_NOISE> 2", "<UNK> 3"]
+    new_lines.extend(add_words)
+
+    logging.info("Starting reading the input file")
+    for i in tqdm(range(len(lines))):
+        x = lines[i]
+        idx = 4 + i
+        new_line = str(x.strip("\n")) + " " + str(idx)
+        new_lines.append(new_line)
+
+    logging.info("Starting writing the words.txt")
+    f_out = open(output_file, "w", encoding="utf-8")
+    for line in new_lines:
+        f_out.write(line)
+        f_out.write("\n")
+
+
+if __name__ == "__main__":
+    main()

egs/alimeeting/ASR/local/test_prepare_lang.py

@@ -0,0 +1,106 @@
+#!/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.
+
+
+# Copyright (c)  2021  Xiaomi Corporation (authors: Fangjun Kuang)
+
+import os
+import tempfile
+
+import k2
+from prepare_lang import (
+    add_disambig_symbols,
+    generate_id_map,
+    get_phones,
+    get_words,
+    lexicon_to_fst,
+    read_lexicon,
+    write_lexicon,
+    write_mapping,
+)
+
+
+def generate_lexicon_file() -> str:
+    fd, filename = tempfile.mkstemp()
+    os.close(fd)
+    s = """
+    !SIL SIL
+    <SPOKEN_NOISE> SPN
+    <UNK> SPN
+    f f
+    a a
+    foo f o o
+    bar b a r
+    bark b a r k
+    food f o o d
+    food2 f o o d
+    fo  f o
+    """.strip()
+    with open(filename, "w") as f:
+        f.write(s)
+    return filename
+
+
+def test_read_lexicon(filename: str):
+    lexicon = read_lexicon(filename)
+    phones = get_phones(lexicon)
+    words = get_words(lexicon)
+    print(lexicon)
+    print(phones)
+    print(words)
+    lexicon_disambig, max_disambig = add_disambig_symbols(lexicon)
+    print(lexicon_disambig)
+    print("max disambig:", f"#{max_disambig}")
+
+    phones = ["<eps>", "SIL", "SPN"] + phones
+    for i in range(max_disambig + 1):
+        phones.append(f"#{i}")
+    words = ["<eps>"] + words
+
+    phone2id = generate_id_map(phones)
+    word2id = generate_id_map(words)
+
+    print(phone2id)
+    print(word2id)
+
+    write_mapping("phones.txt", phone2id)
+    write_mapping("words.txt", word2id)
+
+    write_lexicon("a.txt", lexicon)
+    write_lexicon("a_disambig.txt", lexicon_disambig)
+
+    fsa = lexicon_to_fst(lexicon, phone2id=phone2id, word2id=word2id)
+    fsa.labels_sym = k2.SymbolTable.from_file("phones.txt")
+    fsa.aux_labels_sym = k2.SymbolTable.from_file("words.txt")
+    fsa.draw("L.pdf", title="L")
+
+    fsa_disambig = lexicon_to_fst(
+        lexicon_disambig, phone2id=phone2id, word2id=word2id
+    )
+    fsa_disambig.labels_sym = k2.SymbolTable.from_file("phones.txt")
+    fsa_disambig.aux_labels_sym = k2.SymbolTable.from_file("words.txt")
+    fsa_disambig.draw("L_disambig.pdf", title="L_disambig")
+
+
+def main():
+    filename = generate_lexicon_file()
+    test_read_lexicon(filename)
+    os.remove(filename)
+
+
+if __name__ == "__main__":
+    main()

egs/alimeeting/ASR/local/text2segments.py

@@ -0,0 +1,83 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+
+# Copyright    2021  Xiaomi Corp.        (authors: Mingshuang 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.
+
+
+"""
+This script takes as input "text", which refers to the transcript file for
+WenetSpeech:
+    - text
+and generates the output file text_word_segmentation which is implemented
+with word segmenting:
+    - text_words_segmentation
+"""
+
+
+import argparse
+
+import jieba
+from tqdm import tqdm
+
+jieba.enable_paddle()
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        description="Chinese Word Segmentation for text",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument(
+        "--input-file",
+        default="data/lang_char/text",
+        type=str,
+        help="the input text file for WenetSpeech",
+    )
+    parser.add_argument(
+        "--output-file",
+        default="data/lang_char/text_words_segmentation",
+        type=str,
+        help="the text implemented with words segmenting for WenetSpeech",
+    )
+
+    return parser
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    input_file = args.input_file
+    output_file = args.output_file
+
+    f = open(input_file, "r", encoding="utf-8")
+    lines = f.readlines()
+    new_lines = []
+    for i in tqdm(range(len(lines))):
+        x = lines[i].rstrip()
+        seg_list = jieba.cut(x, use_paddle=True)
+        new_line = " ".join(seg_list)
+        new_lines.append(new_line)
+
+    f_new = open(output_file, "w", encoding="utf-8")
+    for line in new_lines:
+        f_new.write(line)
+        f_new.write("\n")
+
+
+if __name__ == "__main__":
+    main()

egs/alimeeting/ASR/local/text2token.py

@@ -0,0 +1,195 @@
+#!/usr/bin/env python3
+# Copyright    2017  Johns Hopkins University   (authors: Shinji Watanabe)
+#              2022  Xiaomi Corp.               (authors: Mingshuang 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.
+
+
+import argparse
+import codecs
+import re
+import sys
+from typing import List
+
+from pypinyin import lazy_pinyin, pinyin
+
+is_python2 = sys.version_info[0] == 2
+
+
+def exist_or_not(i, match_pos):
+    start_pos = None
+    end_pos = None
+    for pos in match_pos:
+        if pos[0] <= i < pos[1]:
+            start_pos = pos[0]
+            end_pos = pos[1]
+            break
+
+    return start_pos, end_pos
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        description="convert raw text to tokenized text",
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument(
+        "--nchar",
+        "-n",
+        default=1,
+        type=int,
+        help="number of characters to split, i.e., \
+                        aabb -> a a b b with -n 1 and aa bb with -n 2",
+    )
+    parser.add_argument(
+        "--skip-ncols", "-s", default=0, type=int, help="skip first n columns"
+    )
+    parser.add_argument(
+        "--space", default="<space>", type=str, help="space symbol"
+    )
+    parser.add_argument(
+        "--non-lang-syms",
+        "-l",
+        default=None,
+        type=str,
+        help="list of non-linguistic symobles, e.g., <NOISE> etc.",
+    )
+    parser.add_argument(
+        "text", type=str, default=False, nargs="?", help="input text"
+    )
+    parser.add_argument(
+        "--trans_type",
+        "-t",
+        type=str,
+        default="char",
+        choices=["char", "pinyin", "lazy_pinyin"],
+        help="""Transcript type. char/pinyin/lazy_pinyin""",
+    )
+    return parser
+
+
+def token2id(
+    texts, token_table, token_type: str = "lazy_pinyin", oov: str = "<unk>"
+) -> List[List[int]]:
+    """Convert token to id.
+    Args:
+      texts:
+        The input texts, it refers to the chinese text here.
+      token_table:
+        The token table is built based on "data/lang_xxx/token.txt"
+      token_type:
+        The type of token, such as "pinyin" and "lazy_pinyin".
+      oov:
+        Out of vocabulary token. When a word(token) in the transcript
+        does not exist in the token list, it is replaced with `oov`.
+
+    Returns:
+      The list of ids for the input texts.
+    """
+    if texts is None:
+        raise ValueError("texts can't be None!")
+    else:
+        oov_id = token_table[oov]
+        ids: List[List[int]] = []
+        for text in texts:
+            chars_list = list(str(text))
+            if token_type == "lazy_pinyin":
+                text = lazy_pinyin(chars_list)
+                sub_ids = [
+                    token_table[txt] if txt in token_table else oov_id
+                    for txt in text
+                ]
+                ids.append(sub_ids)
+            else:  # token_type = "pinyin"
+                text = pinyin(chars_list)
+                sub_ids = [
+                    token_table[txt[0]] if txt[0] in token_table else oov_id
+                    for txt in text
+                ]
+                ids.append(sub_ids)
+        return ids
+
+
+def main():
+    parser = get_parser()
+    args = parser.parse_args()
+
+    rs = []
+    if args.non_lang_syms is not None:
+        with codecs.open(args.non_lang_syms, "r", encoding="utf-8") as f:
+            nls = [x.rstrip() for x in f.readlines()]
+            rs = [re.compile(re.escape(x)) for x in nls]
+
+    if args.text:
+        f = codecs.open(args.text, encoding="utf-8")
+    else:
+        f = codecs.getreader("utf-8")(
+            sys.stdin if is_python2 else sys.stdin.buffer
+        )
+
+    sys.stdout = codecs.getwriter("utf-8")(
+        sys.stdout if is_python2 else sys.stdout.buffer
+    )
+    line = f.readline()
+    n = args.nchar
+    while line:
+        x = line.split()
+        print(" ".join(x[: args.skip_ncols]), end=" ")
+        a = " ".join(x[args.skip_ncols :])  # noqa E203
+
+        # get all matched positions
+        match_pos = []
+        for r in rs:
+            i = 0
+            while i >= 0:
+                m = r.search(a, i)
+                if m:
+                    match_pos.append([m.start(), m.end()])
+                    i = m.end()
+                else:
+                    break
+        if len(match_pos) > 0:
+            chars = []
+            i = 0
+            while i < len(a):
+                start_pos, end_pos = exist_or_not(i, match_pos)
+                if start_pos is not None:
+                    chars.append(a[start_pos:end_pos])
+                    i = end_pos
+                else:
+                    chars.append(a[i])
+                    i += 1
+            a = chars
+
+        if args.trans_type == "pinyin":
+            a = pinyin(list(str(a)))
+            a = [one[0] for one in a]
+
+        if args.trans_type == "lazy_pinyin":
+            a = lazy_pinyin(list(str(a)))
+
+        a = [a[j : j + n] for j in range(0, len(a), n)]  # noqa E203
+
+        a_flat = []
+        for z in a:
+            a_flat.append("".join(z))
+
+        a_chars = "".join(a_flat)
+        print(a_chars)
+        line = f.readline()
+
+
+if __name__ == "__main__":
+    main()

egs/alimeeting/ASR/prepare.sh

@@ -0,0 +1,133 @@
+#!/usr/bin/env bash
+
+set -eou pipefail
+
+stage=-1
+stop_stage=100
+
+# We assume dl_dir (download dir) contains the following
+# directories and files. If not, they will be downloaded
+# by this script automatically.
+#
+#  - $dl_dir/alimeeting
+#     This directory contains the following files downloaded from
+#       https://openslr.org/62/
+#
+#     - Train_Ali_far.tar.gz
+#     - Train_Ali_near.tar.gz
+#     - Test_Ali.tar.gz
+#     - Eval_Ali.tar.gz
+#
+#  - $dl_dir/musan
+#      This directory contains the following directories downloaded from
+#       http://www.openslr.org/17/
+#
+#     - music
+#     - noise
+#     - speech
+
+dl_dir=$PWD/download
+
+. shared/parse_options.sh || exit 1
+
+# 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
+
+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 0 ] && [ $stop_stage -ge 0 ]; then
+  log "Stage 0: Download data"
+
+  if [ ! -f $dl_dir/alimeeting/Train_Ali_far.tar.gz ]; then
+    lhotse download ali-meeting $dl_dir/alimeeting
+  fi
+fi
+
+if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then
+  log "Stage 1: Prepare alimeeting manifest"
+  # We assume that you have downloaded the alimeeting corpus
+  # to $dl_dir/alimeeting
+  if [ ! -f data/manifests/alimeeting/.manifests.done ]; then
+    mkdir -p data/manifests/alimeeting
+    lhotse prepare ali-meeting $dl_dir/alimeeting data/manifests/alimeeting
+    touch data/manifests/alimeeting/.manifests.done
+  fi
+fi
+
+if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then
+  log "Stage 2: Process alimeeting"
+  if [ ! -f data/fbank/alimeeting/.fbank.done ]; then
+    mkdir -p data/fbank/alimeeting
+    lhotse prepare ali-meeting $dl_dir/alimeeting data/manifests/alimeeting
+  fi
+fi
+
+if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then
+  log "Stage 3: Prepare musan manifest"
+  # We assume that you have downloaded the musan corpus
+  # to data/musan
+  if [ ! -f data/manifests/.musan_manifests.done ]; then
+    log "It may take 6 minutes"
+    mkdir -p data/manifests
+    lhotse prepare musan $dl_dir/musan data/manifests
+    touch data/manifests/.musan_manifests.done
+  fi
+fi
+
+if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then
+  log "Stage 4: Compute fbank for musan"
+  if [ ! -f data/fbank/.msuan.done ]; then
+    mkdir -p data/fbank
+    ./local/compute_fbank_musan.py
+    touch data/fbank/.msuan.done
+  fi
+fi
+
+if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then
+  log "Stage 5: Compute fbank for alimeeting"
+  if [ ! -f data/fbank/.alimeeting.done ]; then
+    mkdir -p data/fbank
+    ./local/compute_fbank_alimeeting.py
+    touch data/fbank/.alimeeting.done
+  fi
+fi
+
+if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then
+  log "Stage 6: Prepare char based lang"
+  lang_char_dir=data/lang_char
+  mkdir -p $lang_char_dir
+
+  # Prepare text.
+  # Note: in Linux, you can install jq with the  following command:
+  # wget -O jq https://github.com/stedolan/jq/releases/download/jq-1.6/jq-linux64
+  gunzip -c data/manifests/alimeeting/supervisions_train.jsonl.gz \
+    | jq ".text" | sed 's/"//g' \
+    | ./local/text2token.py -t "char" > $lang_char_dir/text
+
+  # Prepare words segments
+  python ./local/text2segments.py \
+    --input $lang_char_dir/text \
+    --output $lang_char_dir/text_words_segmentation
+
+  cat $lang_char_dir/text_words_segmentation | sed "s/ /\n/g" \
+    | sort -u | sed "/^$/d" \
+    | uniq > $lang_char_dir/words_no_ids.txt
+
+  # Prepare words.txt
+  if [ ! -f $lang_char_dir/words.txt ]; then
+    ./local/prepare_words.py \
+      --input-file $lang_char_dir/words_no_ids.txt \
+      --output-file $lang_char_dir/words.txt
+  fi
+
+  if [ ! -f $lang_char_dir/L_disambig.pt ]; then
+    ./local/prepare_char.py
+  fi
+fi

egs/alimeeting/ASR/pruned_transducer_stateless2/asr_datamodule.py

@@ -0,0 +1,414 @@
+# Copyright      2021  Piotr Żelasko
+#
+# 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 argparse
+import inspect
+import logging
+from functools import lru_cache
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+
+import torch
+from lhotse import (
+    CutSet,
+    Fbank,
+    FbankConfig,
+    load_manifest,
+    set_caching_enabled,
+)
+from lhotse.dataset import (
+    CutConcatenate,
+    CutMix,
+    DynamicBucketingSampler,
+    K2SpeechRecognitionDataset,
+    PrecomputedFeatures,
+    SingleCutSampler,
+    SpecAugment,
+)
+from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from lhotse.utils import fix_random_seed
+from torch.utils.data import DataLoader
+
+from icefall.utils import str2bool
+
+set_caching_enabled(False)
+torch.set_num_threads(1)
+
+
+class _SeedWorkers:
+    def __init__(self, seed: int):
+        self.seed = seed
+
+    def __call__(self, worker_id: int):
+        fix_random_seed(self.seed + worker_id)
+
+
+class AlimeetingAsrDataModule:
+    """
+    DataModule for k2 ASR experiments.
+    It assumes there is always one train and valid dataloader,
+    but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
+    and test-other).
+    It contains all the common data pipeline modules used in ASR
+    experiments, e.g.:
+    - dynamic batch size,
+    - bucketing samplers,
+    - cut concatenation,
+    - augmentation,
+    - on-the-fly feature extraction
+    This class should be derived for specific corpora used in ASR tasks.
+    """
+
+    def __init__(self, args: argparse.Namespace):
+        self.args = args
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group(
+            title="ASR data related options",
+            description="These options are used for the preparation of "
+            "PyTorch DataLoaders from Lhotse CutSet's -- they control the "
+            "effective batch sizes, sampling strategies, applied data "
+            "augmentations, etc.",
+        )
+        group.add_argument(
+            "--manifest-dir",
+            type=Path,
+            default=Path("data/fbank"),
+            help="Path to directory with train/dev/test cuts.",
+        )
+        group.add_argument(
+            "--max-duration",
+            type=int,
+            default=200.0,
+            help="Maximum pooled recordings duration (seconds) in a "
+            "single batch. You can reduce it if it causes CUDA OOM.",
+        )
+        group.add_argument(
+            "--bucketing-sampler",
+            type=str2bool,
+            default=True,
+            help="When enabled, the batches will come from buckets of "
+            "similar duration (saves padding frames).",
+        )
+        group.add_argument(
+            "--num-buckets",
+            type=int,
+            default=300,
+            help="The number of buckets for the DynamicBucketingSampler"
+            "(you might want to increase it for larger datasets).",
+        )
+        group.add_argument(
+            "--concatenate-cuts",
+            type=str2bool,
+            default=False,
+            help="When enabled, utterances (cuts) will be concatenated "
+            "to minimize the amount of padding.",
+        )
+        group.add_argument(
+            "--duration-factor",
+            type=float,
+            default=1.0,
+            help="Determines the maximum duration of a concatenated cut "
+            "relative to the duration of the longest cut in a batch.",
+        )
+        group.add_argument(
+            "--gap",
+            type=float,
+            default=1.0,
+            help="The amount of padding (in seconds) inserted between "
+            "concatenated cuts. This padding is filled with noise when "
+            "noise augmentation is used.",
+        )
+        group.add_argument(
+            "--on-the-fly-feats",
+            type=str2bool,
+            default=False,
+            help="When enabled, use on-the-fly cut mixing and feature "
+            "extraction. Will drop existing precomputed feature manifests "
+            "if available.",
+        )
+        group.add_argument(
+            "--shuffle",
+            type=str2bool,
+            default=True,
+            help="When enabled (=default), the examples will be "
+            "shuffled for each epoch.",
+        )
+        group.add_argument(
+            "--return-cuts",
+            type=str2bool,
+            default=True,
+            help="When enabled, each batch will have the "
+            "field: batch['supervisions']['cut'] with the cuts that "
+            "were used to construct it.",
+        )
+
+        group.add_argument(
+            "--num-workers",
+            type=int,
+            default=2,
+            help="The number of training dataloader workers that "
+            "collect the batches.",
+        )
+
+        group.add_argument(
+            "--enable-spec-aug",
+            type=str2bool,
+            default=True,
+            help="When enabled, use SpecAugment for training dataset.",
+        )
+
+        group.add_argument(
+            "--spec-aug-time-warp-factor",
+            type=int,
+            default=80,
+            help="Used only when --enable-spec-aug is True. "
+            "It specifies the factor for time warping in SpecAugment. "
+            "Larger values mean more warping. "
+            "A value less than 1 means to disable time warp.",
+        )
+
+        group.add_argument(
+            "--enable-musan",
+            type=str2bool,
+            default=True,
+            help="When enabled, select noise from MUSAN and mix it"
+            "with training dataset. ",
+        )
+
+    def train_dataloaders(
+        self,
+        cuts_train: CutSet,
+        sampler_state_dict: Optional[Dict[str, Any]] = None,
+    ) -> DataLoader:
+        """
+        Args:
+          cuts_train:
+            CutSet for training.
+          sampler_state_dict:
+            The state dict for the training sampler.
+        """
+        logging.info("About to get Musan cuts")
+        cuts_musan = load_manifest(
+            self.args.manifest_dir / "cuts_musan.json.gz"
+        )
+
+        transforms = []
+        if self.args.enable_musan:
+            logging.info("Enable MUSAN")
+            transforms.append(
+                CutMix(
+                    cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True
+                )
+            )
+        else:
+            logging.info("Disable MUSAN")
+
+        if self.args.concatenate_cuts:
+            logging.info(
+                f"Using cut concatenation with duration factor "
+                f"{self.args.duration_factor} and gap {self.args.gap}."
+            )
+            # Cut concatenation should be the first transform in the list,
+            # so that if we e.g. mix noise in, it will fill the gaps between
+            # different utterances.
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        input_transforms = []
+        if self.args.enable_spec_aug:
+            logging.info("Enable SpecAugment")
+            logging.info(
+                f"Time warp factor: {self.args.spec_aug_time_warp_factor}"
+            )
+            # Set the value of num_frame_masks according to Lhotse's version.
+            # In different Lhotse's versions, the default of num_frame_masks is
+            # different.
+            num_frame_masks = 10
+            num_frame_masks_parameter = inspect.signature(
+                SpecAugment.__init__
+            ).parameters["num_frame_masks"]
+            if num_frame_masks_parameter.default == 1:
+                num_frame_masks = 2
+            logging.info(f"Num frame mask: {num_frame_masks}")
+            input_transforms.append(
+                SpecAugment(
+                    time_warp_factor=self.args.spec_aug_time_warp_factor,
+                    num_frame_masks=num_frame_masks,
+                    features_mask_size=27,
+                    num_feature_masks=2,
+                    frames_mask_size=100,
+                )
+            )
+        else:
+            logging.info("Disable SpecAugment")
+
+        logging.info("About to create train dataset")
+        train = K2SpeechRecognitionDataset(
+            cut_transforms=transforms,
+            input_transforms=input_transforms,
+            return_cuts=self.args.return_cuts,
+        )
+
+        if self.args.on_the_fly_feats:
+            # NOTE: the PerturbSpeed transform should be added only if we
+            # remove it from data prep stage.
+            # Add on-the-fly speed perturbation; since originally it would
+            # have increased epoch size by 3, we will apply prob 2/3 and use
+            # 3x more epochs.
+            # Speed perturbation probably should come first before
+            # concatenation, but in principle the transforms order doesn't have
+            # to be strict (e.g. could be randomized)
+            # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms   # noqa
+            # Drop feats to be on the safe side.
+            train = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(
+                    Fbank(FbankConfig(num_mel_bins=80))
+                ),
+                input_transforms=input_transforms,
+                return_cuts=self.args.return_cuts,
+            )
+
+        if self.args.bucketing_sampler:
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+                num_buckets=self.args.num_buckets,
+                buffer_size=30000,
+                drop_last=True,
+            )
+        else:
+            logging.info("Using SingleCutSampler.")
+            train_sampler = SingleCutSampler(
+                cuts_train,
+                max_duration=self.args.max_duration,
+                shuffle=self.args.shuffle,
+            )
+        logging.info("About to create train dataloader")
+
+        # 'seed' is derived from the current random state, which will have
+        # previously been set in the main process.
+        seed = torch.randint(0, 100000, ()).item()
+        worker_init_fn = _SeedWorkers(seed)
+
+        train_dl = DataLoader(
+            train,
+            sampler=train_sampler,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+            worker_init_fn=worker_init_fn,
+        )
+
+        if sampler_state_dict is not None:
+            logging.info("Loading sampler state dict")
+            train_dl.sampler.load_state_dict(sampler_state_dict)
+
+        return train_dl
+
+    def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
+        transforms = []
+        if self.args.concatenate_cuts:
+            transforms = [
+                CutConcatenate(
+                    duration_factor=self.args.duration_factor, gap=self.args.gap
+                )
+            ] + transforms
+
+        logging.info("About to create dev dataset")
+        if self.args.on_the_fly_feats:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                input_strategy=OnTheFlyFeatures(
+                    Fbank(FbankConfig(num_mel_bins=80))
+                ),
+                return_cuts=self.args.return_cuts,
+            )
+        else:
+            validate = K2SpeechRecognitionDataset(
+                cut_transforms=transforms,
+                return_cuts=self.args.return_cuts,
+            )
+        valid_sampler = DynamicBucketingSampler(
+            cuts_valid,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        logging.info("About to create dev dataloader")
+
+        from lhotse.dataset.iterable_dataset import IterableDatasetWrapper
+
+        dev_iter_dataset = IterableDatasetWrapper(
+            dataset=validate,
+            sampler=valid_sampler,
+        )
+        valid_dl = DataLoader(
+            dev_iter_dataset,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+            persistent_workers=False,
+        )
+
+        return valid_dl
+
+    def test_dataloaders(self, cuts: CutSet) -> DataLoader:
+        logging.debug("About to create test dataset")
+        test = K2SpeechRecognitionDataset(
+            input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
+            if self.args.on_the_fly_feats
+            else PrecomputedFeatures(),
+            return_cuts=self.args.return_cuts,
+        )
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+        )
+        from lhotse.dataset.iterable_dataset import IterableDatasetWrapper
+
+        test_iter_dataset = IterableDatasetWrapper(
+            dataset=test,
+            sampler=sampler,
+        )
+        test_dl = DataLoader(
+            test_iter_dataset,
+            batch_size=None,
+            num_workers=self.args.num_workers,
+        )
+        return test_dl
+
+    @lru_cache()
+    def train_cuts(self) -> CutSet:
+        logging.info("About to get train cuts")
+        return load_manifest(self.args.manifest_dir / "cuts_train.json.gz")
+
+    @lru_cache()
+    def valid_cuts(self) -> CutSet:
+        logging.info("About to get dev cuts")
+        return load_manifest(self.args.manifest_dir / "cuts_eval.json.gz")
+
+    @lru_cache()
+    def test_cuts(self) -> List[CutSet]:
+        logging.info("About to get test cuts")
+        return load_manifest(self.args.manifest_dir / "cuts_test.json.gz")

egs/alimeeting/ASR/pruned_transducer_stateless2/beam_search.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/beam_search.py

egs/alimeeting/ASR/pruned_transducer_stateless2/conformer.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/conformer.py

egs/alimeeting/ASR/pruned_transducer_stateless2/decode.py

@@ -0,0 +1,615 @@
+#!/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.
+"""
+When training with the far data, usage:
+(1) greedy search
+./pruned_transducer_stateless2/decode.py \
+        --epoch 29 \
+        --avg 18 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir data/lang_char \
+        --max-duration 100 \
+        --decoding-method greedy_search
+
+(2) modified beam search
+./pruned_transducer_stateless2/decode.py \
+        --epoch 29 \
+        --avg 18 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir data/lang_char \
+        --max-duration 100 \
+        --decoding-method modified_beam_search \
+        --beam-size 4
+
+(3) fast beam search
+./pruned_transducer_stateless2/decode.py \
+        --epoch 29 \
+        --avg 18 \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --lang-dir data/lang_char \
+        --max-duration 1500 \
+        --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 asr_datamodule import AlimeetingAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from lhotse.cut import Cut
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.lexicon import Lexicon
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--batch",
+        type=int,
+        default=None,
+        help="It specifies the batch checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--avg-last-n",
+        type=int,
+        default=0,
+        help="""If positive, --epoch and --avg are ignored and it
+        will use the last n checkpoints exp_dir/checkpoint-xxx.pt
+        where xxx is the number of processed batches while
+        saving that checkpoint.
+        """,
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="The experiment dir",
+    )
+
+    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(
+        "--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 interger 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=2,
+        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""",
+    )
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    lexicon: Lexicon,
+    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.
+      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 = model.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
+    )
+    hyps = []
+
+    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,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    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,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    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,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    else:
+        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}"
+                )
+            hyps.append([lexicon.token_table[idx] for idx in hyp])
+
+    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,
+    lexicon: Lexicon,
+    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.
+      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 = 100
+    else:
+        log_interval = 50
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+        texts = [list(str(text).replace(" ", "")) for text in texts]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            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):
+                this_batch.append((ref_text, 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"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, 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\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER 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()
+    AlimeetingAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_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
+
+    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"-beam-{params.beam_size}"
+    else:
+        params.suffix += f"-context-{params.context_size}"
+        params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
+
+    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 = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    if params.avg_last_n > 0:
+        filenames = find_checkpoints(params.exp_dir)[: params.avg_last_n]
+        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)
+    elif params.batch is not None:
+        filenames = f"{params.exp_dir}/checkpoint-{params.batch}.pt"
+        logging.info(f"averaging {filenames}")
+        model.to(device)
+        model.load_state_dict(average_checkpoints([filenames], device=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))
+
+        average = average_checkpoints(filenames, device=device)
+        checkpoint = {"model": average}
+        torch.save(
+            checkpoint,
+            "pruned_transducer_stateless2/exp/pretrained_epoch_29_avg_18.pt",
+        )
+
+    model.to(device)
+    model.eval()
+    model.device = device
+
+    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}")
+
+    # Note: Please use "pip install webdataset==0.1.103"
+    # for installing the webdataset.
+    import glob
+    import os
+
+    from lhotse import CutSet
+    from lhotse.dataset.webdataset import export_to_webdataset
+
+    alimeeting = AlimeetingAsrDataModule(args)
+
+    dev = "eval"
+    test = "test"
+
+    if not os.path.exists(f"{dev}/shared-0.tar"):
+        os.makedirs(dev)
+        dev_cuts = alimeeting.valid_cuts()
+        export_to_webdataset(
+            dev_cuts,
+            output_path=f"{dev}/shared-%d.tar",
+            shard_size=300,
+        )
+
+    if not os.path.exists(f"{test}/shared-0.tar"):
+        os.makedirs(test)
+        test_cuts = alimeeting.test_cuts()
+        export_to_webdataset(
+            test_cuts,
+            output_path=f"{test}/shared-%d.tar",
+            shard_size=300,
+        )
+
+    dev_shards = [
+        str(path)
+        for path in sorted(glob.glob(os.path.join(dev, "shared-*.tar")))
+    ]
+    cuts_dev_webdataset = CutSet.from_webdataset(
+        dev_shards,
+        split_by_worker=True,
+        split_by_node=True,
+        shuffle_shards=True,
+    )
+
+    test_shards = [
+        str(path)
+        for path in sorted(glob.glob(os.path.join(test, "shared-*.tar")))
+    ]
+    cuts_test_webdataset = CutSet.from_webdataset(
+        test_shards,
+        split_by_worker=True,
+        split_by_node=True,
+        shuffle_shards=True,
+    )
+
+    def remove_short_and_long_utt(c: Cut):
+        return 1.0 <= c.duration
+
+    cuts_dev_webdataset = cuts_dev_webdataset.filter(remove_short_and_long_utt)
+    cuts_test_webdataset = cuts_test_webdataset.filter(
+        remove_short_and_long_utt
+    )
+
+    dev_dl = alimeeting.valid_dataloaders(cuts_dev_webdataset)
+    test_dl = alimeeting.test_dataloaders(cuts_test_webdataset)
+
+    test_sets = ["dev", "test"]
+    test_dl = [dev_dl, test_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            lexicon=lexicon,
+            decoding_graph=decoding_graph,
+        )
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()

egs/alimeeting/ASR/pruned_transducer_stateless2/decoder.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/decoder.py

egs/alimeeting/ASR/pruned_transducer_stateless2/encoder_interface.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/transducer_stateless/encoder_interface.py

egs/alimeeting/ASR/pruned_transducer_stateless2/export.py

@@ -0,0 +1,178 @@
+# 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_stateless2/export.py \
+  --exp-dir ./pruned_transducer_stateless2/exp \
+  --lang-dir data/lang_char \
+  --epoch 29 \
+  --avg 18
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `pruned_transducer_stateless2/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/alimeeting/ASR
+    ./pruned_transducer_stateless2/decode.py \
+        --exp-dir ./pruned_transducer_stateless2/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 get_params, get_transducer_model
+
+from icefall.checkpoint import average_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=28,
+        help="It specifies the checkpoint to use for decoding."
+        "Note: Epoch counts from 0.",
+    )
+
+    parser.add_argument(
+        "--avg",
+        type=int,
+        default=15,
+        help="Number of checkpoints to average. Automatically select "
+        "consecutive checkpoints before the checkpoint specified by "
+        "'--epoch'. ",
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="""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",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; "
+        "2 means tri-gram",
+    )
+
+    return parser
+
+
+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))
+
+    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)
+
+    model.to(device)
+
+    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.eval()
+
+    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()

egs/alimeeting/ASR/pruned_transducer_stateless2/joiner.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/joiner.py

egs/alimeeting/ASR/pruned_transducer_stateless2/model.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/model.py

egs/alimeeting/ASR/pruned_transducer_stateless2/optim.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/optim.py

egs/alimeeting/ASR/pruned_transducer_stateless2/pretrained.py

@@ -0,0 +1,347 @@
+#!/usr/bin/env python3
+# Copyright      2021  Xiaomi Corp.        (authors: Fangjun Kuang)
+# 		         2022  Xiaomi Crop.        (authors: Mingshuang 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.
+"""
+Here, the far data is used for training, usage:
+
+(1) greedy search
+./pruned_transducer_stateless2/pretrained.py \
+        --checkpoint ./pruned_transducer_stateless2/exp/pretrained.pt \
+        --lang-dir ./data/lang_char \
+        --decoding-method greedy_search \
+        --max-sym-per-frame 1 \
+        /path/to/foo.wav \
+        /path/to/bar.wav
+
+(2) modified beam search
+./pruned_transducer_stateless2/pretrained.py \
+        --checkpoint ./pruned_transducer_stateless2/exp/pretrained.pt \
+        --lang-dir ./data/lang_char \
+        --decoding-method modified_beam_search \
+        --beam-size 4 \
+        /path/to/foo.wav \
+        /path/to/bar.wav
+
+(3) fast beam search
+./pruned_transducer_stateless2/pretrained.py \
+        --checkpoint ./pruned_transducer_stateless/exp/pretrained.pt \
+        --lang-dir ./data/lang_char \
+        --decoding-method fast_beam_search \
+        --beam 4 \
+        --max-contexts 4 \
+        --max-states 8 \
+        /path/to/foo.wav \
+        /path/to/bar.wav
+
+You can also use `./pruned_transducer_stateless2/exp/epoch-xx.pt`.
+
+Note: ./pruned_transducer_stateless2/exp/pretrained.pt is generated by
+./pruned_transducer_stateless2/export.py
+"""
+
+
+import argparse
+import logging
+import math
+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 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=str,
+        help="""Path to lang.
+        """,
+    )
+
+    parser.add_argument(
+        "--decoding-method",
+        type=str,
+        default="greedy_search",
+        help="""Possible values are:
+          - greedy_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="Used only when --method is beam_search and 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=2,
+        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
+        --method is greedy_search.
+        """,
+    )
+
+    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))
+
+    lexicon = Lexicon(params.lang_dir)
+    params.blank_id = lexicon.token_table["<blk>"]
+    params.vocab_size = max(lexicon.tokens) + 1
+
+    logging.info(f"{params}")
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+
+    logging.info("Creating 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
+
+    if params.decoding_method == "fast_beam_search":
+        decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
+    else:
+        decoding_graph = None
+
+    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_lengths = [f.size(0) for f in features]
+
+    features = pad_sequence(
+        features, batch_first=True, padding_value=math.log(1e-10)
+    )
+
+    feature_lengths = torch.tensor(feature_lengths, device=device)
+
+    with torch.no_grad():
+        encoder_out, encoder_out_lens = model.encoder(
+            x=features, x_lens=feature_lengths
+        )
+
+    hyps = []
+    msg = f"Using {params.decoding_method}"
+    logging.info(msg)
+
+    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,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    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,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    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,
+        )
+        for i in range(encoder_out.size(0)):
+            hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]])
+    else:
+        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}"
+                )
+            hyps.append([lexicon.token_table[idx] for idx 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()

egs/alimeeting/ASR/pruned_transducer_stateless2/scaling.py

@@ -0,0 +1 @@
+../../../librispeech/ASR/pruned_transducer_stateless2/scaling.py

egs/alimeeting/ASR/pruned_transducer_stateless2/train.py

@@ -0,0 +1,972 @@
+#!/usr/bin/env python3
+# Copyright    2021  Xiaomi Corp.        (authors: Fangjun Kuang,
+#                                                  Wei Kang
+#                                                  Mingshuang 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:
+
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+
+./pruned_transducer_stateless2/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --lang-dir data/lang_char \
+  --max-duration 220 \
+  --save-every-n 1000
+
+# For mix precision training:
+
+./pruned_transducer_stateless2/train.py \
+  --world-size 4 \
+  --num-epochs 30 \
+  --start-epoch 0 \
+  --exp-dir pruned_transducer_stateless2/exp \
+  --lang-dir data/lang_char \
+  --max-duration 220 \
+  --save-every-n 1000
+  --use-fp16 True
+
+"""
+
+import argparse
+import logging
+import os
+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 asr_datamodule import AlimeetingAsrDataModule
+from conformer import Conformer
+from decoder import Decoder
+from joiner import Joiner
+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
+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
+]
+
+os.environ["CUDA_LAUNCH_BLOCKING"] = "1"
+
+
+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=12359,
+        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=0,
+        help="""Resume training from from this epoch.
+        If it is positive, it will load checkpoint from
+        transducer_stateless2/exp/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_stateless2/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=2,
+        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=8000,
+        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=20,
+        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(
+        "--use-fp16",
+        type=str2bool,
+        default=False,
+        help="Whether to use half precision training.",
+    )
+
+    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": 10,
+            "log_interval": 1,
+            "reset_interval": 200,
+            "valid_interval": 400,
+            # parameters for conformer
+            "feature_dim": 80,
+            "subsampling_factor": 4,
+            "encoder_dim": 512,
+            "nhead": 8,
+            "dim_feedforward": 2048,
+            "num_encoder_layers": 12,
+            # parameters for decoder
+            "decoder_dim": 512,
+            # parameters for joiner
+            "joiner_dim": 512,
+            # parameters for Noam
+            "model_warm_step": 200,
+            "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)
+
+    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,
+    )
+    return model
+
+
+def load_checkpoint_if_available(
+    params: AttributeDict,
+    model: nn.Module,
+    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 positive, 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.
+      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 > 0:
+        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,
+        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]
+
+    if params.start_batch > 0:
+        if "cur_epoch" in saved_params:
+            params["start_epoch"] = saved_params["cur_epoch"]
+
+    return saved_params
+
+
+def save_checkpoint(
+    params: AttributeDict,
+    model: nn.Module,
+    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.
+      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,
+        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 compute_loss(
+    params: AttributeDict,
+    model: nn.Module,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    batch: dict,
+    is_training: bool,
+    warmup: float = 1.0,
+) -> Tuple[Tensor, MetricsTracker]:
+    """
+    Compute CTC 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
+    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)
+
+    texts = batch["supervisions"]["text"]
+
+    y = graph_compiler.texts_to_ids(texts)
+    if type(y) == list:
+        y = k2.RaggedTensor(y).to(device)
+    else:
+        y = y.to(device)
+
+    with torch.set_grad_enabled(is_training):
+        simple_loss, pruned_loss = model(
+            x=feature,
+            x_lens=feature_lens,
+            y=y,
+            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: nn.Module,
+    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: nn.Module,
+    optimizer: torch.optim.Optimizer,
+    scheduler: LRSchedulerType,
+    graph_compiler: CharCtcTrainingGraphCompiler,
+    train_dl: torch.utils.data.DataLoader,
+    valid_dl: torch.utils.data.DataLoader,
+    scaler: GradScaler,
+    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.
+      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()
+
+    tot_loss = MetricsTracker()
+
+    for batch_idx, batch in enumerate(train_dl):
+
+        params.batch_idx_train += 1
+        batch_size = len(batch["supervisions"]["text"])
+
+        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
+
+        # 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()
+
+        if params.print_diagnostics and batch_idx == 5:
+            return
+
+        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,
+                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}, loss[{loss_info}], "
+                f"tot_loss[{tot_loss}], 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, "train/current_", params.batch_idx_train
+                )
+                tot_loss.write_summary(
+                    tb_writer, "train/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 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)
+    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,
+    )
+
+    params.blank_id = lexicon.token_table["<blk>"]
+    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}")
+
+    checkpoints = load_checkpoint_if_available(params=params, model=model)
+
+    model.to(device)
+    if world_size > 1:
+        logging.info("Using DDP")
+        model = DDP(model, device_ids=[rank])
+    model.device = device
+
+    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)
+
+    alimeeting = AlimeetingAsrDataModule(args)
+
+    train_cuts = alimeeting.train_cuts()
+    valid_cuts = alimeeting.valid_cuts()
+
+    def remove_short_and_long_utt(c: Cut):
+        # Keep only utterances with duration between 1 second and 15.0 seconds
+        #
+        # Caution: There is a reason to select 10.0 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 <= 15.0
+
+    train_cuts = train_cuts.filter(remove_short_and_long_utt)
+
+    valid_dl = alimeeting.valid_dataloaders(valid_cuts)
+
+    if params.start_batch > 0 and checkpoints and "sampler" in checkpoints:
+        # We only load the sampler's state dict when it loads a checkpoint
+        # saved in the middle of an epoch
+        sampler_state_dict = checkpoints["sampler"]
+    else:
+        sampler_state_dict = None
+
+    train_dl = alimeeting.train_dataloaders(
+        train_cuts, sampler_state_dict=sampler_state_dict
+    )
+
+    if not params.print_diagnostics and params.start_batch == 0:
+        scan_pessimistic_batches_for_oom(
+            model=model,
+            train_dl=train_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"])
+
+    for epoch in range(params.start_epoch, params.num_epochs):
+        scheduler.step_epoch(epoch)
+        fix_random_seed(params.seed + epoch)
+        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,
+            optimizer=optimizer,
+            scheduler=scheduler,
+            graph_compiler=graph_compiler,
+            train_dl=train_dl,
+            valid_dl=valid_dl,
+            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,
+            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 scan_pessimistic_batches_for_oom(
+    model: nn.Module,
+    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 0 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,
+                )
+            loss.backward()
+            optimizer.step()
+            optimizer.zero_grad()
+        except RuntimeError 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]}) ..."
+                )
+            raise
+
+
+def main():
+    parser = get_parser()
+    AlimeetingAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.lang_dir = Path(args.lang_dir)
+    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()

egs/alimeeting/ASR/shared

@@ -0,0 +1 @@
+../../../egs/aishell/ASR/shared

egs/librispeech/ASR/README.md

@@ -21,6 +21,8 @@ The following table lists the differences among them.
 | `pruned_transducer_stateless3`        | Conformer(modified) | Embedding + Conv1d | Using k2 pruned RNN-T loss + using GigaSpeech as extra training data |
 | `pruned_transducer_stateless4`        | Conformer(modified) | Embedding + Conv1d | same as pruned_transducer_stateless2 + save averaged models periodically during training                        |
 | `pruned_transducer_stateless5`        | Conformer(modified) | Embedding + Conv1d | same as pruned_transducer_stateless4 + more layers + random combiner|
+| `pruned_transducer_stateless6`        | Conformer(modified) | Embedding + Conv1d | same as pruned_transducer_stateless4 + distillation with hubert|
+| `pruned_stateless_emformer_rnnt2`     | Emformer(from torchaudio) | Embedding + Conv1d | Using Emformer from torchaudio for streaming ASR|
 
 
 The decoder in `transducer_stateless` is modified from the paper

egs/librispeech/ASR/RESULTS-100hours.md

@@ -3,6 +3,31 @@
 This page shows the WERs for test-clean/test-other using only
 train-clean-100 subset as training data.
 
+## Distillation with hubert
+### 2022-05-27
+Related models/log/tensorboard:
+https://huggingface.co/GuoLiyong/stateless6_baseline_vs_disstillation
+
+Following results are obtained by ./distillation_with_hubert.sh
+
+The only differences is in pruned_transducer_stateless6/train.py.
+
+For baseline: set enable_distillation=False
+
+For distillation: set enable_distillation=True (the default)
+
+Decoding method is modified beam search.
+|                                     | test-clean | test-other | comment                                  |
+|-------------------------------------|------------|------------|------------------------------------------|
+| baseline no vq distillation         | 7.09       | 18.88      | --epoch 20, --avg 10, --max-duration 200 |
+| baseline no vq distillation         | 6.83       | 18.19      | --epoch 30, --avg 10, --max-duration 200 |
+| baseline no vq distillation         | 6.73       | 17.79      | --epoch 40, --avg 10, --max-duration 200 |
+| baseline no vq distillation         | 6.75       | 17.68      | --epoch 50, --avg 10, --max-duration 200 |
+| distillation with hubert            | 5.82       | 15.98      | --epoch 20, --avg 10, --max-duration 200 |
+| distillation with hubert            | 5.52       | 15.15      | --epoch 30, --avg 10, --max-duration 200 |
+| distillation with hubert            | 5.45       | 14.94      | --epoch 40, --avg 10, --max-duration 200 |
+| distillation with hubert            | 5.50       | 14.77      | --epoch 50, --avg 10, --max-duration 200 |
+
 ## Conformer encoder + embedding decoder
 
 ### 2022-02-21

egs/librispeech/ASR/RESULTS.md

@@ -1,5 +1,71 @@
 ## Results
 
+### LibriSpeech BPE training results (Pruned Stateless Emformer RNN-T)
+
+[pruned_stateless_emformer_rnnt2](./pruned_stateless_emformer_rnnt2)
+
+Use <https://github.com/k2-fsa/icefall/pull/390>.
+
+Use [Emformer](https://arxiv.org/abs/2010.10759) from [torchaudio](https://github.com/pytorch/audio)
+for streaming ASR. The Emformer model is imported from torchaudio without modifications.
+
+|                                     | test-clean | test-other | comment                                |
+|-------------------------------------|------------|------------|----------------------------------------|
+| greedy search (max sym per frame 1) | 4.28       | 11.42       | --epoch 39 --avg 6  --max-duration 600 |
+| modified beam search                | 4.22       | 11.16       | --epoch 39 --avg 6  --max-duration 600 |
+| fast beam search                    | 4.29       | 11.26       | --epoch 39 --avg 6 --max-duration 600  |
+
+
+The training commands are:
+```bash
+export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7"
+
+./pruned_stateless_emformer_rnnt2/train.py \
+  --world-size 8 \
+  --num-epochs 40 \
+  --start-epoch 1 \
+  --exp-dir pruned_stateless_emformer_rnnt2/exp-full \
+  --full-libri 1 \
+  --use-fp16 0 \
+  --max-duration 200 \
+  --prune-range 5 \
+  --lm-scale 0.25 \
+  --master-port 12358 \
+  --num-encoder-layers 18 \
+  --left-context-length 128 \
+  --segment-length 8 \
+  --right-context-length 4
+```
+
+The tensorboard log can be found at
+<https://tensorboard.dev/experiment/ZyiqhAhmRjmr49xml4ofLw/>
+
+The decoding commands are:
+```bash
+for m in greedy_search fast_beam_search modified_beam_search; do
+  for epoch in 39; do
+    for avg in 6; do
+      ./pruned_stateless_emformer_rnnt2/decode.py \
+        --epoch $epoch \
+        --avg $avg \
+        --use-averaged-model 1 \
+        --exp-dir pruned_stateless_emformer_rnnt2/exp-full \
+        --max-duration 50 \
+        --decoding-method $m \
+        --num-encoder-layers 18 \
+        --left-context-length 128 \
+        --segment-length 8 \
+        --right-context-length 4
+    done
+  done
+done
+```
+
+You can find a pretrained model, training logs, decoding logs, and decoding
+results at:
+<https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-stateless-emformer-rnnt2-2022-06-01>
+
+
 ### LibriSpeech BPE training results (Pruned Stateless Transducer 5)
 
 [pruned_transducer_stateless5](./pruned_transducer_stateless5)

egs/librispeech/ASR/distillation_with_hubert.sh

@@ -0,0 +1,144 @@
+# A short introduction about distillation framework.
+#
+# A typical traditional distillation method is
+# Loss(teacher embedding, student embedding).
+#
+# Comparing to these, the proposed distillation framework contains two mainly steps:
+# codebook indexes = quantizer.encode(teacher embedding)
+# Loss(codebook indexes, student embedding)
+#
+# Things worth to meantion:
+# 1. The float type teacher embedding is quantized into a sequence of
+#    8-bit integer codebook indexes.
+# 2. a middle layer 36(1-based) out of total 48 layers is used to extract
+#    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
+
+# Set the GPUs available.
+# This script requires at least one GPU.
+# You MUST set environment variable "CUDA_VISIBLE_DEVICES",
+# even you only have ONE GPU. It needed by CodebookIndexExtractor to determine numbert of jobs to extract codebook indexes parallelly.
+
+# Suppose only one GPU exists:
+# export CUDA_VISIBLE_DEVICES="0"
+#
+# Suppose GPU 2,3,4,5 are available.
+export CUDA_VISIBLE_DEVICES="2,3,4,5"
+
+
+if [ $stage -eq 0 ]; then
+  # Preparation stage.
+
+  # Install fairseq according to:
+  # https://github.com/pytorch/fairseq
+  # when testing this code:
+  # commit 806855bf660ea748ed7ffb42fe8dcc881ca3aca0 is used.
+  has_fairseq=$(python3 -c "import importlib; print(importlib.util.find_spec('fairseq') is not None)")
+  if [ $has_fairseq == 'False' ]; then
+    echo "Please install fairseq before running following stages"
+    exit 1
+  fi
+
+  # Install quantization toolkit:
+  # pip install git+https://github.com/danpovey/quantization.git@master
+  # when testing this code:
+  # commit c17ffe67aa2e6ca6b6855c50fde812f2eed7870b is used.
+
+  has_quantization=$(python3 -c "import importlib; print(importlib.util.find_spec('quantization') is not None)")
+  if [ $has_quantization == 'False' ]; then
+    echo "Please install quantization before running following stages"
+    exit 1
+  fi
+
+  echo "Download hubert model."
+  # Parameters about model.
+  exp_dir=./pruned_transducer_stateless6/exp/
+  model_id=hubert_xtralarge_ll60k_finetune_ls960
+  hubert_model_dir=${exp_dir}/hubert_models
+  hubert_model=${hubert_model_dir}/${model_id}.pt
+  mkdir -p ${hubert_model_dir}
+  # For more models refer to: https://github.com/pytorch/fairseq/tree/main/examples/hubert
+  if [ -f ${hubert_model} ]; then
+    echo "hubert model alread exists."
+  else
+    wget -c https://dl.fbaipublicfiles.com/hubert/${model_id} -P ${hubert_model}
+    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"
+  exit 1
+fi
+
+if [ $stage -eq 1 ]; then
+  # 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
+fi
+
+if [ $stage -eq 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.
+  # Training dataset including clean-100h with speed perturb 0.9 and 1.1 has 300 hours.
+  # The output frame rates of Hubert is 50 per second.
+  # Theoretically, 412M = 300 * 3600 * 50 * 8 / 1024 / 1024 is needed.
+  # The actual size of all "*.h5" files storaging codebook index is 450M.
+  # I think the extra "48M" usage is some meta information.
+
+  # Time consumption analysis:
+  # For quantizer training data(teacher embedding) extraction, only 1000 utts from clean-100 are used.
+  # Together with quantizer training, no more than 20 minutes will be used.
+  #
+  # For codebook indexes extraction,
+  # with two pieces of NVIDIA A100 gpus, around three hours needed to process 300 hours training data,
+  # i.e. clean-100 with speed purteb 0.9 and 1.1.
+
+  # GPU usage:
+  # 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.
+  ./pruned_transducer_stateless6/extract_codebook_index.py \
+    --full-libri False
+fi
+
+if [ $stage -eq 3 ]; then
+  # Example training script.
+  # Note: it's better to set spec-aug-time-warpi-factor=-1
+  WORLD_SIZE=$(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 \
+    --spec-aug-time-warp-factor -1 \
+    --max-duration 300 \
+    --world-size ${WORLD_SIZE} \
+    --num-epochs 20
+fi
+
+if [ $stage -eq 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
+  ./pruned_transducer_stateless6/decode.py \
+    --decoding-method "modified_beam_search" \
+    --epoch 20 \
+    --avg 10 \
+    --max-duration 200 \
+    --exp-dir ./pruned_transducer_stateless6/exp
+fi

egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/asr_datamodule.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless/asr_datamodule.py

egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/beam_search.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless/beam_search.py

egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/decode.py

@@ -0,0 +1,645 @@
+#!/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_stateless_emformer_rnnt/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_stateless_emformer_rnnt/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_stateless_emformer_rnnt/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_stateless_emformer_rnnt/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_stateless_emformer_rnnt/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_stateless_emformer_rnnt/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search
+./pruned_stateless_emformer_rnnt/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_stateless_emformer_rnnt/exp \
+    --max-duration 600 \
+    --decoding-method fast_beam_search \
+    --beam 4 \
+    --max-contexts 4 \
+    --max-states 8
+"""
+
+
+import argparse
+import logging
+import math
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Optional, Tuple
+
+import k2
+import sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+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.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    str2bool,
+    write_error_stats,
+)
+
+LOG_EPS = math.log(1e-10)
+
+
+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_stateless_emformer_rnnt/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    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=2,
+        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,
+    sp: spm.SentencePieceProcessor,
+    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.
+      sp:
+        The BPE model.
+      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)
+
+    feature_lens += params.left_context_length
+    feature = torch.nn.functional.pad(
+        feature,
+        pad=(0, 0, 0, params.left_context_length),
+        value=LOG_EPS,
+    )
+
+    encoder_out, encoder_out_lens = model.encoder(
+        x=feature, x_lens=feature_lens
+    )
+    hyps = []
+
+    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,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    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,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    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,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        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}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    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,
+    sp: spm.SentencePieceProcessor,
+    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.
+      sp:
+        The BPE model.
+      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 = 10
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            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"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, 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\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER 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()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_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}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    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(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}")
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()

egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/decoder.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless/decoder.py

egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/emformer.py

@@ -0,0 +1,313 @@
+# Copyright      2022  Xiaomi Corporation     (Author: Mingshuang 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.
+
+import math
+from typing import List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from subsampling import Conv2dSubsampling, VggSubsampling
+
+try:
+    from torchaudio.models import Emformer as _Emformer
+except ImportError:
+    import torchaudio
+
+    print(
+        "Please install torchaudio >= 0.11.0. "
+        f"Current version: {torchaudio.__version__}"
+    )
+    raise
+
+
+def unstack_states(
+    states: List[List[torch.Tensor]],
+) -> List[List[List[torch.Tensor]]]:
+    """Unstack the emformer state corresponding to a batch of utterances
+    into a list of states, were the i-th entry is the state from the i-th
+    utterance in the batch.
+
+    Args:
+      states:
+        A list-of-list of tensors. ``len(states)`` equals to number of
+        layers in the emformer. ``states[i]]`` contains the states for
+        the i-th layer. ``states[i][k]`` is either a 3-D tensor of shape
+        ``(T, N, C)`` or a 2-D tensor of shape ``(C, N)``
+    """
+    batch_size = states[0][0].size(1)
+    num_layers = len(states)
+
+    ans = [None] * batch_size
+    for i in range(batch_size):
+        ans[i] = [[] for _ in range(num_layers)]
+
+    for li, layer in enumerate(states):
+        for s in layer:
+            s_list = s.unbind(dim=1)
+            # We will use stack(dim=1) later in stack_states()
+            for bi, b in enumerate(ans):
+                b[li].append(s_list[bi])
+    return ans
+
+
+def stack_states(
+    state_list: List[List[List[torch.Tensor]]],
+) -> List[List[torch.Tensor]]:
+    """Stack list of emformer states that correspond to separate utterances
+    into a single emformer state so that it can be used as an input for
+    emformer when those utterances are formed into a batch.
+
+    Note:
+      It is the inverse of :func:`unstack_states`.
+
+    Args:
+      state_list:
+        Each element in state_list corresponding to the internal state
+        of the emformer model for a single utterance.
+    Returns:
+      Return a new state corresponding to a batch of utterances.
+      See the input argument of :func:`unstack_states` for the meaning
+      of the returned tensor.
+    """
+    batch_size = len(state_list)
+    ans = []
+    for layer in state_list[0]:
+        # layer is a list of tensors
+        if batch_size > 1:
+            ans.append([[s] for s in layer])
+            # Note: We will stack ans[layer][s][] later to get ans[layer][s]
+        else:
+            ans.append([s.unsqueeze(1) for s in layer])
+
+    for b, states in enumerate(state_list[1:], 1):
+        for li, layer in enumerate(states):
+            for si, s in enumerate(layer):
+                ans[li][si].append(s)
+                if b == batch_size - 1:
+                    ans[li][si] = torch.stack(ans[li][si], dim=1)
+                    # We will use unbind(dim=1) later in unstack_states()
+    return ans
+
+
+class Emformer(EncoderInterface):
+    """This is just a simple wrapper around torchaudio.models.Emformer.
+    We may replace it with our own implementation some time later.
+    """
+
+    def __init__(
+        self,
+        num_features: int,
+        output_dim: int,
+        d_model: int,
+        nhead: int,
+        dim_feedforward: int,
+        num_encoder_layers: int,
+        segment_length: int,
+        left_context_length: int,
+        right_context_length: int,
+        max_memory_size: int = 0,
+        dropout: float = 0.1,
+        subsampling_factor: int = 4,
+        vgg_frontend: bool = False,
+    ) -> None:
+        """
+        Args:
+          num_features:
+            The input dimension of the model.
+          output_dim:
+            The output dimension of the model.
+          d_model:
+            Attention dimension.
+          nhead:
+            Number of heads in multi-head attention.
+          dim_feedforward:
+            The output dimension of the feedforward layers in encoder.
+          num_encoder_layers:
+            Number of encoder layers.
+          segment_length:
+            Number of frames per segment before subsampling.
+          left_context_length:
+            Number of frames in the left context before subsampling.
+          right_context_length:
+            Number of frames in the right context before subsampling.
+          max_memory_size:
+            TODO.
+          dropout:
+            Dropout in encoder.
+          subsampling_factor:
+            Number of output frames is num_in_frames // subsampling_factor.
+            Currently, subsampling_factor MUST be 4.
+          vgg_frontend:
+            True to use vgg style frontend for subsampling.
+        """
+        super().__init__()
+
+        self.subsampling_factor = subsampling_factor
+        if subsampling_factor != 4:
+            raise NotImplementedError("Support only 'subsampling_factor=4'.")
+
+        # self.encoder_embed converts the input of shape (N, T, num_features)
+        # to the shape (N, T//subsampling_factor, d_model).
+        # That is, it does two things simultaneously:
+        #   (1) subsampling: T -> T//subsampling_factor
+        #   (2) embedding: num_features -> d_model
+        if vgg_frontend:
+            self.encoder_embed = VggSubsampling(num_features, d_model)
+        else:
+            self.encoder_embed = Conv2dSubsampling(num_features, d_model)
+
+        self.segment_length = segment_length  # before subsampling
+        self.right_context_length = right_context_length
+
+        assert right_context_length % subsampling_factor == 0
+        assert segment_length % subsampling_factor == 0
+        assert left_context_length % subsampling_factor == 0
+
+        left_context_length = left_context_length // subsampling_factor
+        right_context_length = right_context_length // subsampling_factor
+        segment_length = segment_length // subsampling_factor
+
+        self.model = _Emformer(
+            input_dim=d_model,
+            num_heads=nhead,
+            ffn_dim=dim_feedforward,
+            num_layers=num_encoder_layers,
+            segment_length=segment_length,
+            dropout=dropout,
+            activation="relu",
+            left_context_length=left_context_length,
+            right_context_length=right_context_length,
+            max_memory_size=max_memory_size,
+            weight_init_scale_strategy="depthwise",
+            tanh_on_mem=False,
+            negative_inf=-1e8,
+        )
+
+        self.encoder_output_layer = nn.Sequential(
+            nn.Dropout(p=dropout), nn.Linear(d_model, output_dim)
+        )
+        self.log_eps = math.log(1e-10)
+
+        self._init_state = torch.jit.Attribute([], List[List[torch.Tensor]])
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+          x:
+            Input features of shape (N, T, C).
+          x_lens:
+            A int32 tensor of shape (N,) containing valid frames in `x` before
+            padding. We have `x.size(1) == x_lens.max()`
+        Returns:
+          Return a tuple containing two tensors:
+
+            - encoder_out, a tensor of shape (N, T', C)
+            - encoder_out_lens, a int32 tensor of shape (N,) containing the
+              valid frames in `encoder_out` before padding
+        """
+        x = nn.functional.pad(
+            x,
+            # (left, right, top, bottom)
+            # left/right are for the channel dimension, i.e., axis 2
+            # top/bottom are for the time dimension, i.e., axis 1
+            (0, 0, 0, self.right_context_length),
+            value=self.log_eps,
+        )  # (N, T, C) -> (N, T+right_context_length, C)
+
+        x = self.encoder_embed(x)
+
+        # Caution: We assume the subsampling factor is 4!
+        x_lens = (((x_lens - 1) >> 1) - 1) >> 1
+
+        emformer_out, emformer_out_lens = self.model(x, x_lens)
+        logits = self.encoder_output_layer(emformer_out)
+
+        return logits, emformer_out_lens
+
+    @torch.jit.export
+    def streaming_forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        states: Optional[List[List[torch.Tensor]]] = None,
+    ):
+        """
+        Args:
+          x:
+            A 3-D tensor of shape (N, T, C). Note: x also contains right
+            context frames.
+          x_lens:
+            A 2-D tensor of shap containing the number of valid frames for each
+            element in `x` before padding. Note: It also counts right context
+            frames.
+          states:
+            Internal states of the model.
+        Returns:
+          Return a tuple containing 3 tensors:
+            - encoder_out, a 3-D tensor of shape (N, T, C)
+            - encoder_out_lens: a 1-D tensor of shape (N,)
+            - next_state, internal model states for the next invocation
+        """
+        x = self.encoder_embed(x)
+
+        # Caution: We assume the subsampling factor is 4!
+        x_lens = (((x_lens - 1) >> 1) - 1) >> 1
+
+        emformer_out, emformer_out_lens, states = self.model.infer(
+            x, x_lens, states
+        )
+
+        if x.size(1) != (
+            self.model.segment_length + self.model.right_context_length
+        ):
+            raise ValueError(
+                "Incorrect input shape."
+                f"{x.size(1)} vs {self.model.segment_length} + "
+                f"{self.model.right_context_length}"
+            )
+
+        logits = self.encoder_output_layer(emformer_out)
+
+        return logits, emformer_out_lens, states
+
+    @torch.jit.export
+    def get_init_state(self, device: torch.device) -> List[List[torch.Tensor]]:
+        """Return the initial state of each layer.
+
+        Returns:
+          Return the initial state of each layer. NOTE: the returned
+          tensors are on the given device. `len(ans) == num_emformer_layers`.
+        """
+        if self._init_state:
+            # Note(fangjun): It is OK to share the init state as it is
+            # not going to be modified by the model
+            return self._init_state
+
+        batch_size = 1
+
+        ans: List[List[torch.Tensor]] = []
+        for layer in self.model.emformer_layers:
+            s = layer._init_state(batch_size=batch_size, device=device)
+            ans.append(s)
+
+        self._init_state = ans
+
+        return ans

egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/encoder_interface.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless/encoder_interface.py

egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/export.py

@@ -0,0 +1,281 @@
+#!/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:
+./prunted_stateless_emformer_rnnt/export.py \
+  --exp-dir ./prunted_stateless_emformer_rnnt/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `prunted_stateless_emformer_rnnt/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./prunted_stateless_emformer_rnnt/decode.py \
+        --exp-dir ./prunted_stateless_emformer_rnnt/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 600 \
+        --decoding-method greedy_search \
+        --bpe-model data/lang_bpe_500/bpe.model
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+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.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        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=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="prunted_stateless_emformer_rnnt/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        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()
+    args.exp_dir = Path(args.exp_dir)
+
+    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}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    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.eval()
+
+    model.to("cpu")
+    model.eval()
+    for p in model.parameters():
+        p.requires_grad_(False)
+
+    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"
+        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()

egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/joiner.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless/joiner.py

egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/model.py

@@ -0,0 +1,169 @@
+# 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.
+
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+
+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,
+    ):
+        """
+        Args:
+          encoder:
+            It is the transcription network in the paper. Its accepts
+            two inputs: `x` of (N, T, C) and `x_lens` of shape (N,).
+            It returns two tensors: `logits` of shape (N, T, C) 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, C). It should contain
+            one attribute: `blank_id`.
+          joiner:
+            It has two inputs with shapes: (N, T, C) and (N, U, C). Its
+            output shape is (N, T, U, C). Note that its output contains
+            unnormalized probs, i.e., not processed by log-softmax.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+    ) -> 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.
+          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
+        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, x_lens = self.encoder(x, x_lens)
+        assert torch.all(x_lens > 0)
+
+        # 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 = self.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, C]
+        decoder_out = self.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] = x_lens
+
+        simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
+            lm=decoder_out,
+            am=encoder_out,
+            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, C]
+        # lm_pruned : [B, T, prune_range, C]
+        am_pruned, lm_pruned = k2.do_rnnt_pruning(
+            am=encoder_out, lm=decoder_out, ranges=ranges
+        )
+
+        # logits : [B, T, prune_range, C]
+        logits = self.joiner(am_pruned, lm_pruned)
+
+        pruned_loss = k2.rnnt_loss_pruned(
+            logits=logits,
+            symbols=y_padded,
+            ranges=ranges,
+            termination_symbol=blank_id,
+            boundary=boundary,
+            reduction="sum",
+        )
+
+        return (simple_loss, pruned_loss)

egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/noam.py

@@ -0,0 +1,104 @@
+# Copyright    2021 University of Chinese Academy of Sciences (author: Han Zhu)
+#
+# 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
+
+
+class Noam(object):
+    """
+    Implements Noam optimizer.
+
+    Proposed in
+    "Attention Is All You Need", https://arxiv.org/pdf/1706.03762.pdf
+
+    Modified from
+    https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/optimizer.py  # noqa
+
+    Args:
+      params:
+        iterable of parameters to optimize or dicts defining parameter groups
+      model_size:
+        attention dimension of the transformer model
+      factor:
+        learning rate factor
+      warm_step:
+        warmup steps
+    """
+
+    def __init__(
+        self,
+        params,
+        model_size: int = 256,
+        factor: float = 10.0,
+        warm_step: int = 25000,
+        weight_decay=0,
+    ) -> None:
+        """Construct an Noam object."""
+        self.optimizer = torch.optim.Adam(
+            params, lr=0, betas=(0.9, 0.98), eps=1e-9, weight_decay=weight_decay
+        )
+        self._step = 0
+        self.warmup = warm_step
+        self.factor = factor
+        self.model_size = model_size
+        self._rate = 0
+
+    @property
+    def param_groups(self):
+        """Return param_groups."""
+        return self.optimizer.param_groups
+
+    def step(self):
+        """Update parameters and rate."""
+        self._step += 1
+        rate = self.rate()
+        for p in self.optimizer.param_groups:
+            p["lr"] = rate
+        self._rate = rate
+        self.optimizer.step()
+
+    def rate(self, step=None):
+        """Implement `lrate` above."""
+        if step is None:
+            step = self._step
+        return (
+            self.factor
+            * self.model_size ** (-0.5)
+            * min(step ** (-0.5), step * self.warmup ** (-1.5))
+        )
+
+    def zero_grad(self):
+        """Reset gradient."""
+        self.optimizer.zero_grad()
+
+    def state_dict(self):
+        """Return state_dict."""
+        return {
+            "_step": self._step,
+            "warmup": self.warmup,
+            "factor": self.factor,
+            "model_size": self.model_size,
+            "_rate": self._rate,
+            "optimizer": self.optimizer.state_dict(),
+        }
+
+    def load_state_dict(self, state_dict):
+        """Load state_dict."""
+        for key, value in state_dict.items():
+            if key == "optimizer":
+                self.optimizer.load_state_dict(state_dict["optimizer"])
+            else:
+                setattr(self, key, value)

egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/subsampling.py

@@ -0,0 +1 @@
+../conformer_ctc/subsampling.py

egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/test_emformer.py

@@ -0,0 +1,142 @@
+#!/usr/bin/env python3
+# Copyright    2022  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.
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./pruned_stateless_emformer_rnnt/test_emformer.py
+"""
+
+import torch
+from emformer import Emformer, stack_states, unstack_states
+
+
+def test_emformer():
+    N = 3
+    T = 300
+    C = 80
+
+    output_dim = 500
+
+    encoder = Emformer(
+        num_features=C,
+        output_dim=output_dim,
+        d_model=512,
+        nhead=8,
+        dim_feedforward=2048,
+        num_encoder_layers=20,
+        segment_length=16,
+        left_context_length=120,
+        right_context_length=4,
+        vgg_frontend=False,
+    )
+
+    x = torch.rand(N, T, C)
+    x_lens = torch.randint(100, T, (N,))
+    x_lens[0] = T
+
+    y, y_lens = encoder(x, x_lens)
+
+    y_lens = (((x_lens - 1) >> 1) - 1) >> 1
+    assert x.size(0) == x.size(0)
+    assert y.size(1) == max(y_lens)
+    assert y.size(2) == output_dim
+
+    num_param = sum([p.numel() for p in encoder.parameters()])
+    print(f"Number of encoder parameters: {num_param}")
+
+
+def test_emformer_streaming_forward():
+    N = 3
+    C = 80
+
+    output_dim = 500
+
+    encoder = Emformer(
+        num_features=C,
+        output_dim=output_dim,
+        d_model=512,
+        nhead=8,
+        dim_feedforward=2048,
+        num_encoder_layers=20,
+        segment_length=16,
+        left_context_length=120,
+        right_context_length=4,
+        vgg_frontend=False,
+    )
+
+    x = torch.rand(N, 23, C)
+    x_lens = torch.full((N,), 23)
+    y, y_lens, states = encoder.streaming_forward(x=x, x_lens=x_lens)
+
+    state_list = unstack_states(states)
+    states2 = stack_states(state_list)
+
+    for ss, ss2 in zip(states, states2):
+        for s, s2 in zip(ss, ss2):
+            assert torch.allclose(s, s2), f"{s.sum()}, {s2.sum()}"
+
+
+def test_emformer_init_state():
+    num_encoder_layers = 20
+    d_model = 512
+    encoder = Emformer(
+        num_features=80,
+        output_dim=500,
+        d_model=512,
+        nhead=8,
+        dim_feedforward=2048,
+        num_encoder_layers=num_encoder_layers,
+        segment_length=16,
+        left_context_length=120,
+        right_context_length=4,
+        vgg_frontend=False,
+    )
+    init_state = encoder.get_init_state()
+    assert len(init_state) == num_encoder_layers
+    layer0_state = init_state[0]
+    assert len(layer0_state) == 4
+
+    assert layer0_state[0].shape == (
+        0,  # max_memory_size
+        1,  # batch_size
+        d_model,  # input_dim
+    )
+
+    assert layer0_state[1].shape == (
+        encoder.model.left_context_length,
+        1,  # batch_size
+        d_model,  # input_dim
+    )
+    assert layer0_state[2].shape == layer0_state[1].shape
+    assert layer0_state[3].shape == (
+        1,  # always 1
+        1,  # batch_size
+    )
+
+
+@torch.no_grad()
+def main():
+    test_emformer()
+    test_emformer_streaming_forward()
+    test_emformer_init_state()
+
+
+if __name__ == "__main__":
+    torch.manual_seed(20220329)
+    main()

egs/librispeech/ASR/pruned_stateless_emformer_rnnt2/test_model.py

@@ -0,0 +1,59 @@
+#!/usr/bin/env python3
+# Copyright    2022  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.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./pruned_stateless_emformer_rnnt/test_model.py
+"""
+
+import torch
+from train import get_params, get_transducer_model
+
+
+def test_model():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.unk_id = 2
+
+    params.attention_dim = 512
+    params.nhead = 8
+    params.dim_feedforward = 2048
+    params.num_encoder_layers = 18
+    params.left_context_length = 128
+    params.segment_length = 8
+    params.right_context_length = 4
+    params.memory_size = 0
+
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+    model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+    torch.jit.script(model)
+
+
+def main():
+    test_model()
+
+
+if __name__ == "__main__":
+    main()

egs/librispeech/ASR/pruned_transducer_stateless/joiner.py

@@ -32,13 +32,15 @@ class Joiner(nn.Module):
         """
         Args:
           encoder_out:
-            Output from the encoder. Its shape is (N, T, s_range, C).
+            Output from the encoder. Its shape is (N, T, s_range, C) during
+            training or (N, C) in case of streaming decoding.
           decoder_out:
-            Output from the decoder. Its shape is (N, T, s_range, C).
-        Returns:
+            Output from the decoder. Its shape is (N, T, s_range, C) during
+            training or (N, C) in case of streaming decoding.
           Return a tensor of shape (N, T, s_range, C).
         """
-        assert encoder_out.ndim == decoder_out.ndim == 4
+        assert encoder_out.ndim == decoder_out.ndim
+        assert encoder_out.ndim in (2, 4)
         assert encoder_out.shape == decoder_out.shape
 
         logit = encoder_out + decoder_out

egs/librispeech/ASR/pruned_transducer_stateless3/asr_datamodule.py

@@ -277,10 +277,12 @@ class AsrDataModule:
                 cut_transforms=transforms,
                 return_cuts=self.args.return_cuts,
             )
-        valid_sampler = BucketingSampler(
+        valid_sampler = DynamicBucketingSampler(
             cuts_valid,
             max_duration=self.args.max_duration,
             shuffle=False,
+            num_buckets=self.args.num_buckets,
+            drop_last=False,
         )
         logging.info("About to create dev dataloader")
         valid_dl = DataLoader(
@@ -301,8 +303,12 @@ class AsrDataModule:
             else PrecomputedFeatures(),
             return_cuts=self.args.return_cuts,
         )
-        sampler = BucketingSampler(
-            cuts, max_duration=self.args.max_duration, shuffle=False
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
+            num_buckets=self.args.num_buckets,
+            drop_last=True,
         )
         logging.debug("About to create test dataloader")
         test_dl = DataLoader(

egs/librispeech/ASR/pruned_transducer_stateless3/train.py

@@ -799,7 +799,7 @@ def train_one_epoch(
                 f"tot_loss[{tot_loss}], "
                 f"libri_tot_loss[{libri_tot_loss}], "
                 f"giga_tot_loss[{giga_tot_loss}], "
-                f"batch size: {batch_size}"
+                f"batch size: {batch_size}, "
                 f"lr: {cur_lr:.2e}"
             )
 

egs/librispeech/ASR/pruned_transducer_stateless6/__init__.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless2/__init__.py

egs/librispeech/ASR/pruned_transducer_stateless6/asr_datamodule.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless2/asr_datamodule.py

egs/librispeech/ASR/pruned_transducer_stateless6/beam_search.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless2/beam_search.py

egs/librispeech/ASR/pruned_transducer_stateless6/decode.py

@@ -0,0 +1,634 @@
+#!/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_stateless6/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless6/exp \
+    --max-duration 600 \
+    --decoding-method greedy_search
+
+(2) beam search (not recommended)
+./pruned_transducer_stateless6/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless6/exp \
+    --max-duration 600 \
+    --decoding-method beam_search \
+    --beam-size 4
+
+(3) modified beam search
+./pruned_transducer_stateless6/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless6/exp \
+    --max-duration 600 \
+    --decoding-method modified_beam_search \
+    --beam-size 4
+
+(4) fast beam search
+./pruned_transducer_stateless6/decode.py \
+    --epoch 30 \
+    --avg 15 \
+    --exp-dir ./pruned_transducer_stateless6/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 sentencepiece as spm
+import torch
+import torch.nn as nn
+from asr_datamodule import LibriSpeechAsrDataModule
+from beam_search import (
+    beam_search,
+    fast_beam_search_one_best,
+    greedy_search,
+    greedy_search_batch,
+    modified_beam_search,
+)
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    average_checkpoints_with_averaged_model,
+    find_checkpoints,
+    load_checkpoint,
+)
+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_stateless6/exp",
+        help="The experiment dir",
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    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=2,
+        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""",
+    )
+
+    return parser
+
+
+def decode_one_batch(
+    params: AttributeDict,
+    model: nn.Module,
+    sp: spm.SentencePieceProcessor,
+    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.
+      sp:
+        The BPE model.
+      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)
+
+    layer_results, encoder_out_lens = model.encoder(
+        x=feature, x_lens=feature_lens
+    )
+    encoder_out = layer_results[-1]
+    hyps = []
+
+    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,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    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,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    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,
+        )
+        for hyp in sp.decode(hyp_tokens):
+            hyps.append(hyp.split())
+    else:
+        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}"
+                )
+            hyps.append(sp.decode(hyp).split())
+
+    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,
+    sp: spm.SentencePieceProcessor,
+    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.
+      sp:
+        The BPE model.
+      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 = 10
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+        texts = batch["supervisions"]["text"]
+
+        hyps_dict = decode_one_batch(
+            params=params,
+            model=model,
+            sp=sp,
+            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"
+        )
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, 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\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER 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()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    args = parser.parse_args()
+    args.exp_dir = Path(args.exp_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}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> and <unk> are defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.unk_id = sp.piece_to_id("<unk>")
+    params.vocab_size = sp.get_piece_size()
+
+    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(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}")
+
+    librispeech = LibriSpeechAsrDataModule(args)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            params=params,
+            model=model,
+            sp=sp,
+            decoding_graph=decoding_graph,
+        )
+
+        save_results(
+            params=params,
+            test_set_name=test_set,
+            results_dict=results_dict,
+        )
+
+    logging.info("Done!")
+
+
+if __name__ == "__main__":
+    main()

egs/librispeech/ASR/pruned_transducer_stateless6/decoder.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless2/decoder.py

egs/librispeech/ASR/pruned_transducer_stateless6/encoder_interface.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless2/encoder_interface.py

egs/librispeech/ASR/pruned_transducer_stateless6/export.py

@@ -0,0 +1,217 @@
+#!/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_stateless2/export.py \
+  --exp-dir ./pruned_transducer_stateless2/exp \
+  --bpe-model data/lang_bpe_500/bpe.model \
+  --epoch 20 \
+  --avg 10
+
+It will generate a file exp_dir/pretrained.pt
+
+To use the generated file with `pruned_transducer_stateless2/decode.py`,
+you can do:
+
+    cd /path/to/exp_dir
+    ln -s pretrained.pt epoch-9999.pt
+
+    cd /path/to/egs/librispeech/ASR
+    ./pruned_transducer_stateless2/decode.py \
+        --exp-dir ./pruned_transducer_stateless2/exp \
+        --epoch 9999 \
+        --avg 1 \
+        --max-duration 100 \
+        --bpe-model data/lang_bpe_500/bpe.model
+"""
+
+import argparse
+import logging
+from pathlib import Path
+
+import sentencepiece as spm
+import torch
+from train import get_params, get_transducer_model
+
+from icefall.checkpoint import (
+    average_checkpoints,
+    find_checkpoints,
+    load_checkpoint,
+)
+from icefall.utils import str2bool
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--epoch",
+        type=int,
+        default=28,
+        help="""It specifies the checkpoint to use for averaging.
+        Note: Epoch counts from 0.
+        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(
+        "--exp-dir",
+        type=str,
+        default="pruned_transducer_stateless2/exp",
+        help="""It specifies the directory where all training related
+        files, e.g., checkpoints, log, etc, are saved
+        """,
+    )
+
+    parser.add_argument(
+        "--bpe-model",
+        type=str,
+        default="data/lang_bpe_500/bpe.model",
+        help="Path to the BPE model",
+    )
+
+    parser.add_argument(
+        "--jit",
+        type=str2bool,
+        default=False,
+        help="""True to save a model after applying torch.jit.script.
+        """,
+    )
+
+    parser.add_argument(
+        "--context-size",
+        type=int,
+        default=2,
+        help="The context size in the decoder. 1 means bigram; "
+        "2 means tri-gram",
+    )
+
+    return parser
+
+
+def main():
+    args = get_parser().parse_args()
+    args.exp_dir = Path(args.exp_dir)
+
+    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}")
+
+    sp = spm.SentencePieceProcessor()
+    sp.load(params.bpe_model)
+
+    # <blk> is defined in local/train_bpe_model.py
+    params.blank_id = sp.piece_to_id("<blk>")
+    params.vocab_size = sp.get_piece_size()
+
+    logging.info(params)
+
+    logging.info("About to create model")
+    model = get_transducer_model(params)
+
+    model.to(device)
+
+    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 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()
+
+    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"
+        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()

egs/librispeech/ASR/pruned_transducer_stateless6/extract_codebook_index.py

@@ -0,0 +1,80 @@
+#!/usr/bin/env python3
+# Copyright 2022 Xiaomi Corporation (Author: Liyong Guo)
+#
+# 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 argparse
+import os
+from pathlib import Path
+
+import torch
+from vq_utils import CodebookIndexExtractor
+from asr_datamodule import LibriSpeechAsrDataModule
+from hubert_xlarge import HubertXlargeFineTuned
+from icefall.utils import AttributeDict
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+    parser.add_argument(
+        "--exp-dir",
+        type=Path,
+        default="pruned_transducer_stateless6/exp/",
+        help="The experiment dir",
+    )
+
+    return parser
+
+
+def get_world_size():
+    warn_message = (
+        "It's better to use GPU to extrac codebook indices"
+        "Please set with commonds like: export CUDA_VISIBLE_DEVICES=0,1,2,3"
+    )
+    assert (
+        torch.cuda.is_available() and "CUDA_VISIBLE_DEVICES" in os.environ
+    ), warn_message
+    world_size = len(os.environ["CUDA_VISIBLE_DEVICES"].split(","))
+    assert world_size > 0, warn_message
+    return world_size
+
+
+def main():
+    world_size = get_world_size()
+    parser = get_parser()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    HubertXlargeFineTuned.add_arguments(parser)
+    CodebookIndexExtractor.add_arguments(parser)
+
+    args = parser.parse_args()
+    params = AttributeDict()
+    params.update(vars(args))
+
+    # reset some parameters needed by hubert.
+    params.update(HubertXlargeFineTuned.get_params())
+    params.device = torch.device("cuda", 0)
+    params.world_size = world_size
+
+    extractor = CodebookIndexExtractor(params=params)
+    extractor.extract_and_save_embedding()
+    extractor.train_quantizer()
+    extractor.extract_codebook_indexes()
+
+
+if __name__ == "__main__":
+    main()

egs/librispeech/ASR/pruned_transducer_stateless6/hubert_decode.py

@@ -0,0 +1,205 @@
+#!/usr/bin/env python3
+# Copyright 2022 Xiaomi Corporation (Author: Liyong Guo)
+#
+# 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 argparse
+import logging
+from collections import defaultdict
+from pathlib import Path
+from typing import Dict, List, Tuple
+
+import torch
+
+from asr_datamodule import LibriSpeechAsrDataModule
+from hubert_xlarge import HubertXlargeFineTuned
+
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+    store_transcripts,
+    write_error_stats,
+)
+
+
+def get_parser():
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+
+    parser.add_argument(
+        "--exp-dir",
+        type=Path,
+        default="pruned_transducer_stateless6/exp/",
+        help="The experiment dir",
+    )
+
+    return parser
+
+
+def decode_dataset(
+    dl: torch.utils.data.DataLoader,
+    hubert_model: HubertXlargeFineTuned,
+    params: AttributeDict,
+) -> Dict[str, List[Tuple[List[str], List[str]]]]:
+    """Decode dataset.
+
+    Args:
+      dl:
+        PyTorch's dataloader containing the dataset to decode.
+      model:
+        The neural model.
+
+    Returns:
+      Return a dict, whose key is decoding method "ctc_greedy_search".
+      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.
+    """
+    results = []
+
+    num_cuts = 0
+
+    try:
+        num_batches = len(dl)
+    except TypeError:
+        num_batches = "?"
+
+    results = defaultdict(list)
+    for batch_idx, batch in enumerate(dl):
+
+        hyps = hubert_model.ctc_greedy_search(batch)
+
+        texts = batch["supervisions"]["text"]
+        assert len(hyps) == len(texts)
+        this_batch = []
+
+        for hyp_text, ref_text in zip(hyps, texts):
+            ref_words = ref_text.split()
+            hyp_words = hyp_text.split()
+            this_batch.append((ref_words, hyp_words))
+
+        results["ctc_greedy_search"].extend(this_batch)
+
+        num_cuts += len(texts)
+
+        if batch_idx % 20 == 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}.txt"
+        store_transcripts(filename=recog_path, texts=results)
+
+        # 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}.txt"
+        with open(errs_filename, "w") as f:
+            wer = write_error_stats(
+                f, f"{test_set_name}-{key}", results, 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}.txt"
+    with open(errs_info, "w") as f:
+        print("settings\tWER", file=f)
+        for key, val in test_set_wers:
+            print("{}\t{}".format(key, val), file=f)
+
+    s = "\nFor {}, WER 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()
+    LibriSpeechAsrDataModule.add_arguments(parser)
+    HubertXlargeFineTuned.add_arguments(parser)
+    args = parser.parse_args()
+
+    params = AttributeDict()
+    params.update(vars(args))
+    # reset some parameters needed by hubert.
+    params.update(HubertXlargeFineTuned.get_params())
+
+    params.res_dir = (
+        params.exp_dir / f"ctc_greedy_search-{params.teacher_model_id}"
+    )
+
+    setup_logger(f"{params.res_dir}/log/log-ctc_greedy_search")
+    logging.info("Decoding started")
+    logging.info(params)
+
+    device = torch.device("cpu")
+    if torch.cuda.is_available():
+        device = torch.device("cuda", 0)
+
+    logging.info(f"device: {device}")
+    params.device = device
+
+    hubert_model = HubertXlargeFineTuned(params)
+
+    librispeech = LibriSpeechAsrDataModule(params)
+
+    test_clean_cuts = librispeech.test_clean_cuts()
+    test_other_cuts = librispeech.test_other_cuts()
+
+    test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
+    test_other_dl = librispeech.test_dataloaders(test_other_cuts)
+
+    test_sets = ["test-clean", "test-other"]
+    test_dl = [test_clean_dl, test_other_dl]
+
+    for test_set, test_dl in zip(test_sets, test_dl):
+        results_dict = decode_dataset(
+            dl=test_dl,
+            hubert_model=hubert_model,
+            params=params,
+        )
+
+        save_results(
+            params=params, test_set_name=test_set, results_dict=results_dict
+        )
+
+    logging.info("Done!")
+
+
+torch.set_num_threads(1)
+torch.set_num_interop_threads(1)
+
+if __name__ == "__main__":
+    main()

egs/librispeech/ASR/pruned_transducer_stateless6/hubert_xlarge.py

@@ -0,0 +1,220 @@
+#!/usr/bin/env python3
+# Copyright 2022 Xiaomi Corporation (Author: Liyong Guo)
+#
+# 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 argparse
+import logging
+from pathlib import Path
+from typing import Dict, List, Tuple
+
+import torch
+from fairseq import (
+    checkpoint_utils,
+    tasks,
+    utils,
+)
+from fairseq.data.data_utils import post_process
+from omegaconf import OmegaConf
+
+from icefall.utils import AttributeDict
+
+
+def _load_hubert_model(params: AttributeDict):
+    """
+    Load the hubert model.
+
+    The model loaded is specified by params.hubert_model_dir
+    and params.teacher_model_id.
+
+    Returned model carries hubert,
+    while processor is responsible to map model's output to human readable transcripts.
+    """
+    cfg_task = OmegaConf.create(
+        {
+            "_name": "hubert_pretraining",
+            "single_target": True,
+            "fine_tuning": True,
+            "data": str(params.hubert_model_dir),
+        }
+    )
+    model_path = Path(params.hubert_model_dir) / (
+        params.teacher_model_id + ".pt"
+    )
+    task = tasks.setup_task(cfg_task)
+    processor = task.target_dictionary
+    models, saved_cfg = checkpoint_utils.load_model_ensemble(
+        utils.split_paths(str(model_path), separator="\\"),
+        arg_overrides={},
+        strict=True,
+        suffix="",
+        num_shards=1,
+    )
+    model = models[0]
+    model.to(params.device)
+    model.eval()
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    logging.info(f"Number of model parameters: {num_param}")
+
+    return model, processor
+
+
+class HubertXlargeFineTuned:
+    """
+    A wrapper of hubert extra large fine-tuned model.
+
+    A teacher model is responsible for:
+        1. load teacher model
+        2. extracting embeddings to train quantizer.
+        3. extract codebook indices
+        4. verify its performance with ctc_greedy_search method.
+    """
+
+    def __init__(self, params: AttributeDict):
+        self.model, self.processor = _load_hubert_model(params)
+        self.w2v_model = self.model.w2v_encoder.w2v_model
+        self.params = params
+
+    @staticmethod
+    def get_params() -> AttributeDict:
+        """Return a dict containing parameters defined in other modules.
+
+        Their default value conflits to hubert's requirements so they are reset as following.
+        """
+        params = AttributeDict(
+            {
+                # parameters defined in asr_datamodule.py
+                "input_strategy": "AudioSamples",
+                "enable_musan": False,
+                "enable_spec_aug": False,
+                "return_cuts": True,
+                "drop_last": False,
+                # parameters used by quantizer
+                "embedding_dim": 1280,
+            }
+        )
+        return params
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        # Options about model loading.
+        parser.add_argument(
+            "--hubert-model-dir",
+            type=Path,
+            default="./pruned_transducer_stateless6/exp/hubert_models/",
+            help="path to save downloaded hubert models.",
+        )
+
+        parser.add_argument(
+            "--teacher-model-id",
+            type=str,
+            default="hubert_xtralarge_ll60k_finetune_ls960",
+            help="""could be one of:
+            [
+                "hubert_xtralarge_ll60k_finetune_ls960",  # fine-tuned model.
+                "hubert_xtralarge_ll60k.pt",  # pretrained model without fintuing.
+            ]""",
+        )
+        parser.add_argument(
+            "--total-layers",
+            type=int,
+            default=48,
+        )
+
+    # Modified from HubertModel.forward to extract all middle layers output
+    def extract_layers_result(
+        self,
+        batch: Dict,
+    ) -> List[torch.Tensor]:
+        """
+        Extract activations from all layers.
+        """
+        features = batch["inputs"]
+
+        # corresponding task.normalize in fairseq
+        features = torch.nn.functional.layer_norm(features, features.shape)
+
+        supervisions = batch["supervisions"]
+        num_samples = supervisions["num_samples"]
+        B, T = features.shape
+        padding_mask = torch.arange(0, T).expand(B, T) > num_samples.reshape(
+            [-1, 1]
+        )
+
+        padding_mask = padding_mask.to(self.params.device)
+        features = features.to(self.params.device)
+
+        features = self.w2v_model.forward_features(features)
+
+        features = features.transpose(1, 2)
+        features = self.w2v_model.layer_norm(features)
+
+        padding_mask = self.w2v_model.forward_padding_mask(
+            features, padding_mask
+        )
+
+        if self.w2v_model.post_extract_proj is not None:
+            features = self.w2v_model.post_extract_proj(features)
+
+        _, layer_results = self.w2v_model.encoder(
+            features,
+            padding_mask=padding_mask,
+        )
+        return layer_results
+
+    def extract_embedding(self, batch) -> Tuple[torch.tensor, List[int]]:
+        """
+        Eextract embeddings specified by self.params.embedding_layer.
+
+        These embeddings could be used to train quantizer
+        or to extract codebook indexes.
+
+        The returned List[int] is valid length of each embedding.
+        We only want to store codebook indexes related to
+        these valid embeddings.
+        """
+        supervisions = batch["supervisions"]
+        cut_list = supervisions["cut"]
+        assert all(c.start == 0 for c in cut_list)
+        layer_results = self.extract_layers_result(batch)
+        embeddings = layer_results[self.params.embedding_layer - 1][0]
+        encoder_embedding = embeddings.transpose(0, 1)  # N, T, C
+        N = encoder_embedding.shape[0]
+        assert len(cut_list) == N
+        # 320 is from: 16,000 / 50 = sample_rate / hbuert output frame rate
+        num_frames = (supervisions["num_samples"] // 320).tolist()
+        return encoder_embedding, num_frames
+
+    def ctc_greedy_search(self, batch):
+        """
+        Mainly used to verify hubert model is used correctly.
+        """
+        layer_results = self.extract_layers_result(batch=batch)
+        encoder_out = self.w2v_model.encoder.layer_norm(
+            layer_results[self.params.total_layers - 1][0]
+        )
+        encoder_out = self.model.w2v_encoder.proj(encoder_out.transpose(0, 1))
+
+        toks = encoder_out.argmax(dim=-1)
+        blank = 0
+        toks = [tok.unique_consecutive() for tok in toks]
+        hyps = [
+            self.processor.string(tok[tok != blank].int().cpu()) for tok in toks
+        ]
+        hyps = [post_process(hyp, "letter") for hyp in hyps]
+
+        return hyps

egs/librispeech/ASR/pruned_transducer_stateless6/joiner.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless2/joiner.py

egs/librispeech/ASR/pruned_transducer_stateless6/model.py

@@ -0,0 +1,249 @@
+# 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.
+
+
+import k2
+import torch
+import torch.nn as nn
+from encoder_interface import EncoderInterface
+from scaling import ScaledLinear
+
+from icefall.utils import add_sos
+
+from quantization.prediction import JointCodebookLoss
+
+
+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,
+        num_codebooks: int = 0,
+    ):
+        """
+        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.
+          num_codebooks:
+            Used by distillation loss.
+        """
+        super().__init__()
+        assert isinstance(encoder, EncoderInterface), type(encoder)
+        assert hasattr(decoder, "blank_id")
+
+        self.encoder = encoder
+        self.decoder = decoder
+        self.joiner = joiner
+
+        self.simple_am_proj = ScaledLinear(
+            encoder_dim, vocab_size, initial_speed=0.5
+        )
+        self.simple_lm_proj = ScaledLinear(decoder_dim, vocab_size)
+        if num_codebooks > 0:
+            self.codebook_loss_net = JointCodebookLoss(
+                predictor_channels=encoder_dim, num_codebooks=num_codebooks
+            )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_lens: torch.Tensor,
+        y: k2.RaggedTensor,
+        prune_range: int = 5,
+        am_scale: float = 0.0,
+        lm_scale: float = 0.0,
+        warmup: float = 1.0,
+        codebook_indexes: torch.Tensor = None,
+    ) -> 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.
+          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.
+          codebook_indexes:
+            codebook_indexes extracted from a teacher model.
+        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
+
+        layer_results, x_lens = self.encoder(x, x_lens, warmup=warmup)
+        encoder_out = layer_results[-1]
+        middle_layer_output = layer_results[0]
+        if self.training and codebook_indexes is not None:
+            assert hasattr(self, "codebook_loss_net")
+            if codebook_indexes.shape[1] != middle_layer_output.shape[1]:
+                codebook_indexes = self.concat_successive_codebook_indexes(
+                    middle_layer_output, codebook_indexes
+                )
+            codebook_loss = self.codebook_loss_net(
+                middle_layer_output, codebook_indexes
+            )
+        else:
+            # when codebook index is not available.
+            codebook_loss = None
+
+        assert torch.all(x_lens > 0)
+
+        # 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 = self.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 = self.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] = x_lens
+
+        lm = self.simple_lm_proj(decoder_out)
+        am = self.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=self.joiner.encoder_proj(encoder_out),
+            lm=self.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 = self.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, codebook_loss)
+
+    @staticmethod
+    def concat_successive_codebook_indexes(
+        middle_layer_output, codebook_indexes
+    ):
+        # Output rate of hubert is 50 frames per second,
+        # while that of current encoder is 25.
+        # Following code handling two issues:
+        # 1.
+        #   Roughly speaking, to generate another frame output,
+        #   hubert needes extra two frames,
+        #   while current encoder needs extra four frames.
+        #   Suppose there are only extra three frames provided,
+        #   hubert will generate another frame while current encoder does nothing.
+        # 2.
+        #   codebook loss is a frame-wise loss, to enalbe 25 frames studnet output
+        #   learns from 50 frames teacher output, two successive frames of teacher model
+        #   output is concatenated together.
+        t_expected = middle_layer_output.shape[1]
+        N, T, C = codebook_indexes.shape
+
+        # Handling issue 1.
+        if T >= t_expected * 2:
+            codebook_indexes = codebook_indexes[:, : t_expected * 2, :]
+        # Handling issue 2.
+        codebook_indexes = codebook_indexes.reshape(N, t_expected, C * 2)
+        assert middle_layer_output.shape[1] == codebook_indexes.shape[1]
+        return codebook_indexes

egs/librispeech/ASR/pruned_transducer_stateless6/optim.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless2/optim.py

egs/librispeech/ASR/pruned_transducer_stateless6/scaling.py

@@ -0,0 +1 @@
+../pruned_transducer_stateless2/scaling.py

egs/librispeech/ASR/pruned_transducer_stateless6/test_model.py

@@ -0,0 +1,51 @@
+#!/usr/bin/env python3
+# Copyright    2022  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.
+
+
+"""
+To run this file, do:
+
+    cd icefall/egs/librispeech/ASR
+    python ./pruned_transducer_stateless6/test_model.py
+"""
+
+import torch
+from train import get_params, get_transducer_model
+
+
+def test_model():
+    params = get_params()
+    params.vocab_size = 500
+    params.blank_id = 0
+    params.context_size = 2
+    params.unk_id = 2
+    params.enable_distiallation = False
+
+    model = get_transducer_model(params)
+
+    num_param = sum([p.numel() for p in model.parameters()])
+    print(f"Number of model parameters: {num_param}")
+    model.__class__.forward = torch.jit.ignore(model.__class__.forward)
+    torch.jit.script(model)
+
+
+def main():
+    test_model()
+
+
+if __name__ == "__main__":
+    main()

egs/librispeech/ASR/pruned_transducer_stateless6/vq_utils.py

@@ -0,0 +1,399 @@
+#!/usr/bin/env python3
+# Copyright 2022 Xiaomi Corporation (Author: Liyong Guo)
+#
+# 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 argparse
+import copy
+import glob
+import logging
+import os
+from functools import cached_property
+from pathlib import Path
+from typing import List, Tuple
+
+import numpy as np
+import torch
+import torch.multiprocessing as mp
+import quantization
+
+from asr_datamodule import LibriSpeechAsrDataModule
+from hubert_xlarge import HubertXlargeFineTuned
+from icefall.utils import (
+    AttributeDict,
+    setup_logger,
+)
+from lhotse import CutSet, load_manifest
+from lhotse.features.io import NumpyHdf5Writer
+
+
+class CodebookIndexExtractor:
+    """
+    A wrapper of quantiation.Quantizer.
+
+    It's responsible for:
+        1. extract and save activations from a teacher model.
+        2. train quantizer from previous activations.
+        3. extract codebook indexes for whole training set.
+           Normally this step needs multi GPUs.
+    """
+
+    def __init__(self, params: AttributeDict):
+
+        self.params = params
+        params.subsets = ["clean-100"]
+        if self.params.full_libri:
+            self.params.subsets += ["clean-360", "other-500"]
+
+        self.init_dirs()
+        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
+        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
+        self.manifest_dir = self.vq_dir / f"splits{self.params.world_size}"
+        self.manifest_dir.mkdir(parents=True, exist_ok=True)
+
+        # It's doesn't matter whether ori_manifest_dir is str or Path.
+        # Set it to Path to be consistent.
+        self.ori_manifest_dir = Path("./data/fbank/")
+        self.dst_manifest_dir = Path("./data/vq_fbank/")
+
+        self.dst_manifest_dir.mkdir(parents=True, exist_ok=True)
+
+    @classmethod
+    def add_arguments(cls, parser: argparse.ArgumentParser):
+        # Options about teacher embeddings eatraction.
+        parser.add_argument(
+            "--embedding-layer",
+            type=int,
+            help="layer to extract teacher embeddings, 1-based.",
+            default=36,
+        )
+
+        parser.add_argument(
+            "--num-utts",
+            type=int,
+            default=1000,
+            help="num utts to train quantizer",
+        )
+
+        parser.add_argument(
+            "--num-codebooks",
+            type=int,
+            default=8,
+            help="""number of codebooks,
+            i.e. number of codebook indexes each teacher embedding is compressed.
+            """,
+        )
+
+    @property
+    def embedding_file_path(self):
+        """
+        The saved embedding is used to train quantizer.
+        """
+        embedding_file_id = (
+            f"num_utts_{self.params.num_utts}"
+            + f"-layer_{self.params.embedding_layer}"
+            + "-embedding_embeddings.h5"
+        )
+
+        embedding_file_path = self.vq_dir / embedding_file_id
+        return embedding_file_path
+
+    @torch.no_grad()
+    def extract_and_save_embedding(self):
+        """
+        The extract embedding is used to train quantizer.
+        """
+        if self.embedding_file_path.exists():
+            warn_message = (
+                f"{self.embedding_file_path} already exists."
+                + " Skip extracting embeddings from teacher model"
+            )
+            logging.warn(warn_message)
+            return
+
+        total_cuts = 0
+        with NumpyHdf5Writer(self.embedding_file_path) as writer:
+            for batch_idx, batch in enumerate(self.quantizer_train_dl):
+                cut_list = batch["supervisions"]["cut"]
+                (
+                    encoder_embedding,
+                    num_frames,
+                ) = self.teacher_model.extract_embedding(batch)
+                encoder_embedding = encoder_embedding.cpu().numpy()
+                for idx, cut in enumerate(cut_list):
+                    cut.encoder_embedding = writer.store_array(
+                        key=cut.id,
+                        value=encoder_embedding[idx][: num_frames[idx]],
+                    )
+                total_cuts += len(cut_list)
+                logging.info(
+                    f"Processed {total_cuts} output of {self.params.num_utts} cuts."
+                )
+
+        logging.info(f"Processed all {total_cuts} cuts.")
+
+    @property
+    def quantizer_train_dl(self):
+        # used to train quantizer.
+        librispeech = LibriSpeechAsrDataModule(self.params)
+        quantizer_trian_cuts = librispeech.train_clean_100_cuts().subset(
+            first=self.params.num_utts
+        )
+        return librispeech.train_dataloaders(quantizer_trian_cuts)
+
+    @cached_property
+    def quantizer_file_path(self):
+        quantizer_file_id = (
+            f"num_utts-{self.params.num_utts}"
+            + f"-layer-{self.params.embedding_layer}"
+            + f"-num_codebooks_{self.params.num_codebooks}"
+            + "-quantizer.pt"
+        )
+        quantizer_file_path = Path(self.vq_dir) / quantizer_file_id
+
+        return quantizer_file_path
+
+    def train_quantizer(self):
+        if self.quantizer_file_path.exists():
+            warn_message = (
+                f"{self.quantizer_file_path} already exists."
+                + " Skip trainning quantizer."
+            )
+            logging.warn(warn_message)
+            return
+
+        assert self.embedding_file_path.exists()
+        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.
+
+        def minibatch_generator(data: torch.Tensor, repeat: bool):
+            assert 3 * B < data.shape[0]
+            cur_offset = 0
+            while True if repeat else cur_offset + B <= data.shape[0]:
+                start = cur_offset % (data.shape[0] + 1 - B)
+                end = start + B
+                cur_offset += B
+                yield data[start:end, :].to(self.params.device).to(
+                    dtype=torch.float
+                )
+
+        for x in minibatch_generator(train, repeat=True):
+            trainer.step(x)
+            if trainer.done():
+                break
+
+        quantizer = trainer.get_quantizer()
+        torch.save(quantizer.state_dict(), self.quantizer_file_path)
+
+    def split_ori_manifests(self):
+        """
+        When multi gpus are available, split original manifests
+        and extract codebook indexes in a prallel way.
+        """
+        for subset in self.params.subsets:
+            logging.info(f"About to split {subset}.")
+            ori_manifest = f"./data/fbank/cuts_train-{subset}.json.gz"
+            split_cmd = f"lhotse split {self.params.world_size} {ori_manifest} {self.manifest_dir}"
+            os.system(f"{split_cmd}")
+
+    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"
+            dst_vq_manifest = (
+                self.dst_manifest_dir / f"cuts_train-{subset}.json.gz"
+            )
+            if 1 == self.params.world_size:
+                merge_cmd = f"cp {vq_manifests} {dst_vq_manifest}"
+            else:
+                merge_cmd = f"lhotse combine {vq_manifests} {dst_vq_manifest}"
+            os.system(f"{merge_cmd}")
+
+    def reuse_manifests(self):
+        """
+        Only train-* subsets are extracted codebook indexes from.
+        The reset subsets are just a link from ./data/fbank.
+        """
+
+        def is_train(manifest: str) -> bool:
+            for train_subset in ["clean-100", "clean-360", "other-500"]:
+                if train_subset in manifest:
+                    return True
+            return False
+
+        # Type of self.ori_nanifest_dir is Path
+        # and result type of glob.glob is str.
+        reusable_manifests = [
+            manifest
+            for manifest in glob.glob(f"{self.ori_manifest_dir}/*.gz")
+            if not is_train(manifest)
+        ]
+        for manifest_path in reusable_manifests:
+            ori_manifest_path = Path(manifest_path).resolve()
+            # Path cannot used as a parameter of str.replace.
+            # Cast them to str.
+            dst_manifest_path = Path(
+                manifest_path.replace(
+                    str(self.ori_manifest_dir), str(self.dst_manifest_dir)
+                )
+            ).resolve()
+            if not dst_manifest_path.exists():
+                os.symlink(ori_manifest_path, dst_manifest_path)
+
+    def create_vq_fbank(self):
+        self.reuse_manifests()
+        self.merge_vq_manifests()
+
+    @cached_property
+    def teacher_model(self):
+        return HubertXlargeFineTuned(self.params)
+
+    @cached_property
+    def quantizer(self):
+        assert self.quantizer_file_path.exists()
+        quantizer = quantization.Quantizer(
+            dim=self.params.embedding_dim,
+            num_codebooks=self.params.num_codebooks,
+            codebook_size=256,
+        )
+        quantizer.load_state_dict(torch.load(self.quantizer_file_path))
+        quantizer.to(self.params.device)
+        return quantizer
+
+    def load_ori_dl(self, subset):
+        if self.params.world_size == 1:
+            ori_manifest_path = f"./data/fbank/cuts_train-{subset}.json.gz"
+        else:
+            ori_manifest_path = (
+                self.manifest_dir
+                / f"cuts_train-{subset}.{self.params.manifest_index}.json.gz"
+            )
+
+        cuts = load_manifest(ori_manifest_path)
+        dl = LibriSpeechAsrDataModule(self.params).train_dataloaders(cuts)
+        return dl
+
+    def _release_gpu_memory(self):
+        self.__dict__.pop("teacher_model", None)
+        self.__dict__.pop("quantizer", None)
+        torch.cuda.empty_cache()
+
+    def extract_codebook_indexes(self):
+        if self.params.world_size == 1:
+            self.extract_codebook_indexes_imp()
+        else:
+            # Since a new extractor will be created for each rank in
+            # compute_codebook_indexes_parallel, it's better to
+            # release the GPU memory occupied by current extractor.
+            self._release_gpu_memory()
+
+            # Prepare split manifests for each job.
+            self.split_ori_manifests()
+            mp.spawn(
+                compute_codebook_indexes_parallel,
+                args=(self.params,),
+                nprocs=self.params.world_size,
+                join=True,
+            )
+        self.create_vq_fbank()
+
+    @torch.no_grad()
+    def extract_codebook_indexes_imp(self):
+        for subset in self.params.subsets:
+            num_cuts = 0
+            cuts = []
+            if self.params.world_size == 1:
+                manifest_file_id = f"{subset}"
+            else:
+                manifest_file_id = f"{subset}-{self.params.manifest_index}"
+
+            manifest_file_path = self.manifest_dir / manifest_file_id
+            with NumpyHdf5Writer(manifest_file_path) as writer:
+                for batch_idx, batch in enumerate(self.load_ori_dl(subset)):
+                    (
+                        encoder_embedding,
+                        num_frames,
+                    ) = self.teacher_model.extract_embedding(batch)
+                    codebook_indexes = self.quantizer.encode(encoder_embedding)
+                    # [N, T, C]
+                    codebook_indexes = codebook_indexes.to("cpu").numpy()
+                    assert np.min(codebook_indexes) >= 0
+                    assert np.max(codebook_indexes) < 256
+                    supervisions = batch["supervisions"]
+                    cut_list = supervisions["cut"]
+                    assert len(cut_list) == codebook_indexes.shape[0]
+                    assert all(c.start == 0 for c in supervisions["cut"])
+
+                    for idx, cut in enumerate(cut_list):
+                        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
+                    num_cuts += len(cut_list)
+                    message = f"Processed {num_cuts} cuts from {subset}"
+                    if self.params.world_size > 1:
+                        message += f" by job {self.params.manifest_index}"
+                    logging.info(f"{message}.")
+
+                json_file_path = (
+                    self.manifest_dir
+                    / f"with_codebook_indexes-cuts_train-{manifest_file_id}.json.gz"
+                )
+                CutSet.from_cuts(cuts).to_json(json_file_path)
+
+
+@torch.no_grad()
+def compute_codebook_indexes_parallel(
+    rank: int,
+    params,
+) -> List[Tuple[str, List[int]]]:
+    """Create an extractor for each rank and extract codebook indexes parallelly.
+
+    Normally, this function is called by torch.multiprocessing
+    when multi GPUs are available.
+    """
+    params = copy.deepcopy(params)
+    device = torch.device("cuda", rank)
+    params.device = device
+
+    # rank is 0-based while split manifests by "lhotse split" is 1-based.
+    params.manifest_index = rank + 1
+
+    extractor = CodebookIndexExtractor(params=params)
+    extractor.extract_codebook_indexes_imp()

egs/librispeech/ASR/tdnn_lstm_ctc/asr_datamodule.py

@@ -25,16 +25,19 @@ from typing import Any, Dict, Optional
 
 import torch
 from lhotse import CutSet, Fbank, FbankConfig, load_manifest
-from lhotse.dataset import (
-    BucketingSampler,
+from lhotse.dataset import (  # noqa F401 for PrecomputedFeatures
     CutConcatenate,
     CutMix,
+    DynamicBucketingSampler,
     K2SpeechRecognitionDataset,
     PrecomputedFeatures,
     SingleCutSampler,
     SpecAugment,
 )
-from lhotse.dataset.input_strategies import OnTheFlyFeatures
+from lhotse.dataset.input_strategies import (  # noqa F401 For AudioSamples
+    AudioSamples,
+    OnTheFlyFeatures,
+)
 from lhotse.utils import fix_random_seed
 from torch.utils.data import DataLoader
 
@@ -110,7 +113,7 @@ class LibriSpeechAsrDataModule:
             "--num-buckets",
             type=int,
             default=30,
-            help="The number of buckets for the BucketingSampler"
+            help="The number of buckets for the DynamicBucketingSampler"
             "(you might want to increase it for larger datasets).",
         )
         group.add_argument(
@@ -150,6 +153,12 @@ class LibriSpeechAsrDataModule:
             help="When enabled (=default), the examples will be "
             "shuffled for each epoch.",
         )
+        group.add_argument(
+            "--drop-last",
+            type=str2bool,
+            default=True,
+            help="Whether to drop last batch. Used by sampler.",
+        )
         group.add_argument(
             "--return-cuts",
             type=str2bool,
@@ -192,6 +201,13 @@ class LibriSpeechAsrDataModule:
             "with training dataset. ",
         )
 
+        group.add_argument(
+            "--input-strategy",
+            type=str,
+            default="PrecomputedFeatures",
+            help="AudioSamples or PrecomputedFeatures",
+        )
+
     def train_dataloaders(
         self,
         cuts_train: CutSet,
@@ -263,6 +279,7 @@ class LibriSpeechAsrDataModule:
 
         logging.info("About to create train dataset")
         train = K2SpeechRecognitionDataset(
+            input_strategy=eval(self.args.input_strategy)(),
             cut_transforms=transforms,
             input_transforms=input_transforms,
             return_cuts=self.args.return_cuts,
@@ -289,14 +306,13 @@ class LibriSpeechAsrDataModule:
             )
 
         if self.args.bucketing_sampler:
-            logging.info("Using BucketingSampler.")
-            train_sampler = BucketingSampler(
+            logging.info("Using DynamicBucketingSampler.")
+            train_sampler = DynamicBucketingSampler(
                 cuts_train,
                 max_duration=self.args.max_duration,
                 shuffle=self.args.shuffle,
                 num_buckets=self.args.num_buckets,
-                bucket_method="equal_duration",
-                drop_last=True,
+                drop_last=self.args.drop_last,
             )
         else:
             logging.info("Using SingleCutSampler.")
@@ -350,7 +366,7 @@ class LibriSpeechAsrDataModule:
                 cut_transforms=transforms,
                 return_cuts=self.args.return_cuts,
             )
-        valid_sampler = BucketingSampler(
+        valid_sampler = DynamicBucketingSampler(
             cuts_valid,
             max_duration=self.args.max_duration,
             shuffle=False,
@@ -371,11 +387,13 @@ class LibriSpeechAsrDataModule:
         test = K2SpeechRecognitionDataset(
             input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
             if self.args.on_the_fly_feats
-            else PrecomputedFeatures(),
+            else eval(self.args.input_strategy)(),
             return_cuts=self.args.return_cuts,
         )
-        sampler = BucketingSampler(
-            cuts, max_duration=self.args.max_duration, shuffle=False
+        sampler = DynamicBucketingSampler(
+            cuts,
+            max_duration=self.args.max_duration,
+            shuffle=False,
         )
         logging.debug("About to create test dataloader")
         test_dl = DataLoader(

icefall/utils.py

@@ -127,7 +127,11 @@ def setup_logger(
         level = logging.CRITICAL
 
     logging.basicConfig(
-        filename=log_filename, format=formatter, level=level, filemode="w"
+        filename=log_filename,
+        format=formatter,
+        level=level,
+        filemode="w",
+        force=True,
     )
     if use_console:
         console = logging.StreamHandler()