train/decode with codebook loss

egs/librispeech/ASR/vq_pruned_transducer_stateless2/decode.py

@@ -18,36 +18,36 @@
 """
 Usage:
 (1) greedy search
-./pruned_transducer_stateless2/decode.py \
+./vq_pruned_transducer_stateless2/decode.py \
         --epoch 28 \
         --avg 15 \
-        --exp-dir ./pruned_transducer_stateless2/exp \
+        --exp-dir ./vq_pruned_transducer_stateless2/exp \
         --max-duration 100 \
         --decoding-method greedy_search
 
 (2) beam search
-./pruned_transducer_stateless2/decode.py \
+./vq_pruned_transducer_stateless2/decode.py \
         --epoch 28 \
         --avg 15 \
-        --exp-dir ./pruned_transducer_stateless2/exp \
+        --exp-dir ./vq_pruned_transducer_stateless2/exp \
         --max-duration 100 \
         --decoding-method beam_search \
         --beam-size 4
 
 (3) modified beam search
-./pruned_transducer_stateless2/decode.py \
+./vq_pruned_transducer_stateless2/decode.py \
         --epoch 28 \
         --avg 15 \
-        --exp-dir ./pruned_transducer_stateless2/exp \
+        --exp-dir ./vq_pruned_transducer_stateless2/exp \
         --max-duration 100 \
         --decoding-method modified_beam_search \
         --beam-size 4
 
 (4) fast beam search
-./pruned_transducer_stateless2/decode.py \
+./vq_pruned_transducer_stateless2/decode.py \
         --epoch 28 \
         --avg 15 \
-        --exp-dir ./pruned_transducer_stateless2/exp \
+        --exp-dir ./vq_pruned_transducer_stateless2/exp \
         --max-duration 1500 \
         --decoding-method fast_beam_search \
         --beam 4 \
@@ -124,7 +124,7 @@ def get_parser():
     parser.add_argument(
         "--exp-dir",
         type=str,
-        default="pruned_transducer_stateless2/exp",
+        default="vq_pruned_transducer_stateless2/exp",
         help="The experiment dir",
     )
 

egs/librispeech/ASR/vq_pruned_transducer_stateless2/train.py

@@ -19,27 +19,19 @@
 """
 Usage:
 
-export CUDA_VISIBLE_DEVICES="0,1,2,3"
+export CUDA_VISIBLE_DEVICES="0,1,2"
 
-./pruned_transducer_stateless2/train.py \
+./vq_pruned_transducer_stateless2/train.py \
-  --world-size 4 \
+  --world-size 3 \
   --num-epochs 30 \
   --start-epoch 0 \
-  --exp-dir pruned_transducer_stateless2/exp \
+  --manifest-dir data/quantizer-finetuned_hubert_xtralarge-36layer-f0238f68-bytes_per_frame-8/ \
-  --full-libri 1 \
+  --codebook-loss-scale 0.1 \
+  --num-codebooks=8 \
+  --exp-dir vq_pruned_transducer_stateless2/exp \
+  --full-libri 0 \
   --max-duration 300
 
-# For mix precision training:
-
-./pruned_transducer_stateless2/train.py \
-  --world-size 4 \
-  --num-epochs 30 \
-  --start-epoch 0 \
-  --use_fp16 1 \
-  --exp-dir pruned_transducer_stateless2/exp \
-  --full-libri 1 \
-  --max-duration 550
-
 """
 
 
@@ -60,9 +52,10 @@ from asr_datamodule import LibriSpeechAsrDataModule
 from conformer import Conformer
 from decoder import Decoder
 from joiner import Joiner
-from lhotse.cut import Cut
+from lhotse.cut import Cut, MonoCut
 from lhotse.dataset.sampling.base import CutSampler
 from lhotse.utils import fix_random_seed
+from lhotse.dataset.collation import collate_custom_field
 from model import Transducer
 from optim import Eden, Eve
 from torch import Tensor
@@ -138,7 +131,7 @@ def get_parser():
     parser.add_argument(
         "--exp-dir",
         type=str,
-        default="pruned_transducer_stateless2/exp",
+        default="vq_pruned_transducer_stateless2/exp",
         help="""The experiment dir.
         It specifies the directory where all training related
         files, e.g., checkpoints, log, etc, are saved
@@ -262,6 +255,23 @@ def get_parser():
         help="Whether to use half precision training.",
     )
 
+    parser.add_argument(
+        "--codebook-loss-scale",
+        type=float,
+        default=0.1,
+        help="""The weight of code book loss.
+            Note: Currently rate of ctc_loss +  rate of att_loss = 1.0
+            codebook_loss is independent with them.
+            """,
+    )
+
+    parser.add_argument(
+        "--num-codebooks",
+        type=int,
+        default=8,
+        help="number of code books",
+    )
+
     return parser
 
 
@@ -333,6 +343,8 @@ def get_params() -> AttributeDict:
             # parameters for Noam
             "model_warm_step": 3000,  # arg given to model, not for lrate
             "env_info": get_env_info(),
+            "extra_output_layer": 5,  # 0-based index
+            "num_codebooks": 8,
         }
     )
 
@@ -348,6 +360,7 @@ def get_encoder_model(params: AttributeDict) -> nn.Module:
         nhead=params.nhead,
         dim_feedforward=params.dim_feedforward,
         num_encoder_layers=params.num_encoder_layers,
+        extra_output_layer=params.extra_output_layer,
     )
     return encoder
 
@@ -385,6 +398,7 @@ def get_transducer_model(params: AttributeDict) -> nn.Module:
         decoder_dim=params.decoder_dim,
         joiner_dim=params.joiner_dim,
         vocab_size=params.vocab_size,
+        num_codebooks=params.num_codebooks,
     )
     return model
 
@@ -539,8 +553,20 @@ def compute_loss(
     y = sp.encode(texts, out_type=int)
     y = k2.RaggedTensor(y).to(device)
 
+    if is_training:
+        cuts = batch["supervisions"]["cut"]
+        # -100 is identical to ignore_value in CE loss computation.
+        cuts_pre_mixed = [
+            c if isinstance(c, MonoCut) else c.tracks[0].cut for c in cuts
+        ]
+        codebook_indices, codebook_indices_lens = collate_custom_field(
+            cuts_pre_mixed, "codebook_indices", pad_value=-100
+        )
+        codebook_indices = codebook_indices.to(device)
+    else:
+        codebook_indices = None
     with torch.set_grad_enabled(is_training):
-        simple_loss, pruned_loss = model(
+        simple_loss, pruned_loss, codebook_loss = model(
             x=feature,
             x_lens=feature_lens,
             y=y,
@@ -548,6 +574,7 @@ def compute_loss(
             am_scale=params.am_scale,
             lm_scale=params.lm_scale,
             warmup=warmup,
+            codebook_indices=codebook_indices,
         )
         # after the main warmup step, we keep pruned_loss_scale small
         # for the same amount of time (model_warm_step), to avoid
@@ -562,6 +589,9 @@ def compute_loss(
             params.simple_loss_scale * simple_loss
             + pruned_loss_scale * pruned_loss
         )
+        if is_training:
+            assert codebook_loss is not None
+            loss += params.codebook_loss_scale * codebook_loss
 
     assert loss.requires_grad == is_training
 
@@ -576,6 +606,8 @@ def compute_loss(
     info["loss"] = loss.detach().cpu().item()
     info["simple_loss"] = simple_loss.detach().cpu().item()
     info["pruned_loss"] = pruned_loss.detach().cpu().item()
+    if is_training:
+        info["codebook_loss"] = codebook_loss.detach().cpu().item()
 
     return loss, info
 
@@ -835,7 +867,7 @@ def run(rank, world_size, args):
 
     if params.print_diagnostics:
         opts = diagnostics.TensorDiagnosticOptions(
-            2 ** 22
+            2**22
         )  # allow 4 megabytes per sub-module
         diagnostic = diagnostics.attach_diagnostics(model, opts)