fix bugs

egs/librispeech/ASR/tdnn_lstm_ctc2/decode.py

@@ -125,7 +125,7 @@ def get_parser():
 def get_params() -> AttributeDict:
     params = AttributeDict(
         {
-            "exp_dir": Path("tdnn_lstm_ctc/exp/"),
+            "exp_dir": Path("tdnn_lstm_ctc2/exp/"),
             "lang_dir": Path("data/lang_phone"),
             "lm_dir": Path("data/lm"),
             "feature_dim": 80,
@@ -136,6 +136,11 @@ def get_params() -> AttributeDict:
             "max_active_states": 10000,
             "use_double_scores": True,
             "env_info": get_env_info(),
+            "grad_norm_threshold": 10.0,
+            # For each sequence element in batch, its gradient will be
+            # filtered out if the gradient norm is larger than
+            # `grad_norm_threshold * median`, where `median` is the median
+            # value of gradient norms of all elememts in batch.
         }
     )
     return params
@@ -452,6 +457,7 @@ def main():
         num_features=params.feature_dim,
         num_classes=max_phone_id + 1,  # +1 for the blank symbol
         subsampling_factor=params.subsampling_factor,
+        grad_norm_threshold=params.grad_norm_threshold,
     )
     if params.avg == 1:
         load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)

egs/librispeech/ASR/tdnn_lstm_ctc2/model.py

@@ -144,7 +144,10 @@ class TdnnLstm(nn.Module):
             [nn.BatchNorm1d(num_features=500, affine=False) for _ in range(5)]
         )
         self.grad_filters = nn.ModuleList(
-            [GradientFilter(batch_dim=1, threshold=grad_norm_threshold)]
+            [
+                GradientFilter(batch_dim=1, threshold=grad_norm_threshold)
+                for _ in range(5)
+            ]
         )
 
         self.dropout = nn.Dropout(0.2)

egs/librispeech/ASR/tdnn_lstm_ctc2/pretrained.py

@@ -124,6 +124,11 @@ def get_params() -> AttributeDict:
             "min_active_states": 30,
             "max_active_states": 10000,
             "use_double_scores": True,
+            "grad_norm_threshold": 10.0,
+            # For each sequence element in batch, its gradient will be
+            # filtered out if the gradient norm is larger than
+            # `grad_norm_threshold * median`, where `median` is the median
+            # value of gradient norms of all elememts in batch.
         }
     )
     return params
@@ -172,6 +177,7 @@ def main():
         num_features=params.feature_dim,
         num_classes=params.num_classes,
         subsampling_factor=params.subsampling_factor,
+        grad_norm_threshold=params.grad_norm_threshold,
     )
 
     checkpoint = torch.load(args.checkpoint, map_location="cpu")

egs/librispeech/ASR/tdnn_lstm_ctc2/train.py

@@ -19,7 +19,7 @@
 """
 Usage:
   export CUDA_VISIBLE_DEVICES="0,1,2,3"
-  ./tdnn_lstm_ctc/train.py \
+  ./tdnn_lstm_ctc2/train.py \
      --world-size 4 \
      --full-libri 1 \
      --max-duration 300 \
@@ -112,16 +112,6 @@ def get_parser():
         help="The seed for random generators intended for reproducibility",
     )
 
-    parser.add_argument(
-        "--grad-norm-threshold",
-        type=float,
-        default=10.0,
-        help="""For each sequence element in batch, its gradient will be
-        filtered out if the gradient norm is larger than
-        `grad_norm_threshold * median`, where `median` is the median
-        value of gradient norms of all elememts in batch.""",
-    )
-
     return parser
 
 
@@ -199,6 +189,11 @@ def get_params() -> AttributeDict:
             "reduction": "sum",
             "use_double_scores": True,
             "env_info": get_env_info(),
+            "grad_norm_threshold": 10.0,
+            # For each sequence element in batch, its gradient will be
+            # filtered out if the gradient norm is larger than
+            # `grad_norm_threshold * median`, where `median` is the median
+            # value of gradient norms of all elememts in batch.
         }
     )