Disable CUDA_LAUNCH_BLOCKING in wenetspeech recipes. (#554)

* Disable CUDA_LAUNCH_BLOCKING in wenetspeech recipes.

* minor fixes

Author: csukuangfj

egs/wenetspeech/ASR/local/preprocess_wenetspeech.py

@@ -23,6 +23,8 @@
 from lhotse import CutSet, SupervisionSegment
 from lhotse.recipes.utils import read_manifests_if_cached
 
+from icefall import setup_logger
+
 # Similar text filtering and normalization procedure as in:
 # https://github.com/SpeechColab/WenetSpeech/blob/main/toolkits/kaldi/wenetspeech_data_prep.sh
 
@@ -48,13 +50,17 @@ def preprocess_wenet_speech():
     output_dir = Path("data/fbank")
     output_dir.mkdir(exist_ok=True)
 
+    # Note: By default, we preprocess all sub-parts.
+    # You can delete those that you don't need.
+    # For instance, if you don't want to use the L subpart, just remove
+    # the line below containing "L"
     dataset_parts = (
-        "L",
-        "M",
-        "S",
         "DEV",
         "TEST_NET",
         "TEST_MEETING",
+        "S",
+        "M",
+        "L",
     )
 
     logging.info("Loading manifest (may take 10 minutes)")
@@ -81,10 +87,13 @@ def preprocess_wenet_speech():
         logging.info(f"Normalizing text in {partition}")
         for sup in m["supervisions"]:
             text = str(sup.text)
-            logging.info(f"Original text: {text}")
+            orig_text = text
             sup.text = normalize_text(sup.text)
             text = str(sup.text)
-            logging.info(f"Normalize text: {text}")
+            if len(orig_text) != len(text):
+                logging.info(
+                    f"\nOriginal text vs normalized text:\n{orig_text}\n{text}"
+                )
 
         # Create long-recording cut manifests.
         logging.info(f"Processing {partition}")
@@ -109,12 +118,10 @@ def preprocess_wenet_speech():
 
 
 def main():
-    formatter = (
-        "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
-    )
-    logging.basicConfig(format=formatter, level=logging.INFO)
+    setup_logger(log_filename="./log-preprocess-wenetspeech")
 
     preprocess_wenet_speech()
+    logging.info("Done")
 
 
 if __name__ == "__main__":

egs/wenetspeech/ASR/pruned_transducer_stateless2/train.py

@@ -81,7 +81,6 @@
 
 import argparse
 import logging
-import os
 import warnings
 from pathlib import Path
 from shutil import copyfile
@@ -120,8 +119,6 @@
     torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler
 ]
 
-os.environ["CUDA_LAUNCH_BLOCKING"] = "1"
-
 
 def get_parser():
     parser = argparse.ArgumentParser(
@@ -162,7 +159,7 @@ def get_parser():
         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
+        pruned_transducer_stateless2/exp/epoch-{start_epoch-1}.pt
         """,
     )
 
@@ -361,8 +358,8 @@ def get_params() -> AttributeDict:
             "best_valid_loss": float("inf"),
             "best_train_epoch": -1,
             "best_valid_epoch": -1,
-            "batch_idx_train": 10,
-            "log_interval": 1,
+            "batch_idx_train": 0,
+            "log_interval": 50,
             "reset_interval": 200,
             # parameters for conformer
             "feature_dim": 80,
@@ -545,7 +542,7 @@ def compute_loss(
     warmup: float = 1.0,
 ) -> Tuple[Tensor, MetricsTracker]:
     """
-    Compute CTC loss given the model and its inputs.
+    Compute RNN-T loss given the model and its inputs.
     Args:
       params:
         Parameters for training. See :func:`get_params`.
@@ -573,7 +570,7 @@ def compute_loss(
     texts = batch["supervisions"]["text"]
 
     y = graph_compiler.texts_to_ids(texts)
-    if type(y) == list:
+    if isinstance(y, list):
         y = k2.RaggedTensor(y).to(device)
     else:
         y = y.to(device)
@@ -697,7 +694,6 @@ def train_one_epoch(
     tot_loss = MetricsTracker()
 
     for batch_idx, batch in enumerate(train_dl):
-
         params.batch_idx_train += 1
         batch_size = len(batch["supervisions"]["text"])
 

egs/wenetspeech/ASR/pruned_transducer_stateless5/train.py

@@ -61,7 +61,6 @@
 import argparse
 import copy
 import logging
-import os
 import warnings
 from pathlib import Path
 from shutil import copyfile
@@ -103,8 +102,6 @@
     torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler
 ]
 
-os.environ["CUDA_LAUNCH_BLOCKING"] = "1"
-
 
 def add_model_arguments(parser: argparse.ArgumentParser):
     parser.add_argument(
@@ -684,7 +681,7 @@ def compute_loss(
     texts = batch["supervisions"]["text"]
 
     y = graph_compiler.texts_to_ids(texts)
-    if type(y) == list:
+    if isinstance(y, list):
         y = k2.RaggedTensor(y).to(device)
     else:
         y = y.to(device)