update examples and model summary print

nglehuy · nglehuy · commit 65acbf71347d · 2022-03-12T14:28:22.000+07:00
diff --git a/examples/conformer/README.md b/examples/conformer/README.md
@@ -10,11 +10,11 @@ Go to [config.yml](./config.yml)
 
 ## Usage
 
-Training, see `python examples/conformer/train_*.py --help`
+Training, see `python examples/conformer/train.py --help`
 
-Testing, see `python examples/conformer/test_*.py --help`
+Testing, see `python examples/conformer/test.py --help`
 
-TFLite Conversion, see `python examples/conformer/tflite_*.py --help`
+TFLite Conversion, see `python examples/conformer/inference/gen_tflite_model.py --help`
 
 ## Conformer Subwords - Results on LibriSpeech
 
diff --git a/examples/conformer/config.yml b/examples/conformer/config.yml
@@ -31,7 +31,7 @@ decoder_config:
   beam_width: 0
   norm_score: True
   corpus_files:
-    - /h/MLDL/Datasets/ASR/Raw/LibriSpeech/train-clean-100/transcripts.tsv
+    - H:/MLDL/Datasets/ASR/Raw/LibriSpeech/train-clean-100/transcripts.tsv
 
 model_config:
   name: conformer
@@ -75,7 +75,7 @@ learning_config:
           num_masks: 1
           mask_factor: 27
     data_paths:
-      - /h/MLDL/Datasets/ASR/Raw/LibriSpeech/train-clean-100/transcripts.tsv
+      - H:/MLDL/Datasets/ASR/Raw/LibriSpeech/train-clean-100/transcripts.tsv
     tfrecords_dir: null
     shuffle: True
     cache: True
@@ -85,7 +85,8 @@ learning_config:
 
   eval_dataset_config:
     use_tf: True
-    data_paths: null
+    data_paths:
+      - H:/MLDL/Datasets/ASR/Raw/LibriSpeech/dev-clean/transcripts.tsv
     tfrecords_dir: null
     shuffle: False
     cache: True
@@ -95,7 +96,8 @@ learning_config:
 
   test_dataset_config:
     use_tf: True
-    data_paths: null
+    data_paths:
+      - H:/MLDL/Datasets/ASR/Raw/LibriSpeech/test-clean/transcripts.tsv
     tfrecords_dir: null
     shuffle: False
     cache: True
@@ -113,13 +115,13 @@ learning_config:
     batch_size: 2
     num_epochs: 50
     checkpoint:
-      filepath: /d/Models/local/conformer/checkpoints/{epoch:02d}.h5
+      filepath: D:/Models/local/conformer/checkpoints/{epoch:02d}.h5
       save_best_only: False
       save_weights_only: True
       save_freq: epoch
-    states_dir: /d/Models/local/conformer/states
+    states_dir: D:/Models/local/conformer/states
     tensorboard:
-      log_dir: /d/Models/local/conformer/tensorboard
+      log_dir: D:/Models/local/conformer/tensorboard
       histogram_freq: 1
       write_graph: True
       write_images: True
diff --git a/examples/conformer/train.py b/examples/conformer/train.py
@@ -75,7 +75,6 @@ def main(
     )
 
     with strategy.scope():
-        # build model
         conformer = Conformer(**config.model_config, vocabulary_size=text_featurizer.num_classes)
         conformer.make(speech_featurizer.shape, prediction_shape=text_featurizer.prepand_shape, batch_size=global_batch_size)
         if pretrained:
diff --git a/examples/contextnet/README.md b/examples/contextnet/README.md
@@ -12,11 +12,11 @@ Go to [config.yml](./config.yml)
 
 ## Usage
 
-Training, see `python examples/contextnet/train_*.py --help`
+Training, see `python examples/contextnet/train.py --help`
 
-Testing, see `python examples/contextnet/test_*.py --help`
+Testing, see `python examples/contextnet/test.py --help`
 
-TFLite Conversion, see `python examples/contextnet/tflite_*.py --help`
+TFLite Conversion, see `python examples/contextnet/inference/gen_tflite_model.py --help`
 
 ## RNN Transducer Subwords - Results on LibriSpeech
 
diff --git a/examples/contextnet/config.yml b/examples/contextnet/config.yml
@@ -24,7 +24,7 @@ speech_config:
   normalize_per_frame: False
 
 decoder_config:
-  vocabulary: null
+  vocabulary: ./vocabularies/librispeech/librispeech_train_4_1030.subwords
   target_vocab_size: 1024
   max_subword_length: 4
   blank_at_zero: True
@@ -207,7 +207,7 @@ learning_config:
           num_masks: 1
           mask_factor: 27
     data_paths:
-      - /mnt/h/ML/Datasets/ASR/Raw/LibriSpeech/train-clean-100/transcripts.tsv
+      - H:/MLDL/Datasets/ASR/Raw/LibriSpeech/train-clean-100/transcripts.tsv
     tfrecords_dir: null
     shuffle: True
     cache: True
@@ -217,7 +217,8 @@ learning_config:
 
   eval_dataset_config:
     use_tf: True
-    data_paths: null
+    data_paths:
+      - H:/MLDL/Datasets/ASR/Raw/LibriSpeech/dev-clean/transcripts.tsv
     tfrecords_dir: null
     shuffle: False
     cache: True
@@ -228,7 +229,7 @@ learning_config:
   test_dataset_config:
     use_tf: True
     data_paths:
-      - /mnt/h/ML/Datasets/ASR/Raw/LibriSpeech/test-clean/transcripts.tsv
+      - H:/MLDL/Datasets/ASR/Raw/LibriSpeech/test-clean/transcripts.tsv
     tfrecords_dir: null
     shuffle: False
     cache: True
@@ -246,13 +247,13 @@ learning_config:
     batch_size: 2
     num_epochs: 20
     checkpoint:
-      filepath: /mnt/e/Models/local/contextnet/checkpoints/{epoch:02d}.h5
+      filepath: D:/Models/local/contextnet/checkpoints/{epoch:02d}.h5
       save_best_only: False
       save_weights_only: True
       save_freq: epoch
-    states_dir: /mnt/e/Models/local/contextnet/states
+    states_dir: D:/Models/local/contextnet/states
     tensorboard:
-      log_dir: /mnt/e/Models/local/contextnet/tensorboard
+      log_dir: D:/Models/local/contextnet/tensorboard
       histogram_freq: 1
       write_graph: True
       write_images: True
diff --git a/examples/contextnet/train.py b/examples/contextnet/train.py
@@ -75,7 +75,6 @@ def main(
     )
 
     with strategy.scope():
-        # build model
         contextnet = ContextNet(**config.model_config, vocabulary_size=text_featurizer.num_classes)
         contextnet.make(speech_featurizer.shape, prediction_shape=text_featurizer.prepand_shape, batch_size=global_batch_size)
         if pretrained:
diff --git a/examples/deepspeech2/config.yml b/examples/deepspeech2/config.yml
@@ -52,7 +52,7 @@ learning_config:
   train_dataset_config:
     use_tf: True
     data_paths:
-      - /mnt/h/ML/Datasets/ASR/Raw/LibriSpeech/train-clean-100/transcripts.tsv
+      - H:/MLDL/Datasets/ASR/Raw/LibriSpeech/train-clean-100/transcripts.tsv
     tfrecords_dir: null
     shuffle: True
     cache: True
@@ -62,7 +62,8 @@ learning_config:
 
   eval_dataset_config:
     use_tf: True
-    data_paths: null
+    data_paths:
+      - H:/MLDL/Datasets/ASR/Raw/LibriSpeech/dev-clean/transcripts.tsv
     tfrecords_dir: null
     shuffle: False
     cache: True
@@ -73,7 +74,7 @@ learning_config:
   test_dataset_config:
     use_tf: True
     data_paths:
-      - /mnt/h/ML/Datasets/ASR/Raw/LibriSpeech/test-clean/transcripts.tsv
+      - H:/MLDL/Datasets/ASR/Raw/LibriSpeech/test-clean/transcripts.tsv
     tfrecords_dir: null
     shuffle: False
     cache: True
@@ -90,13 +91,13 @@ learning_config:
     batch_size: 4
     num_epochs: 20
     checkpoint:
-      filepath: /mnt/e/Models/local/deepspeech2/checkpoints/{epoch:02d}.h5
+      filepath: D:/Models/local/deepspeech2/checkpoints/{epoch:02d}.h5
       save_best_only: False
       save_weights_only: True
       save_freq: epoch
-    states_dir: /mnt/e/Models/local/deepspeech2/states
+    states_dir: D:/Models/local/deepspeech2/states
     tensorboard:
-      log_dir: /mnt/e/Models/local/deepspeech2/tensorboard
+      log_dir: D:/Models/local/deepspeech2/tensorboard
       histogram_freq: 1
       write_graph: True
       write_images: True
diff --git a/examples/deepspeech2/train.py b/examples/deepspeech2/train.py
@@ -31,7 +31,7 @@ def main(
     config: str = DEFAULT_YAML,
     tfrecords: bool = False,
     sentence_piece: bool = False,
-    subwords: bool = True,
+    subwords: bool = False,
     bs: int = None,
     spx: int = 1,
     metadata: str = None,
diff --git a/examples/jasper/config.yml b/examples/jasper/config.yml
@@ -59,7 +59,7 @@ learning_config:
   train_dataset_config:
     use_tf: True
     data_paths:
-      - /mnt/h/ML/Datasets/ASR/Raw/LibriSpeech/train-clean-100/transcripts.tsv
+      - H:/MLDL/Datasets/ASR/Raw/LibriSpeech/train-clean-100/transcripts.tsv
     tfrecords_dir: null
     shuffle: True
     cache: True
@@ -69,7 +69,8 @@ learning_config:
 
   eval_dataset_config:
     use_tf: True
-    data_paths: null
+    data_paths:
+      - H:/MLDL/Datasets/ASR/Raw/LibriSpeech/dev-clean/transcripts.tsv
     tfrecords_dir: null
     shuffle: False
     cache: True
@@ -80,7 +81,7 @@ learning_config:
   test_dataset_config:
     use_tf: True
     data_paths:
-      - /mnt/h/ML/Datasets/ASR/Raw/LibriSpeech/test-clean/transcripts.tsv
+      - H:/MLDL/Datasets/ASR/Raw/LibriSpeech/test-clean/transcripts.tsv
     tfrecords_dir: null
     shuffle: False
     cache: True
@@ -97,13 +98,13 @@ learning_config:
     batch_size: 4
     num_epochs: 20
     checkpoint:
-      filepath: /mnt/e/Models/local/jasper/checkpoints/{epoch:02d}.h5
+      filepath: D:/Models/local/jasper/checkpoints/{epoch:02d}.h5
       save_best_only: False
       save_weights_only: True
       save_freq: epoch
-    states_dir: /mnt/e/Models/local/jasper/states
+    states_dir: D:/Models/local/jasper/states
     tensorboard:
-      log_dir: /mnt/e/Models/local/jasper/tensorboard
+      log_dir: D:/Models/local/jasper/tensorboard
       histogram_freq: 1
       write_graph: True
       write_images: True
diff --git a/examples/jasper/train.py b/examples/jasper/train.py
@@ -30,7 +30,7 @@ def main(
     config: str = DEFAULT_YAML,
     tfrecords: bool = False,
     sentence_piece: bool = False,
-    subwords: bool = True,
+    subwords: bool = False,
     bs: int = None,
     spx: int = 1,
     metadata: str = None,
diff --git a/examples/rnn_transducer/README.md b/examples/rnn_transducer/README.md
@@ -8,11 +8,11 @@ Go to [config.yml](./config.yml)
 
 ## Usage
 
-Training, see `python examples/streamingTransducer/train_*.py --help`
+Training, see `python examples/rnn_transducer/train.py --help`
 
-Testing, see `python examples/streamingTransducer/test_*.py --help`
+Testing, see `python examples/rnn_transducer/test.py --help`
 
-TFLite Conversion, see `python examples/streamingTransducer/tflite_*.py --help`
+TFLite Conversion, see `python examples/rnn_transducer/inference/gen_tflite_model.py --help`
 
 ## RNN Transducer Subwords - Results on LibriSpeech
 
diff --git a/examples/rnn_transducer/config.yml b/examples/rnn_transducer/config.yml
@@ -26,15 +26,15 @@ speech_config:
   top_db: null
 
 decoder_config:
-  vocabulary: null
+  vocabulary: ./vocabularies/librispeech/librispeech_train_4_1030.subwords
   target_vocab_size: 1024
   max_subword_length: 4
   blank_at_zero: True
   beam_width: 5
   norm_score: True
 
 model_config:
-  name: streaming_transducer
+  name: rnn_transducer
   encoder_reductions:
     0: 3
     1: 2
@@ -43,7 +43,6 @@ model_config:
   encoder_rnn_units: 1024
   encoder_nlayers: 8
   encoder_layer_norm: True
-  encoder_feature_norm: True
   prediction_embed_dim: 320
   prediction_embed_dropout: 0.0
   prediction_num_rnns: 2
@@ -67,7 +66,7 @@ learning_config:
           num_masks: 1
           mask_factor: 27
     data_paths:
-      - /mnt/h/ML/Datasets/ASR/Raw/LibriSpeech/train-clean-100/transcripts.tsv
+      - H:/MLDL/Datasets/ASR/Raw/LibriSpeech/train-clean-100/transcripts.tsv
     tfrecords_dir: null
     shuffle: True
     cache: True
@@ -77,7 +76,8 @@ learning_config:
 
   eval_dataset_config:
     use_tf: True
-    data_paths: null
+    data_paths:
+      - H:/MLDL/Datasets/ASR/Raw/LibriSpeech/dev-clean/transcripts.tsv
     tfrecords_dir: null
     shuffle: False
     cache: True
@@ -88,7 +88,7 @@ learning_config:
   test_dataset_config:
     use_tf: True
     data_paths:
-      - /mnt/h/ML/Datasets/ASR/Raw/LibriSpeech/test-clean/transcripts.tsv
+      - H:/MLDL/Datasets/ASR/Raw/LibriSpeech/test-clean/transcripts.tsv
     tfrecords_dir: null
     shuffle: False
     cache: True
@@ -105,13 +105,13 @@ learning_config:
     batch_size: 2
     num_epochs: 20
     checkpoint:
-      filepath: /mnt/e/Models/local/rnn_transducer/checkpoints/{epoch:02d}.h5
+      filepath: D:/Models/local/rnn_transducer/checkpoints/{epoch:02d}.h5
       save_best_only: False
       save_weights_only: True
       save_freq: epoch
-    states_dir: /mnt/e/Models/local/rnn_transducer/states
+    states_dir: D:/Models/local/rnn_transducer/states
     tensorboard:
-      log_dir: /mnt/e/Models/local/rnn_transducer/tensorboard
+      log_dir: D:/Models/local/rnn_transducer/tensorboard
       histogram_freq: 1
       write_graph: True
       write_images: True
diff --git a/examples/rnn_transducer/train.py b/examples/rnn_transducer/train.py
@@ -13,7 +13,6 @@
 # limitations under the License.
 
 import os
-import math
 import fire
 from tensorflow_asr.utils import env_util
 
@@ -24,7 +23,6 @@
 from tensorflow_asr.configs.config import Config
 from tensorflow_asr.helpers import featurizer_helpers, dataset_helpers
 from tensorflow_asr.models.transducer.rnn_transducer import RnnTransducer
-from tensorflow_asr.optimizers.schedules import TransformerSchedule
 
 DEFAULT_YAML = os.path.join(os.path.abspath(os.path.dirname(__file__)), "config.yml")
 
@@ -75,24 +73,15 @@ def main(
     )
 
     with strategy.scope():
-        # build model
         rnn_transducer = RnnTransducer(**config.model_config, vocabulary_size=text_featurizer.num_classes)
         rnn_transducer.make(
             speech_featurizer.shape, prediction_shape=text_featurizer.prepand_shape, batch_size=global_batch_size
         )
         if pretrained:
             rnn_transducer.load_weights(pretrained, by_name=True, skip_mismatch=True)
         rnn_transducer.summary(line_length=100)
-        optimizer = tf.keras.optimizers.Adam(
-            TransformerSchedule(
-                d_model=rnn_transducer.dmodel,
-                warmup_steps=config.learning_config.optimizer_config.pop("warmup_steps", 10000),
-                max_lr=(0.05 / math.sqrt(rnn_transducer.dmodel)),
-            ),
-            **config.learning_config.optimizer_config
-        )
         rnn_transducer.compile(
-            optimizer=optimizer,
+            optimizer=config.learning_config.optimizer_config,
             experimental_steps_per_execution=spx,
             global_batch_size=global_batch_size,
             blank=text_featurizer.blank,
diff --git a/setup.py b/setup.py
@@ -41,7 +41,7 @@ def parse_requirements(lines: List[str]):
 
 setup(
     name="TensorFlowASR",
-    version="1.0.2",
+    version="1.0.3",
     author="Huy Le Nguyen",
     author_email="nlhuy.cs.16@gmail.com",
     description="Almost State-of-the-art Automatic Speech Recognition using Tensorflow 2",
diff --git a/tensorflow_asr/models/ctc/base_ctc.py b/tensorflow_asr/models/ctc/base_ctc.py
diff --git a/tensorflow_asr/models/transducer/base_transducer.py b/tensorflow_asr/models/transducer/base_transducer.py

Original file line number	Diff line number	Diff line change
`@@ -75,7 +75,6 @@ def main(`
`75`	`75`	`)`
`76`	`76`
`77`	`77`	`with strategy.scope():`
`78`		`- # build model`
`79`	`78`	`conformer = Conformer(**config.model_config, vocabulary_size=text_featurizer.num_classes)`
`80`	`79`	`conformer.make(speech_featurizer.shape, prediction_shape=text_featurizer.prepand_shape, batch_size=global_batch_size)`
`81`	`80`	`if pretrained:`