Merge pull request #151 from TensorSpeech/dev/gcs

Support for save model to cloud and infinite dataset

README.md

@@ -44,7 +44,7 @@ TensorFlowASR implements some automatic speech recognition architectures such as
 - [TFLite Convertion](#tflite-convertion)
 - [Features Extraction](#features-extraction)
 - [Augmentations](#augmentations)
-- [Training & Testing](#training--testing)
+- [Training & Testing Tutorial](#training--testing-tutorial)
 - [Corpus Sources and Pretrained Models](#corpus-sources-and-pretrained-models)
   - [English](#english)
   - [Vietnamese](#vietnamese)
@@ -164,34 +164,17 @@ See [features_extraction](./tensorflow_asr/featurizers/README.md)
 
 See [augmentations](./tensorflow_asr/augmentations/README.md)
 
-## Training & Testing
-
-**Example YAML Config Structure**
-
-```yaml
-speech_config: ...
-model_config: ...
-decoder_config: ...
-learning_config:
-  train_dataset_config:
-    augmentation_config: ...
-    data_paths: ...
-    tfrecords_dir: ...
-  eval_dataset_config:
-    augmentation_config: ...
-    data_paths: ...
-    tfrecords_dir: ...
-  test_dataset_config:
-    augmentation_config: ...
-    data_paths: ...
-    tfrecords_dir: ...
-  optimizer_config: ...
-  running_config:
-    batch_size: 8
-    num_epochs: 20
-    outdir: ...
-    log_interval_steps: 500
-```
+## Training & Testing Tutorial
+
+1. Define config YAML file, see the `config.yml` files in the [example folder](./examples) for reference (you can copy and modify values such as parameters, paths, etc.. to match your local machine configuration)
+2. Download your corpus (a.k.a datasets) and create a script to generate `transcripts.tsv` files from your corpus (this is general format used in this project because each dataset has different format). For more detail, see [datasets](./tensorflow_asr/datasets/README.md). **Note:** Make sure your data contain only characters in your language, for example, english has `a` to `z` and `'`. **Do not use `cache` if your dataset size is not fit in the RAM**.
+3. [Optional] Generate TFRecords to use `tf.data.TFRecordDataset` for better performance by using the script [create_tfrecords.py](./scripts/create_tfrecords.py)
+4. Create vocabulary file (characters or subwords/wordpieces) by defining `language.characters`, using the scripts [generate_vocab_subwords.py](./scripts/generate_vocab_subwords.py) or [generate_vocab_sentencepiece.py](./scripts/generate_vocab_sentencepiece.py). There're predefined ones in [vocabularies](./vocabularies)
+5. [Optional] Generate metadata file for your dataset by using script [generate_metadata.py](./scripts/generate_metadata.py). This metadata file contains maximum lengths calculated with your `config.yml` and total number of elements in each dataset, for static shape training and precalculated steps per epoch.
+6. For training, see `train_*.py` files in the [example folder](./examples) to see the options
+7. For testing, see `test_.*.py` files in the [example folder](./examples) to see the options. **Note:** Testing is currently not supported for TPUs. It will print nothing other than the progress bar in the console, but it will store the predicted transcripts to the file `output_name.tsv` in the `outdir` defined in the config yaml file. After testing is done, the metrics (WER and CER) are calculated from `output_name.tsv`. **If you define the same `output_name`, it will resume the testing from the previous tested batch, which means if the testing is done then it will only calculate the metrics, if you want to run a new test, define a new `output_name` that the file `output.tsv` is not exists or only contains the header**
+
+**Recommendation**: For better performance, please use **keras builtin training functions** as in `train_keras_*.py` files and/or tfrecords. Keras builtin training uses **infinite dataset**, which avoids the potential last partial batch.
 
 See [examples](./examples/) for some predefined ASR models and results
 

examples/conformer/train_keras_subword_conformer.py

@@ -38,6 +38,10 @@
 
 parser.add_argument("--ebs", type=int, default=None, help="Evaluation batch size per replica")
 
+parser.add_argument("--spx", type=int, default=1, help="Steps per execution for maximizing performance")
+
+parser.add_argument("--metadata_prefix", type=str, default=None, help="Path to file containing metadata")
+
 parser.add_argument("--devices", type=int, nargs="*", default=[0], help="Devices' ids to apply distributed training")
 
 parser.add_argument("--mxp", default=False, action="store_true", help="Enable mixed precision")
@@ -79,25 +83,38 @@
 if args.tfrecords:
     train_dataset = ASRTFRecordDatasetKeras(
         speech_featurizer=speech_featurizer, text_featurizer=text_featurizer,
-        **vars(config.learning_config.train_dataset_config)
+        **vars(config.learning_config.train_dataset_config),
+        indefinite=True
     )
     eval_dataset = ASRTFRecordDatasetKeras(
         speech_featurizer=speech_featurizer, text_featurizer=text_featurizer,
         **vars(config.learning_config.eval_dataset_config)
     )
+    # Update metadata calculated from both train and eval datasets
+    train_dataset.load_metadata(args.metadata_prefix)
+    eval_dataset.load_metadata(args.metadata_prefix)
+    # Use dynamic length
+    speech_featurizer.reset_length()
+    text_featurizer.reset_length()
 else:
     train_dataset = ASRSliceDatasetKeras(
         speech_featurizer=speech_featurizer, text_featurizer=text_featurizer,
-        **vars(config.learning_config.train_dataset_config)
+        **vars(config.learning_config.train_dataset_config),
+        indefinite=True
     )
     eval_dataset = ASRSliceDatasetKeras(
         speech_featurizer=speech_featurizer, text_featurizer=text_featurizer,
-        **vars(config.learning_config.train_dataset_config)
+        **vars(config.learning_config.train_dataset_config),
+        indefinite=True
     )
 
+global_batch_size = config.learning_config.running_config.batch_size
+global_batch_size *= strategy.num_replicas_in_sync
+
+train_data_loader = train_dataset.create(global_batch_size)
+eval_data_loader = eval_dataset.create(global_batch_size)
+
 with strategy.scope():
-    global_batch_size = config.learning_config.running_config.batch_size
-    global_batch_size *= strategy.num_replicas_in_sync
     # build model
     conformer = Conformer(**config.model_config, vocabulary_size=text_featurizer.num_classes)
     conformer._build(speech_featurizer.shape)
@@ -114,19 +131,21 @@
         epsilon=config.learning_config.optimizer_config["epsilon"]
     )
 
-    conformer.compile(optimizer=optimizer, global_batch_size=global_batch_size, blank=text_featurizer.blank)
-
-    train_data_loader = train_dataset.create(global_batch_size)
-    eval_data_loader = eval_dataset.create(global_batch_size)
-
-    callbacks = [
-        tf.keras.callbacks.ModelCheckpoint(**config.learning_config.running_config.checkpoint),
-        tf.keras.callbacks.experimental.BackupAndRestore(config.learning_config.running_config.states_dir),
-        tf.keras.callbacks.TensorBoard(**config.learning_config.running_config.tensorboard)
-    ]
-
-    conformer.fit(
-        train_data_loader, epochs=config.learning_config.running_config.num_epochs,
-        validation_data=eval_data_loader, callbacks=callbacks,
-        steps_per_epoch=train_dataset.total_steps
+    conformer.compile(
+        optimizer=optimizer,
+        experimental_steps_per_execution=args.spx,
+        global_batch_size=global_batch_size,
+        blank=text_featurizer.blank
     )
+
+callbacks = [
+    tf.keras.callbacks.ModelCheckpoint(**config.learning_config.running_config.checkpoint),
+    tf.keras.callbacks.experimental.BackupAndRestore(config.learning_config.running_config.states_dir),
+    tf.keras.callbacks.TensorBoard(**config.learning_config.running_config.tensorboard)
+]
+
+conformer.fit(
+    train_data_loader, epochs=config.learning_config.running_config.num_epochs,
+    validation_data=eval_data_loader, callbacks=callbacks,
+    steps_per_epoch=train_dataset.total_steps, validation_steps=eval_dataset.total_steps
+)

examples/conformer/train_tpu_keras_subword_conformer.py

@@ -32,7 +32,7 @@
 
 parser.add_argument("--bs", type=int, default=None, help="Batch size per replica")
 
-parser.add_argument("--spx", type=int, default=1, help="Steps per execution for maximizing TPU performance")
+parser.add_argument("--spx", type=int, default=50, help="Steps per execution for maximizing TPU performance")
 
 parser.add_argument("--tpu_address", type=str, default=None, help="TPU address. Leave None on Colab")
 
@@ -78,11 +78,13 @@
 
 train_dataset = ASRTFRecordDatasetKeras(
     speech_featurizer=speech_featurizer, text_featurizer=text_featurizer,
-    **vars(config.learning_config.train_dataset_config)
+    **vars(config.learning_config.train_dataset_config),
+    indefinite=True
 )
 eval_dataset = ASRTFRecordDatasetKeras(
     speech_featurizer=speech_featurizer, text_featurizer=text_featurizer,
-    **vars(config.learning_config.eval_dataset_config)
+    **vars(config.learning_config.eval_dataset_config),
+    indefinite=True
 )
 
 if args.compute_lengths:
@@ -93,10 +95,14 @@
 train_dataset.load_metadata(args.metadata_prefix)
 eval_dataset.load_metadata(args.metadata_prefix)
 
+batch_size = args.bs if args.bs is not None else config.learning_config.running_config.batch_size
+global_batch_size = batch_size
+global_batch_size *= strategy.num_replicas_in_sync
+
+train_data_loader = train_dataset.create(global_batch_size)
+eval_data_loader = eval_dataset.create(global_batch_size)
+
 with strategy.scope():
-    batch_size = args.bs if args.bs is not None else config.learning_config.running_config.batch_size
-    global_batch_size = batch_size
-    global_batch_size *= strategy.num_replicas_in_sync
     # build model
     conformer = Conformer(**config.model_config, vocabulary_size=text_featurizer.num_classes)
     conformer._build(speech_featurizer.shape, prediction_shape=text_featurizer.prepand_shape, batch_size=global_batch_size)
@@ -120,17 +126,14 @@
         blank=text_featurizer.blank
     )
 
-    train_data_loader = train_dataset.create(global_batch_size)
-    eval_data_loader = eval_dataset.create(global_batch_size)
-
-    callbacks = [
-        tf.keras.callbacks.ModelCheckpoint(**config.learning_config.running_config.checkpoint),
-        tf.keras.callbacks.experimental.BackupAndRestore(config.learning_config.running_config.states_dir),
-        tf.keras.callbacks.TensorBoard(**config.learning_config.running_config.tensorboard)
-    ]
+callbacks = [
+    tf.keras.callbacks.ModelCheckpoint(**config.learning_config.running_config.checkpoint),
+    tf.keras.callbacks.experimental.BackupAndRestore(config.learning_config.running_config.states_dir),
+    tf.keras.callbacks.TensorBoard(**config.learning_config.running_config.tensorboard)
+]
 
-    conformer.fit(
-        train_data_loader, epochs=config.learning_config.running_config.num_epochs,
-        validation_data=eval_data_loader, callbacks=callbacks,
-        steps_per_epoch=train_dataset.total_steps, validation_steps=eval_dataset.total_steps
-    )
+conformer.fit(
+    train_data_loader, epochs=config.learning_config.running_config.num_epochs,
+    validation_data=eval_data_loader, callbacks=callbacks,
+    steps_per_epoch=train_dataset.total_steps, validation_steps=eval_dataset.total_steps
+)