def main(
config: str = DEFAULT_YAML,
h5: str = None,
subwords: bool = False,
sentence_piece: bool = False,
output: str = None,
):
assert h5 and output
tf.keras.backend.clear_session()
tf.compat.v1.enable_control_flow_v2()
config = Config(config)
speech_featurizer, text_featurizer = featurizer_helpers.prepare_featurizers(
config=config,
subwords=subwords,
sentence_piece=sentence_piece,
)
contextnet = ContextNet(**config.model_config,
vocabulary_size=text_featurizer.num_classes)
contextnet.make(speech_featurizer.shape)
contextnet.load_weights(h5, by_name=True)
contextnet.summary(line_length=100)
contextnet.add_featurizers(speech_featurizer, text_featurizer)
exec_helpers.convert_tflite(model=contextnet, output=output)
def main(
config: str = DEFAULT_YAML,
saved: str = None,
mxp: bool = False,
bs: int = None,
sentence_piece: bool = False,
subwords: bool = False,
device: int = 0,
cpu: bool = False,
output: str = "test.tsv",
):
assert saved and output
tf.random.set_seed(0)
tf.keras.backend.clear_session()
tf.config.optimizer.set_experimental_options({"auto_mixed_precision": mxp})
env_util.setup_devices([device], cpu=cpu)
config = Config(config)
speech_featurizer, text_featurizer = featurizer_helpers.prepare_featurizers(
config=config,
subwords=subwords,
sentence_piece=sentence_piece,
)
contextnet = ContextNet(**config.model_config,
vocabulary_size=text_featurizer.num_classes)
contextnet.make(speech_featurizer.shape)
contextnet.load_weights(saved, by_name=True)
contextnet.summary(line_length=100)
contextnet.add_featurizers(speech_featurizer, text_featurizer)
test_dataset = dataset_helpers.prepare_testing_datasets(
config=config,
speech_featurizer=speech_featurizer,
text_featurizer=text_featurizer)
batch_size = bs or config.learning_config.running_config.batch_size
test_data_loader = test_dataset.create(batch_size)
exec_helpers.run_testing(model=contextnet,
test_dataset=test_dataset,
test_data_loader=test_data_loader,
output=output)
def test_contextnet():
config = Config(DEFAULT_YAML)
text_featurizer = CharFeaturizer(config.decoder_config)
speech_featurizer = TFSpeechFeaturizer(config.speech_config)
model = ContextNet(vocabulary_size=text_featurizer.num_classes,
**config.model_config)
model.make(speech_featurizer.shape)
model.summary(line_length=150)
model.add_featurizers(speech_featurizer=speech_featurizer,
text_featurizer=text_featurizer)
concrete_func = model.make_tflite_function(
timestamp=False).get_concrete_function()
converter = tf.lite.TFLiteConverter.from_concrete_functions(
[concrete_func])
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.experimental_new_converter = True
converter.target_spec.supported_ops = [
tf.lite.OpsSet.TFLITE_BUILTINS, tf.lite.OpsSet.SELECT_TF_OPS
]
tflite = converter.convert()
logger.info("Converted successfully with no timestamp")
concrete_func = model.make_tflite_function(
timestamp=True).get_concrete_function()
converter = tf.lite.TFLiteConverter.from_concrete_functions(
[concrete_func])
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.experimental_new_converter = True
converter.target_spec.supported_ops = [
tf.lite.OpsSet.TFLITE_BUILTINS, tf.lite.OpsSet.SELECT_TF_OPS
]
converter.convert()
logger.info("Converted successfully with timestamp")
tflitemodel = tf.lite.Interpreter(model_content=tflite)
signal = tf.random.normal([4000])
input_details = tflitemodel.get_input_details()
output_details = tflitemodel.get_output_details()
tflitemodel.resize_tensor_input(input_details[0]["index"], [4000])
tflitemodel.allocate_tensors()
tflitemodel.set_tensor(input_details[0]["index"], signal)
tflitemodel.set_tensor(input_details[1]["index"],
tf.constant(text_featurizer.blank, dtype=tf.int32))
tflitemodel.set_tensor(
input_details[2]["index"],
tf.zeros([
config.model_config["prediction_num_rnns"], 2, 1,
config.model_config["prediction_rnn_units"]
],
dtype=tf.float32))
tflitemodel.invoke()
hyp = tflitemodel.get_tensor(output_details[0]["index"])
logger.info(hyp)
train_dataset.load_metadata(args.metadata)
eval_dataset.load_metadata(args.metadata)
if not args.static_length:
speech_featurizer.reset_length()
text_featurizer.reset_length()
global_batch_size = args.bs or 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():
# 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 args.pretrained:
contextnet.load_weights(args.pretrained,
by_name=True,
skip_mismatch=True)
contextnet.summary(line_length=100)
optimizer = tf.keras.optimizers.Adam(
TransformerSchedule(
d_model=contextnet.dmodel,
warmup_steps=config.learning_config.optimizer_config.pop(
"warmup_steps", 10000),
max_lr=(0.05 / math.sqrt(contextnet.dmodel))),
**config.learning_config.optimizer_config)
def main(
config: str = DEFAULT_YAML,
tfrecords: bool = False,
sentence_piece: bool = False,
subwords: bool = True,
bs: int = None,
spx: int = 1,
metadata: str = None,
static_length: bool = False,
devices: list = [0],
mxp: bool = False,
pretrained: str = None,
):
tf.keras.backend.clear_session()
tf.config.optimizer.set_experimental_options({"auto_mixed_precision": mxp})
strategy = env_util.setup_strategy(devices)
config = Config(config)
speech_featurizer, text_featurizer = featurizer_helpers.prepare_featurizers(
config=config,
subwords=subwords,
sentence_piece=sentence_piece,
)
train_dataset, eval_dataset = dataset_helpers.prepare_training_datasets(
config=config,
speech_featurizer=speech_featurizer,
text_featurizer=text_featurizer,
tfrecords=tfrecords,
metadata=metadata,
)
if not static_length:
speech_featurizer.reset_length()
text_featurizer.reset_length()
train_data_loader, eval_data_loader, global_batch_size = dataset_helpers.prepare_training_data_loaders(
config=config,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
strategy=strategy,
batch_size=bs,
)
with strategy.scope():
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:
contextnet.load_weights(pretrained, by_name=True, skip_mismatch=True)
contextnet.summary(line_length=100)
optimizer = tf.keras.optimizers.Adam(
TransformerSchedule(
d_model=contextnet.dmodel,
warmup_steps=config.learning_config.optimizer_config.pop("warmup_steps", 10000),
max_lr=(0.05 / math.sqrt(contextnet.dmodel)),
),
**config.learning_config.optimizer_config
)
contextnet.compile(
optimizer=optimizer,
experimental_steps_per_execution=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),
]
contextnet.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 if eval_data_loader else None,
)