def main(): parser = argparse.ArgumentParser(prog="Conformer Training") parser.add_argument("--config", type=str, default=DEFAULT_YAML, help="The file path of model configuration file") parser.add_argument("--max_ckpts", type=int, default=10, help="Max number of checkpoints to keep") parser.add_argument("--tbs", type=int, default=None, help="Train batch size per replica") parser.add_argument("--ebs", type=int, default=None, help="Evaluation batch size per replica") parser.add_argument("--acs", type=int, default=None, help="Train accumulation steps") 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") parser.add_argument("--subwords", type=str, default=None, help="Path to file that stores generated subwords") parser.add_argument("--subwords_corpus", nargs="*", type=str, default=[], help="Transcript files for generating subwords") parser.add_argument("--train-dir", '-td', nargs='*', required=True) parser.add_argument("--dev-dir", '-dd', nargs='*', required=True) args = parser.parse_args() tf.config.optimizer.set_experimental_options( {"auto_mixed_precision": args.mxp}) strategy = setup_strategy(args.devices) config = Config(args.config, learning=True) with open(config.speech_config) as f: speech_config = yaml.load(f, Loader=yaml.Loader) speech_featurizer = TFSpeechFeaturizer(speech_config) if args.subwords and os.path.exists(args.subwords): print("Loading subwords ...") text_featurizer = SubwordFeaturizer.load_from_file( config.decoder_config, args.subwords) else: print("Generating subwords ...") text_featurizer = SubwordFeaturizer.build_from_corpus( config.decoder_config, corpus_files=args.subwords_corpus) text_featurizer.save_to_file(args.subwords) train_dataset = Dataset(data_paths=args.train_dir, speech_featurizer=speech_featurizer, text_featurizer=text_featurizer, augmentations=config.learning_config.augmentations, stage="train", cache=False, shuffle=False) eval_dataset = Dataset(data_paths=args.dev_dir, speech_featurizer=speech_featurizer, text_featurizer=text_featurizer, stage="eval", cache=False, shuffle=False) conformer_trainer = TransducerTrainer( config=config.learning_config.running_config, text_featurizer=text_featurizer, strategy=strategy) with conformer_trainer.strategy.scope(): # build model conformer = Conformer(**config.model_config, vocabulary_size=text_featurizer.num_classes) conformer._build(speech_featurizer.shape) conformer.summary(line_length=120) optimizer = tf.keras.optimizers.Adam( TransformerSchedule(d_model=conformer.dmodel, warmup_steps=config.learning_config. optimizer_config["warmup_steps"], max_lr=(0.05 / math.sqrt(conformer.dmodel))), beta_1=config.learning_config.optimizer_config["beta1"], beta_2=config.learning_config.optimizer_config["beta2"], epsilon=config.learning_config.optimizer_config["epsilon"]) conformer_trainer.compile(model=conformer, optimizer=optimizer, max_to_keep=args.max_ckpts) conformer_trainer.fit(train_dataset, eval_dataset, train_bs=args.tbs, eval_bs=args.ebs, train_acs=args.acs)
eval_dataset = ASRSliceDataset( data_paths=config.learning_config.dataset_config.eval_paths, speech_featurizer=speech_featurizer, text_featurizer=text_featurizer, stage="eval", cache=args.cache, shuffle=True ) conformer_trainer = TrainerWithMaskingGA( config=config.learning_config.running_config, text_featurizer=text_featurizer, strategy=strategy ) with conformer_trainer.strategy.scope(): # build model conformer = Conformer(**config.model_config, vocabulary_size=text_featurizer.num_classes) conformer._build(speech_featurizer.shape) conformer.summary(line_length=120) optimizer = tf.keras.optimizers.Adam( TransformerSchedule( d_model=config.model_config["encoder_dmodel"], warmup_steps=config.learning_config.optimizer_config["warmup_steps"], max_lr=(0.05 / math.sqrt(config.model_config["encoder_dmodel"])) ), beta_1=config.learning_config.optimizer_config["beta1"], beta_2=config.learning_config.optimizer_config["beta2"], epsilon=config.learning_config.optimizer_config["epsilon"] ) conformer_trainer.compile(model=conformer, optimizer=optimizer, max_to_keep=args.max_ckpts)
def test_conformer(): config = Config(DEFAULT_YAML, learning=False) text_featurizer = CharFeaturizer(config.decoder_config) speech_featurizer = TFSpeechFeaturizer(config.speech_config) model = Conformer(vocabulary_size=text_featurizer.num_classes, **config.model_config) model._build(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() print("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() print("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"]) print(hyp)