def wav_to_text(manifest, greedy=True): from ruamel.yaml import YAML yaml = YAML(typ="safe") with open(MODEL_YAML) as f: jasper_model_definition = yaml.load(f) labels = jasper_model_definition['labels'] data_layer = nemo_asr.AudioToTextDataLayer(shuffle=False, manifest_filepath=manifest, labels=labels, batch_size=1) audio_signal, audio_signal_len, _, _ = data_layer() processed_signal, processed_signal_len = data_preprocessor( input_signal=audio_signal, length=audio_signal_len) encoded, encoded_len = jasper_encoder(audio_signal=processed_signal, length=processed_signal_len) log_probs = jasper_decoder(encoder_output=encoded) predictions = greedy_decoder(log_probs=log_probs) if ENABLE_NGRAM: print('Running with beam search') beam_predictions = beam_search_with_lm( log_probs=log_probs, log_probs_length=encoded_len) eval_tensors = [beam_predictions] if greedy: eval_tensors = [predictions] tensors = neural_factory.infer(tensors=eval_tensors) if greedy: from nemo_asr.helpers import post_process_predictions prediction = post_process_predictions(tensors[0], labels) else: prediction = tensors[0][0][0][0][1] return prediction
def transcribe(self, manifest_path): """Reads audio file and returns the recognized transcrition""" self.nf.logger.info('Started Transcribing Speech') data_layer = nemo_asr.AudioToTextDataLayer( manifest_filepath = manifest_path, sample_rate = self.sample_rate, labels = self.labels, batch_size = 1, **self.eval_dl_params) # os.remove("data.json") self.nf.logger.info('Loading {0} examples'.format(len(data_layer))) audio_sig_e1, a_sig_length_e1, transcript_e1, transcript_len_e1 = data_layer() # apply pre-processing processed_signal_e1, p_length_e1 = self.preprocessor( input_signal = audio_sig_e1, length = a_sig_length_e1) # encode audio signal encoded_e1, encoded_len_e1 = self.jasper_encoder( audio_signal=processed_signal_e1, length=p_length_e1) # decode encoded signal log_probs_e1 = self.jasper_decoder(encoder_output=encoded_e1) # apply CTC decode if self.asr_conf["decoder"] == "beam": beam_predictions_e1 = self.ctc_decoder( log_probs=log_probs_e1, log_probs_length=encoded_len_e1) evaluated_tensors = self.nf.infer( tensors=[beam_predictions_e1,encoded_e1], use_cache=False) hypotheses = [] # Over mini-batch print("done1") return evaluated_tensors for i in evaluated_tensors[1]: hypotheses.append(i) else: greedy_predictions_e1 = self.ctc_decoder(log_probs=log_probs_e1) eval_tensors = [log_probs_e1, greedy_predictions_e1, transcript_e1, transcript_len_e1, encoded_len_e1,encoded_e1] evaluated_tensors = self.nf.infer( tensors = eval_tensors, cache = True ) print("done2") return evaluated_tensors hypotheses = post_process_predictions( evaluated_tensors[1], self.labels) return hypotheses
def infer(self,filepaths=None,waveforms=None,return_logits=False): """ Perform ASR inference on either a list of files or waveforms Arguments: filepaths: List of absolute filepaths to the .wav files transcribe waveforms: List of waveforms to transcribe. If filepaths is None, then waveforms must be specified. return_logits: If true, also return the logits output by the decoder Returns: return_dict: A dictionary with the following fields, where each field is a list with an element for each element of either filepaths or waveforms. greedy_prediction: The result of greedy ctc decoding greedy_transcript: The transcript form of the greedy prediction logits: decoder output logits """ if filepaths is not None: waveforms = [] for filepath in filepaths: waveform,sr = librosa.core.load(filepath,sr=self.model_definition['sample_rate']) waveforms.append(waveform) self.data_layer.set_signal(waveforms) elif waveforms is not None: self.data_layer.set_signal(waveforms) else: raise ValueError("Need filepaths or waveforms") tensors_to_evaluate = [self.predictions] if return_logits: tensors_to_evaluate.append(self.log_probs) evaluated_tensors = self.neural_factory.infer(tensors_to_evaluate,verbose=False) greedy_transcript = post_process_predictions(evaluated_tensors[0],self.vocab) result_dict = {'greedy prediction':evaluated_tensors[0]} result_dict['greedy transcript']=greedy_transcript if return_logits: result_dict['logits']=evaluated_tensors[1] return result_dict
def main(): parser = argparse.ArgumentParser(description='Jasper') parser.add_argument("--local_rank", default=None, type=int) parser.add_argument("--batch_size", default=32, type=int) parser.add_argument("--model_config", type=str, required=True) parser.add_argument("--eval_datasets", type=str, required=True) parser.add_argument("--load_dir", type=str, required=True) parser.add_argument("--vocab_file", type=str, required=True) parser.add_argument("--save_logprob", default=None, type=str) parser.add_argument("--lm_path", default=None, type=str) parser.add_argument("--beam_width", default=50, type=int) parser.add_argument("--alpha", default=2.0, type=float) parser.add_argument("--beta", default=1.0, type=float) parser.add_argument("--cutoff_prob", default=0.99, type=float) parser.add_argument("--cutoff_top_n", default=40, type=int) args = parser.parse_args() batch_size = args.batch_size load_dir = args.load_dir if args.local_rank is not None: if args.lm_path: raise NotImplementedError( "Beam search decoder with LM does not currently support " "evaluation on multi-gpu.") device = nemo.core.DeviceType.AllGpu else: device = nemo.core.DeviceType.GPU # Instantiate Neural Factory with supported backend neural_factory = nemo.core.NeuralModuleFactory( backend=nemo.core.Backend.PyTorch, local_rank=args.local_rank, optimization_level=nemo.core.Optimization.mxprO1, placement=device) logger = neural_factory.logger if args.local_rank is not None: logger.info('Doing ALL GPU') yaml = YAML(typ="safe") with open(args.model_config) as f: jasper_params = yaml.load(f) vocab = load_vocab(args.vocab_file) sample_rate = jasper_params['sample_rate'] eval_datasets = args.eval_datasets eval_dl_params = copy.deepcopy(jasper_params["AudioToTextDataLayer"]) eval_dl_params.update(jasper_params["AudioToTextDataLayer"]["eval"]) eval_dl_params["normalize_transcripts"] = False del eval_dl_params["train"] del eval_dl_params["eval"] data_layer = nemo_asr.AudioToTextDataLayer( manifest_filepath=eval_datasets, sample_rate=sample_rate, labels=vocab, batch_size=batch_size, **eval_dl_params) n = len(data_layer) logger.info('Evaluating {0} examples'.format(n)) data_preprocessor = nemo_asr.AudioPreprocessing( sample_rate=sample_rate, **jasper_params["AudioPreprocessing"]) jasper_encoder = nemo_asr.JasperEncoder( feat_in=jasper_params["AudioPreprocessing"]["features"], **jasper_params["JasperEncoder"]) jasper_decoder = nemo_asr.JasperDecoderForCTC( feat_in=jasper_params["JasperEncoder"]["jasper"][-1]["filters"], num_classes=len(vocab)) greedy_decoder = nemo_asr.GreedyCTCDecoder() if args.lm_path: beam_width = args.beam_width alpha = args.alpha beta = args.beta cutoff_prob = args.cutoff_prob cutoff_top_n = args.cutoff_top_n beam_search_with_lm = nemo_asr.BeamSearchDecoderWithLM( vocab=vocab, beam_width=beam_width, alpha=alpha, beta=beta, cutoff_prob=cutoff_prob, cutoff_top_n=cutoff_top_n, lm_path=args.lm_path, num_cpus=max(os.cpu_count(), 1)) logger.info('================================') logger.info( f"Number of parameters in encoder: {jasper_encoder.num_weights}") logger.info( f"Number of parameters in decoder: {jasper_decoder.num_weights}") logger.info( f"Total number of parameters in decoder: " f"{jasper_decoder.num_weights + jasper_encoder.num_weights}") logger.info('================================') audio_signal_e1, a_sig_length_e1, transcript_e1, transcript_len_e1 = \ data_layer() processed_signal_e1, p_length_e1 = data_preprocessor( input_signal=audio_signal_e1, length=a_sig_length_e1) encoded_e1, encoded_len_e1 = jasper_encoder( audio_signal=processed_signal_e1, length=p_length_e1) log_probs_e1 = jasper_decoder(encoder_output=encoded_e1) predictions_e1 = greedy_decoder(log_probs=log_probs_e1) eval_tensors = [log_probs_e1, predictions_e1, transcript_e1, transcript_len_e1, encoded_len_e1] if args.lm_path: beam_predictions_e1 = beam_search_with_lm( log_probs=log_probs_e1, log_probs_length=encoded_len_e1) eval_tensors.append(beam_predictions_e1) evaluated_tensors = neural_factory.infer( tensors=eval_tensors, checkpoint_dir=load_dir, ) greedy_hypotheses = post_process_predictions(evaluated_tensors[1], vocab) references = post_process_transcripts( evaluated_tensors[2], evaluated_tensors[3], vocab) cer = word_error_rate(hypotheses=greedy_hypotheses, references=references, use_cer=True) logger.info("Greedy CER {:.2f}%".format(cer * 100)) if args.lm_path: beam_hypotheses = [] # Over mini-batch for i in evaluated_tensors[-1]: # Over samples for j in i: beam_hypotheses.append(j[0][1]) cer = word_error_rate( hypotheses=beam_hypotheses, references=references, use_cer=True) logger.info("Beam CER {:.2f}".format(cer * 100)) if args.save_logprob: # Convert logits to list of numpy arrays logprob = [] for i, batch in enumerate(evaluated_tensors[0]): for j in range(batch.shape[0]): logprob.append( batch[j][:evaluated_tensors[4][i][j], :].cpu().numpy()) with open(args.save_logprob, 'wb') as f: pickle.dump(logprob, f, protocol=pickle.HIGHEST_PROTOCOL)
def main(): parser = argparse.ArgumentParser(parents=[nm_argparse.NemoArgParser()], description='AN4 ASR', conflict_handler='resolve') # Overwrite default args parser.add_argument("--train_dataset", type=str, help="training dataset path") parser.add_argument("--eval_datasets", type=str, nargs=1, help="validation dataset path") # Create new args parser.add_argument("--lm", default="./an4-lm.3gram.binary", type=str) parser.add_argument("--test_after_training", action='store_true') parser.add_argument("--momentum", type=float) parser.add_argument("--beta1", default=0.95, type=float) parser.add_argument("--beta2", default=0.25, type=float) parser.set_defaults( model_config="./configs/jasper_an4.yaml", train_dataset="/home/mrjenkins/TestData/an4_dataset/an4_train.json", eval_datasets="/home/mrjenkins/TestData/an4_dataset/an4_val.json", work_dir="./tmp", checkpoint_dir="./tmp", optimizer="novograd", num_epochs=50, batch_size=32, eval_batch_size=16, lr=0.02, weight_decay=0.005, checkpoint_save_freq=1000, eval_freq=100, amp_opt_level="O1") args = parser.parse_args() betas = (args.beta1, args.beta2) wer_thr = 0.20 beam_wer_thr = 0.15 nf = nemo.core.NeuralModuleFactory(local_rank=args.local_rank, optimization_level=args.amp_opt_level, random_seed=0, log_dir=args.work_dir, checkpoint_dir=args.checkpoint_dir, create_tb_writer=True, cudnn_benchmark=args.cudnn_benchmark) tb_writer = nf.tb_writer checkpoint_dir = nf.checkpoint_dir args.checkpoint_dir = nf.checkpoint_dir # Load model definition yaml = YAML(typ="safe") with open(args.model_config) as f: jasper_params = yaml.load(f) vocab = jasper_params['labels'] sample_rate = jasper_params['sample_rate'] # build train and eval model train_dl_params = copy.deepcopy(jasper_params["AudioToTextDataLayer"]) train_dl_params.update(jasper_params["AudioToTextDataLayer"]["train"]) del train_dl_params["train"] del train_dl_params["eval"] data_layer = nemo_asr.AudioToTextDataLayer( manifest_filepath=args.train_dataset, sample_rate=sample_rate, labels=vocab, batch_size=args.batch_size, **train_dl_params) num_samples = len(data_layer) total_steps = int(num_samples * args.num_epochs / args.batch_size) print("Train samples=", num_samples, "num_steps=", total_steps) data_preprocessor = nemo_asr.AudioPreprocessing( sample_rate=sample_rate, **jasper_params["AudioPreprocessing"]) # data_augmentation = nemo_asr.SpectrogramAugmentation( # **jasper_params['SpectrogramAugmentation'] # ) eval_dl_params = copy.deepcopy(jasper_params["AudioToTextDataLayer"]) eval_dl_params.update(jasper_params["AudioToTextDataLayer"]["eval"]) del eval_dl_params["train"] del eval_dl_params["eval"] data_layer_eval = nemo_asr.AudioToTextDataLayer( manifest_filepath=args.eval_datasets, sample_rate=sample_rate, labels=vocab, batch_size=args.eval_batch_size, **eval_dl_params) num_samples = len(data_layer_eval) nf.logger.info(f"Eval samples={num_samples}") jasper_encoder = nemo_asr.JasperEncoder( feat_in=jasper_params["AudioPreprocessing"]["features"], **jasper_params["JasperEncoder"]) jasper_decoder = nemo_asr.JasperDecoderForCTC( feat_in=jasper_params["JasperEncoder"]["jasper"][-1]["filters"], num_classes=len(vocab)) ctc_loss = nemo_asr.CTCLossNM(num_classes=len(vocab)) greedy_decoder = nemo_asr.GreedyCTCDecoder() # Training model audio, audio_len, transcript, transcript_len = data_layer() processed, processed_len = data_preprocessor(input_signal=audio, length=audio_len) encoded, encoded_len = jasper_encoder(audio_signal=processed, length=processed_len) log_probs = jasper_decoder(encoder_output=encoded) predictions = greedy_decoder(log_probs=log_probs) loss = ctc_loss(log_probs=log_probs, targets=transcript, input_length=encoded_len, target_length=transcript_len) # Evaluation model audio_e, audio_len_e, transcript_e, transcript_len_e = data_layer_eval() processed_e, processed_len_e = data_preprocessor(input_signal=audio_e, length=audio_len_e) encoded_e, encoded_len_e = jasper_encoder(audio_signal=processed_e, length=processed_len_e) log_probs_e = jasper_decoder(encoder_output=encoded_e) predictions_e = greedy_decoder(log_probs=log_probs_e) loss_e = ctc_loss(log_probs=log_probs_e, targets=transcript_e, input_length=encoded_len_e, target_length=transcript_len_e) nf.logger.info("Num of params in encoder: {0}".format( jasper_encoder.num_weights)) # Callbacks to print info to console and Tensorboard train_callback = nemo.core.SimpleLossLoggerCallback( tensors=[loss, predictions, transcript, transcript_len], print_func=lambda x: monitor_asr_train_progress(x, labels=vocab), get_tb_values=lambda x: [["loss", x[0]]], tb_writer=tb_writer, ) checkpointer_callback = nemo.core.CheckpointCallback( folder=checkpoint_dir, step_freq=args.checkpoint_save_freq) eval_tensors = [loss_e, predictions_e, transcript_e, transcript_len_e] eval_callback = nemo.core.EvaluatorCallback( eval_tensors=eval_tensors, user_iter_callback=lambda x, y: process_evaluation_batch( x, y, labels=vocab), user_epochs_done_callback=process_evaluation_epoch, eval_step=args.eval_freq, tb_writer=tb_writer) nf.train(tensors_to_optimize=[loss], callbacks=[train_callback, eval_callback, checkpointer_callback], optimizer=args.optimizer, lr_policy=CosineAnnealing(total_steps=total_steps), optimization_params={ "num_epochs": args.num_epochs, "max_steps": args.max_steps, "lr": args.lr, "momentum": args.momentum, "betas": betas, "weight_decay": args.weight_decay, "grad_norm_clip": None }, batches_per_step=args.iter_per_step) if args.test_after_training: # Create BeamSearch NM beam_search_with_lm = nemo_asr.BeamSearchDecoderWithLM( vocab=vocab, beam_width=64, alpha=2., beta=1.5, lm_path=args.lm, num_cpus=max(os.cpu_count(), 1)) beam_predictions = beam_search_with_lm(log_probs=log_probs_e, log_probs_length=encoded_len_e) eval_tensors.append(beam_predictions) evaluated_tensors = nf.infer(eval_tensors) greedy_hypotheses = post_process_predictions(evaluated_tensors[1], vocab) references = post_process_transcripts(evaluated_tensors[2], evaluated_tensors[3], vocab) wer = word_error_rate(hypotheses=greedy_hypotheses, references=references) nf.logger.info("Greedy WER: {:.2f}".format(wer * 100)) assert wer <= wer_thr, ( "Final eval greedy WER {:.2f}% > than {:.2f}%".format( wer * 100, wer_thr * 100)) beam_hypotheses = [] # Over mini-batch for i in evaluated_tensors[-1]: # Over samples for j in i: beam_hypotheses.append(j[0][1]) beam_wer = word_error_rate(hypotheses=beam_hypotheses, references=references) nf.logger.info("Beam WER {:.2f}%".format(beam_wer * 100)) assert beam_wer <= beam_wer_thr, ( "Final eval beam WER {:.2f}% > than {:.2f}%".format( beam_wer * 100, beam_wer_thr * 100)) assert beam_wer <= wer, ("Final eval beam WER > than the greedy WER.") # Reload model weights and train for extra 10 epochs checkpointer_callback = nemo.core.CheckpointCallback( folder=checkpoint_dir, step_freq=args.checkpoint_save_freq, force_load=True) nf.reset_trainer() nf.train(tensors_to_optimize=[loss], callbacks=[train_callback, checkpointer_callback], optimizer=args.optimizer, optimization_params={ "num_epochs": args.num_epochs + 10, "lr": args.lr, "momentum": args.momentum, "betas": betas, "weight_decay": args.weight_decay, "grad_norm_clip": None }, reset=True) evaluated_tensors = nf.infer(eval_tensors[:-1]) greedy_hypotheses = post_process_predictions(evaluated_tensors[1], vocab) references = post_process_transcripts(evaluated_tensors[2], evaluated_tensors[3], vocab) wer_new = word_error_rate(hypotheses=greedy_hypotheses, references=references) nf.logger.info("New greedy WER: {:.2f}%".format(wer_new * 100)) assert wer_new <= wer * 1.1, ( f"Fine tuning: new WER {wer * 100:.2f}% > than the previous WER " f"{wer_new * 100:.2f}%")
def main(): parser = argparse.ArgumentParser(description='Jasper') parser.add_argument("--local_rank", default=None, type=int) parser.add_argument("--batch_size", default=32, type=int) parser.add_argument("--model_config", type=str, required=True) parser.add_argument("--eval_datasets", type=str, required=True) parser.add_argument("--load_dir", type=str, required=True) parser.add_argument("--save_logprob", default=None, type=str) parser.add_argument("--lm_path", default=None, type=str) parser.add_argument('--alpha', default=2., type=float, help='value of LM weight', required=False) parser.add_argument( '--alpha_max', type=float, help='maximum value of LM weight (for a grid search in \'eval\' mode)', required=False) parser.add_argument('--alpha_step', type=float, help='step for LM weight\'s tuning in \'eval\' mode', required=False, default=0.1) parser.add_argument('--beta', default=1.5, type=float, help='value of word count weight', required=False) parser.add_argument( '--beta_max', type=float, help='maximum value of word count weight (for a grid search in \ \'eval\' mode', required=False) parser.add_argument( '--beta_step', type=float, help='step for word count weight\'s tuning in \'eval\' mode', required=False, default=0.1) parser.add_argument("--beam_width", default=128, type=int) args = parser.parse_args() batch_size = args.batch_size load_dir = args.load_dir if args.local_rank is not None: if args.lm_path: raise NotImplementedError( "Beam search decoder with LM does not currently support " "evaluation on multi-gpu.") device = nemo.core.DeviceType.AllGpu else: device = nemo.core.DeviceType.GPU # Instantiate Neural Factory with supported backend neural_factory = nemo.core.NeuralModuleFactory( backend=nemo.core.Backend.PyTorch, local_rank=args.local_rank, optimization_level=nemo.core.Optimization.mxprO1, placement=device) logger = neural_factory.logger if args.local_rank is not None: logger.info('Doing ALL GPU') yaml = YAML(typ="safe") with open(args.model_config) as f: jasper_params = yaml.load(f) vocab = jasper_params['labels'] sample_rate = jasper_params['sample_rate'] eval_datasets = args.eval_datasets eval_dl_params = copy.deepcopy(jasper_params["AudioToTextDataLayer"]) eval_dl_params.update(jasper_params["AudioToTextDataLayer"]["eval"]) del eval_dl_params["train"] del eval_dl_params["eval"] data_layer = nemo_asr.AudioToTextDataLayer(manifest_filepath=eval_datasets, sample_rate=sample_rate, labels=vocab, batch_size=batch_size, **eval_dl_params) N = len(data_layer) logger.info('Evaluating {0} examples'.format(N)) data_preprocessor = nemo_asr.AudioToMelSpectrogramPreprocessor( sample_rate=sample_rate, **jasper_params["AudioToMelSpectrogramPreprocessor"]) jasper_encoder = nemo_asr.JasperEncoder( feat_in=jasper_params["AudioToMelSpectrogramPreprocessor"]["features"], **jasper_params["JasperEncoder"]) jasper_decoder = nemo_asr.JasperDecoderForCTC( feat_in=jasper_params["JasperEncoder"]["jasper"][-1]["filters"], num_classes=len(vocab)) greedy_decoder = nemo_asr.GreedyCTCDecoder() logger.info('================================') logger.info( f"Number of parameters in encoder: {jasper_encoder.num_weights}") logger.info( f"Number of parameters in decoder: {jasper_decoder.num_weights}") logger.info(f"Total number of parameters in decoder: " f"{jasper_decoder.num_weights + jasper_encoder.num_weights}") logger.info('================================') audio_signal_e1, a_sig_length_e1, transcript_e1, transcript_len_e1 =\ data_layer() processed_signal_e1, p_length_e1 = data_preprocessor( input_signal=audio_signal_e1, length=a_sig_length_e1) encoded_e1, encoded_len_e1 = jasper_encoder( audio_signal=processed_signal_e1, length=p_length_e1) log_probs_e1 = jasper_decoder(encoder_output=encoded_e1) predictions_e1 = greedy_decoder(log_probs=log_probs_e1) eval_tensors = [ log_probs_e1, predictions_e1, transcript_e1, transcript_len_e1, encoded_len_e1 ] evaluated_tensors = neural_factory.infer(tensors=eval_tensors, checkpoint_dir=load_dir, cache=True) greedy_hypotheses = post_process_predictions(evaluated_tensors[1], vocab) references = post_process_transcripts(evaluated_tensors[2], evaluated_tensors[3], vocab) wer = word_error_rate(hypotheses=greedy_hypotheses, references=references) logger.info("Greedy WER {:.2f}%".format(wer * 100)) if args.lm_path: if args.alpha_max is None: args.alpha_max = args.alpha # include alpha_max in tuning range args.alpha_max += args.alpha_step / 10.0 if args.beta_max is None: args.beta_max = args.beta # include beta_max in tuning range args.beta_max += args.beta_step / 10.0 beam_wers = [] for alpha in np.arange(args.alpha, args.alpha_max, args.alpha_step): for beta in np.arange(args.beta, args.beta_max, args.beta_step): logger.info('================================') logger.info(f'Infering with (alpha, beta): ({alpha}, {beta})') beam_search_with_lm = nemo_asr.BeamSearchDecoderWithLM( vocab=vocab, beam_width=args.beam_width, alpha=alpha, beta=beta, lm_path=args.lm_path, num_cpus=max(os.cpu_count(), 1)) beam_predictions_e1 = beam_search_with_lm( log_probs=log_probs_e1, log_probs_length=encoded_len_e1) evaluated_tensors = neural_factory.infer( tensors=[beam_predictions_e1], use_cache=True, verbose=False) beam_hypotheses = [] # Over mini-batch for i in evaluated_tensors[-1]: # Over samples for j in i: beam_hypotheses.append(j[0][1]) wer = word_error_rate(hypotheses=beam_hypotheses, references=references) logger.info("Beam WER {:.2f}%".format(wer * 100)) beam_wers.append(((alpha, beta), wer * 100)) logger.info('Beam WER for (alpha, beta)') logger.info('================================') logger.info('\n' + '\n'.join([str(e) for e in beam_wers])) logger.info('================================') best_beam_wer = min(beam_wers, key=lambda x: x[1]) logger.info('Best (alpha, beta): ' f'{best_beam_wer[0]}, ' f'WER: {best_beam_wer[1]:.2f}%')
def main(): parser = argparse.ArgumentParser(parents=[nm_argparse.NemoArgParser()], description='AN4 ASR', conflict_handler='resolve') # Overwrite default args parser.add_argument("--train_dataset", type=str, help="training dataset path") parser.add_argument("--eval_datasets", type=str, nargs=1, help="validation dataset path") # Create new args parser.add_argument("--lm", default="./an4-lm.3gram.binary", type=str) parser.add_argument("--test_after_training", action='store_true') parser.add_argument("--momentum", type=float) parser.add_argument("--beta1", default=0.95, type=float) parser.add_argument("--beta2", default=0.25, type=float) parser.set_defaults( model_config="./configs/jasper_an4.yaml", train_dataset="/home/mrjenkins/TestData/an4_dataset/an4_train.json", eval_datasets="/home/mrjenkins/TestData/an4_dataset/an4_val.json", work_dir="./tmp", optimizer="novograd", num_epochs=50, batch_size=48, eval_batch_size=64, lr=0.02, weight_decay=0.005, checkpoint_save_freq=1000, eval_freq=100, amp_opt_level="O1") args = parser.parse_args() betas = (args.beta1, args.beta2) wer_thr = 0.20 beam_wer_thr = 0.15 nf = nemo.core.NeuralModuleFactory(local_rank=args.local_rank, files_to_copy=[__file__], optimization_level=args.amp_opt_level, random_seed=0, log_dir=args.work_dir, create_tb_writer=True, cudnn_benchmark=args.cudnn_benchmark) tb_writer = nf.tb_writer checkpoint_dir = nf.checkpoint_dir # Load model definition yaml = YAML(typ="safe") with open(args.model_config) as f: jasper_params = yaml.load(f) (loss, eval_tensors, callbacks, total_steps, vocab, log_probs_e, encoded_len_e) = create_dags(jasper_params, args, nf) nf.train( tensors_to_optimize=[loss], callbacks=callbacks, optimizer=args.optimizer, lr_policy=CosineAnnealing(total_steps=total_steps, min_lr=args.lr / 100), optimization_params={ "num_epochs": args.num_epochs, "max_steps": args.max_steps, "lr": args.lr, "momentum": args.momentum, "betas": betas, "weight_decay": args.weight_decay, "grad_norm_clip": None }, batches_per_step=args.iter_per_step, amp_max_loss_scale=256., # synced_batchnorm=(nf.global_rank is not None), ) if args.test_after_training: nemo.logging.info("Testing greedy and beam search with LM WER.") # Create BeamSearch NM if nf.world_size > 1: nemo.logging.warning("Skipping beam search WER as it does not " "work if doing distributed training.") else: beam_search_with_lm = nemo_asr.BeamSearchDecoderWithLM( vocab=vocab, beam_width=64, alpha=2., beta=1.5, lm_path=args.lm, num_cpus=max(os.cpu_count(), 1)) beam_predictions = beam_search_with_lm( log_probs=log_probs_e, log_probs_length=encoded_len_e) eval_tensors.append(beam_predictions) evaluated_tensors = nf.infer(eval_tensors) if nf.global_rank in [0, None]: greedy_hypotheses = post_process_predictions( evaluated_tensors[1], vocab) references = post_process_transcripts(evaluated_tensors[2], evaluated_tensors[3], vocab) wer = word_error_rate(hypotheses=greedy_hypotheses, references=references) nemo.logging.info("Greedy WER: {:.2f}%".format(wer * 100)) if wer > wer_thr: nf.sync_all_processes(False) raise ValueError(f"Final eval greedy WER {wer*100:.2f}% > :" f"than {wer_thr*100:.2f}%") nf.sync_all_processes() if nf.world_size == 1: beam_hypotheses = [] # Over mini-batch for i in evaluated_tensors[-1]: # Over samples for j in i: beam_hypotheses.append(j[0][1]) beam_wer = word_error_rate(hypotheses=beam_hypotheses, references=references) nemo.logging.info("Beam WER {:.2f}%".format(beam_wer * 100)) assert beam_wer <= beam_wer_thr, ( "Final eval beam WER {:.2f}% > than {:.2f}%".format( beam_wer * 100, beam_wer_thr * 100)) assert beam_wer <= wer, ( "Final eval beam WER > than the greedy WER.") # Reload model weights and train for extra 10 epochs checkpointer_callback = nemo.core.CheckpointCallback( folder=checkpoint_dir, step_freq=args.checkpoint_save_freq, force_load=True) # Distributed Data Parallel changes the underlying class so we need # to reinstantiate Encoder and Decoder args.num_epochs += 10 previous_step_count = total_steps loss, eval_tensors, callbacks, total_steps, vocab, _, _ = create_dags( jasper_params, args, nf) nf.reset_trainer() nf.train( tensors_to_optimize=[loss], callbacks=callbacks, optimizer=args.optimizer, lr_policy=CosineAnnealing(warmup_steps=previous_step_count, total_steps=total_steps), optimization_params={ "num_epochs": args.num_epochs, "lr": args.lr / 100, "momentum": args.momentum, "betas": betas, "weight_decay": args.weight_decay, "grad_norm_clip": None }, reset=True, amp_max_loss_scale=256., # synced_batchnorm=(nf.global_rank is not None), ) evaluated_tensors = nf.infer(eval_tensors) if nf.global_rank in [0, None]: greedy_hypotheses = post_process_predictions( evaluated_tensors[1], vocab) references = post_process_transcripts(evaluated_tensors[2], evaluated_tensors[3], vocab) wer_new = word_error_rate(hypotheses=greedy_hypotheses, references=references) nemo.logging.info("New greedy WER: {:.2f}%".format(wer_new * 100)) if wer_new > wer * 1.1: nf.sync_all_processes(False) raise ValueError( f"Fine tuning: new WER {wer_new* 100:.2f}% > than the " f"previous WER {wer * 100:.2f}%") nf.sync_all_processes() # Open the log file and ensure that epochs is strictly increasing if nf._exp_manager.log_file: epochs = [] with open(nf._exp_manager.log_file, "r") as log_file: line = log_file.readline() while line: index = line.find("Starting epoch") if index != -1: epochs.append(int(line[index + len("Starting epoch"):])) line = log_file.readline() for i, e in enumerate(epochs): if i != e: raise ValueError("Epochs from logfile was not understood")
def offline_inference(config, encoder, decoder, audio_file): MODEL_YAML = config CHECKPOINT_ENCODER = encoder CHECKPOINT_DECODER = decoder sample_rate, signal = wave.read(audio_file) # get labels (vocab) yaml = YAML(typ="safe") with open(MODEL_YAML) as f: jasper_model_definition = yaml.load(f) labels = jasper_model_definition['labels'] # build neural factory and neural modules neural_factory = nemo.core.NeuralModuleFactory( placement=nemo.core.DeviceType.GPU, backend=nemo.core.Backend.PyTorch) data_preprocessor = nemo_asr.AudioToMelSpectrogramPreprocessor( factory=neural_factory, **jasper_model_definition["AudioToMelSpectrogramPreprocessor"]) jasper_encoder = nemo_asr.JasperEncoder( feat_in=jasper_model_definition["AudioToMelSpectrogramPreprocessor"]["features"], **jasper_model_definition["JasperEncoder"]) jasper_decoder = nemo_asr.JasperDecoderForCTC( feat_in=jasper_model_definition["JasperEncoder"]["jasper"][-1]["filters"], num_classes=len(labels)) greedy_decoder = nemo_asr.GreedyCTCDecoder() # load model jasper_encoder.restore_from(CHECKPOINT_ENCODER) jasper_decoder.restore_from(CHECKPOINT_DECODER) # AudioDataLayer class AudioDataLayer(DataLayerNM): @staticmethod def create_ports(): input_ports = {} output_ports = { "audio_signal": NeuralType({0: AxisType(BatchTag), 1: AxisType(TimeTag)}), "a_sig_length": NeuralType({0: AxisType(BatchTag)}), } return input_ports, output_ports def __init__(self, **kwargs): DataLayerNM.__init__(self, **kwargs) self.output_enable = False def __iter__(self): return self def __next__(self): if not self.output_enable: raise StopIteration self.output_enable = False return torch.as_tensor(self.signal, dtype=torch.float32), \ torch.as_tensor(self.signal_shape, dtype=torch.int64) def set_signal(self, signal): self.signal = np.reshape(signal.astype(np.float32)/32768., [1, -1]) self.signal_shape = np.expand_dims(self.signal.size, 0).astype(np.int64) self.output_enable = True def __len__(self): return 1 @property def dataset(self): return None @property def data_iterator(self): return self # Instantiate necessary neural modules data_layer = AudioDataLayer() # Define inference DAG audio_signal, audio_signal_len = data_layer() processed_signal, processed_signal_len = data_preprocessor( input_signal=audio_signal, length=audio_signal_len) encoded, encoded_len = jasper_encoder(audio_signal=processed_signal, length=processed_signal_len) log_probs = jasper_decoder(encoder_output=encoded) predictions = greedy_decoder(log_probs=log_probs) # audio inference data_layer.set_signal(signal) tensors = neural_factory.infer([ audio_signal, processed_signal, encoded, log_probs, predictions], verbose=False) # results audio = tensors[0][0][0].cpu().numpy() features = tensors[1][0][0].cpu().numpy() encoded_features = tensors[2][0][0].cpu().numpy(), probs = tensors[3][0][0].cpu().numpy() preds = tensors[4][0] transcript = post_process_predictions([preds], labels) return transcript, audio, features, encoded_features, probs, preds