def main(): parser = argparse.ArgumentParser(description='PyTorch Tacotron 2 Training') parser = parse_args(parser) args, _ = parser.parse_known_args() LOGGER.set_model_name("Tacotron2_PyT") LOGGER.set_backends([ dllg.StdOutBackend(log_file=None, logging_scope=dllg.TRAIN_ITER_SCOPE, iteration_interval=1), dllg.JsonBackend(log_file=os.path.join(args.output_directory, args.log_file) if args.rank == 0 else None, logging_scope=dllg.TRAIN_ITER_SCOPE, iteration_interval=1) ]) LOGGER.timed_block_start("run") LOGGER.register_metric(tags.TRAIN_ITERATION_LOSS, metric_scope=dllg.TRAIN_ITER_SCOPE) LOGGER.register_metric("iter_time", metric_scope=dllg.TRAIN_ITER_SCOPE) LOGGER.register_metric("epoch_time", metric_scope=dllg.EPOCH_SCOPE) LOGGER.register_metric("run_time", metric_scope=dllg.RUN_SCOPE) LOGGER.register_metric("val_iter_loss", metric_scope=dllg.EPOCH_SCOPE) LOGGER.register_metric("train_epoch_frames/sec", metric_scope=dllg.EPOCH_SCOPE) LOGGER.register_metric("train_epoch_avg_frames/sec", metric_scope=dllg.EPOCH_SCOPE) LOGGER.register_metric("train_epoch_avg_loss", metric_scope=dllg.EPOCH_SCOPE) log_hardware() parser = parse_tacotron2_args(parser) args = parser.parse_args() log_args(args) torch.backends.cudnn.enabled = args.cudnn_enabled torch.backends.cudnn.benchmark = args.cudnn_benchmark distributed_run = args.world_size > 1 if distributed_run: init_distributed(args, args.world_size, args.rank, args.group_name) os.makedirs(args.output_directory, exist_ok=True) LOGGER.log(key=tags.RUN_START) run_start_time = time.time() model = get_tacotron2_model(args, len(args.training_anchor_dirs), is_training=True) if not args.amp_run and distributed_run: model = DDP(model) model.restore_checkpoint(os.path.join(args.output_directory, args.latest_checkpoint_file)) optimizer = torch.optim.Adam(model.parameters(), lr=args.init_lr, weight_decay=args.weight_decay) writer = SummaryWriter(args.output_directory) if args.amp_run: model, optimizer = amp.initialize(model, optimizer, opt_level='O0') if distributed_run: model = DDP(model) criterion = Tacotron2Loss() collate_fn = TextMelCollate(args) train_dataset = TextMelDataset(args, args.training_anchor_dirs) train_loader = DataLoader(train_dataset, num_workers=2, shuffle=False, batch_size=args.batch_size//len(args.training_anchor_dirs), pin_memory=False, drop_last=True, collate_fn=collate_fn) # valate_dataset = TextMelDataset(args, args.validation_anchor_dirs) model.train() elapsed_epochs = model.get_elapsed_epochs() epochs = args.epochs - elapsed_epochs iteration = elapsed_epochs * len(train_loader) LOGGER.log(key=tags.TRAIN_LOOP) for epoch in range(1, epochs + 1): LOGGER.epoch_start() epoch_start_time = time.time() epoch += elapsed_epochs LOGGER.log(key=tags.TRAIN_EPOCH_START, value=epoch) # used to calculate avg frames/sec over epoch reduced_num_frames_epoch = 0 # used to calculate avg loss over epoch train_epoch_avg_loss = 0.0 train_epoch_avg_frames_per_sec = 0.0 num_iters = 0 adjust_learning_rate(optimizer, epoch, args) for i, batch in enumerate(train_loader): print(f"Batch: {i}/{len(train_loader)} epoch {epoch}") LOGGER.iteration_start() iter_start_time = time.time() LOGGER.log(key=tags.TRAIN_ITER_START, value=i) # start = time.perf_counter() optimizer.zero_grad() x, y, num_frames = batch_to_gpu(batch) outputs = model(x) y_pred = [output.cpu() for output in outputs] loss = criterion(y_pred, y) if distributed_run: reduced_loss = reduce_tensor(loss.data, args.world_size).item() reduced_num_frames = reduce_tensor(num_frames.data, 1).item() else: reduced_loss = loss.item() reduced_num_frames = num_frames.item() if np.isnan(reduced_loss): raise Exception("loss is NaN") LOGGER.log(key=tags.TRAIN_ITERATION_LOSS, value=reduced_loss) train_epoch_avg_loss += reduced_loss num_iters += 1 # accumulate number of frames processed in this epoch reduced_num_frames_epoch += reduced_num_frames if args.amp_run: with amp.scale_loss(loss, optimizer) as scaled_loss: scaled_loss.backward() grad_norm = torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.grad_clip_thresh) else: loss.backward() grad_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), args.grad_clip_thresh) optimizer.step() iteration += 1 writer.add_scalar('Training/Loss', reduced_loss, iteration) LOGGER.log(key=tags.TRAIN_ITER_STOP, value=i) iter_stop_time = time.time() iter_time = iter_stop_time - iter_start_time frames_per_sec = reduced_num_frames/iter_time train_epoch_avg_frames_per_sec += frames_per_sec LOGGER.log(key="train_iter_frames/sec", value=frames_per_sec) LOGGER.log(key="iter_time", value=iter_time) LOGGER.iteration_stop() LOGGER.log(key=tags.TRAIN_EPOCH_STOP, value=epoch) epoch_stop_time = time.time() epoch_time = epoch_stop_time - epoch_start_time LOGGER.log(key="train_epoch_frames/sec", value=(reduced_num_frames_epoch/epoch_time)) LOGGER.log(key="train_epoch_avg_frames/sec", value=(train_epoch_avg_frames_per_sec/num_iters if num_iters > 0 else 0.0)) LOGGER.log(key="train_epoch_avg_loss", value=(train_epoch_avg_loss/num_iters if num_iters > 0 else 0.0)) LOGGER.log(key="epoch_time", value=epoch_time) LOGGER.log(key=tags.EVAL_START, value=epoch) # validate(model, criterion, valate_dataset, iteration, collate_fn, distributed_run, args) LOGGER.log(key=tags.EVAL_STOP, value=epoch) # Store latest checkpoint in each epoch model.elapse_epoch() checkpoint_path = os.path.join(args.output_directory, args.latest_checkpoint_file) model.save_checkpoint(checkpoint_path) # Plot alignemnt if epoch % args.epochs_per_alignment == 0 and args.rank == 0: alignments = y_pred[3].data.numpy() index = np.random.randint(len(alignments)) plot_alignment(alignments[index], # [enc_step, dec_step] os.path.join(args.output_directory, f"align_{epoch:04d}_{iteration}.png"), info=f"{datetime.now().strftime('%Y-%m-%d %H:%M')} Epoch={epoch:04d} Iteration={iteration} Average loss={train_epoch_avg_loss/num_iters:.5f}") # Save checkpoint if epoch % args.epochs_per_checkpoint == 0 and args.rank == 0: checkpoint_path = os.path.join(args.output_directory, f"checkpoint_{epoch:04d}.pt") print(f"Saving model and optimizer state at epoch {epoch:04d} to {checkpoint_path}") model.save_checkpoint(checkpoint_path) # Save evaluation # save_sample(model, args.tacotron2_checkpoint, args.phrase_path, # os.path.join(args.output_directory, f"sample_{epoch:04d}_{iteration}.wav"), args.sampling_rate) LOGGER.epoch_stop() run_stop_time = time.time() run_time = run_stop_time - run_start_time LOGGER.log(key="run_time", value=run_time) LOGGER.log(key=tags.RUN_FINAL) print("training time", run_stop_time - run_start_time) writer.close() LOGGER.timed_block_stop("run") if args.rank == 0: LOGGER.finish()
def main(): """ Launches text to speech (inference). Inference is executed on a single GPU. """ parser = argparse.ArgumentParser( description='PyTorch Tacotron 2 Inference') parser = parse_args(parser) args, _ = parser.parse_known_args() LOGGER.set_model_name("Tacotron2_PyT") LOGGER.set_backends([ dllg.StdOutBackend(log_file=None, logging_scope=dllg.TRAIN_ITER_SCOPE, iteration_interval=1), dllg.JsonBackend(log_file=args.log_file, logging_scope=dllg.TRAIN_ITER_SCOPE, iteration_interval=1) ]) LOGGER.register_metric("tacotron2_items_per_sec", metric_scope=dllg.TRAIN_ITER_SCOPE) LOGGER.register_metric("tacotron2_latency", metric_scope=dllg.TRAIN_ITER_SCOPE) LOGGER.register_metric("waveglow_items_per_sec", metric_scope=dllg.TRAIN_ITER_SCOPE) LOGGER.register_metric("waveglow_latency", metric_scope=dllg.TRAIN_ITER_SCOPE) LOGGER.register_metric("latency", metric_scope=dllg.TRAIN_ITER_SCOPE) log_hardware() log_args(args) tacotron2 = load_and_setup_model('Tacotron2', parser, args.tacotron2, args.amp_run) waveglow = load_and_setup_model('WaveGlow', parser, args.waveglow, args.amp_run) denoiser = Denoiser(waveglow).cuda() tacotron2.forward = tacotron2.infer type(tacotron2).forward = type(tacotron2).infer jitted_tacotron2 = torch.jit.script(tacotron2) texts = [] try: f = open(args.input, 'r') texts = f.readlines() except: print("Could not read file") sys.exit(1) if args.include_warmup: sequence = torch.randint(low=0, high=148, size=(1,50), dtype=torch.long).cuda() input_lengths = torch.IntTensor([sequence.size(1)]).cuda().long() for i in range(3): with torch.no_grad(): _, mel, _, _, mel_lengths = jitted_tacotron2(sequence, input_lengths) _ = waveglow.infer(mel) LOGGER.iteration_start() measurements = {} sequences_padded, input_lengths = prepare_input_sequence(texts) with torch.no_grad(), MeasureTime(measurements, "tacotron2_time"): _, mel, _, _, mel_lengths = jitted_tacotron2(sequences_padded, input_lengths) with torch.no_grad(), MeasureTime(measurements, "waveglow_time"): audios = waveglow.infer(mel, sigma=args.sigma_infer) audios = audios.float() audios = denoiser(audios, strength=args.denoising_strength).squeeze(1) tacotron2_infer_perf = mel.size(0)*mel.size(2)/measurements['tacotron2_time'] waveglow_infer_perf = audios.size(0)*audios.size(1)/measurements['waveglow_time'] LOGGER.log(key="tacotron2_items_per_sec", value=tacotron2_infer_perf) LOGGER.log(key="tacotron2_latency", value=measurements['tacotron2_time']) LOGGER.log(key="waveglow_items_per_sec", value=waveglow_infer_perf) LOGGER.log(key="waveglow_latency", value=measurements['waveglow_time']) LOGGER.log(key="latency", value=(measurements['tacotron2_time']+ measurements['waveglow_time'])) for i, audio in enumerate(audios): audio = audio[:mel_lengths[i]*args.stft_hop_length] audio = audio/torch.max(torch.abs(audio)) audio_path = args.output + "audio_"+str(i)+".wav" write(audio_path, args.sampling_rate, audio.cpu().numpy()) LOGGER.iteration_stop() LOGGER.finish()