コード例 #1
0
    def test_train_step():
        input_dummy = torch.randint(0, 24, (8, 128)).long().to(device)
        input_lengths = torch.randint(100, 129, (8, )).long().to(device)
        input_lengths[-1] = 128
        mel_spec = torch.rand(8, 30, c.audio['num_mels']).to(device)
        linear_spec = torch.rand(8, 30, c.audio['num_freq']).to(device)
        mel_lengths = torch.randint(20, 30, (8, )).long().to(device)
        stop_targets = torch.zeros(8, 30, 1).float().to(device)
        speaker_ids = torch.randint(0, 5, (8, )).long().to(device)

        for idx in mel_lengths:
            stop_targets[:, int(idx.item()):, 0] = 1.0

        stop_targets = stop_targets.view(input_dummy.shape[0],
                                         stop_targets.size(1) // c.r, -1)
        stop_targets = (stop_targets.sum(2) >
                        0.0).unsqueeze(2).float().squeeze()

        criterion = L1LossMasked(seq_len_norm=False).to(device)
        criterion_st = nn.BCEWithLogitsLoss().to(device)
        model = Tacotron(
            num_chars=32,
            num_speakers=5,
            postnet_output_dim=c.audio['num_freq'],
            decoder_output_dim=c.audio['num_mels'],
            r=c.r,
            memory_size=c.memory_size
        ).to(device)  #FIXME: missing num_speakers parameter to Tacotron ctor
        model.train()
        print(" > Num parameters for Tacotron model:%s" %
              (count_parameters(model)))
        model_ref = copy.deepcopy(model)
        count = 0
        for param, param_ref in zip(model.parameters(),
                                    model_ref.parameters()):
            assert (param - param_ref).sum() == 0, param
            count += 1
        optimizer = optim.Adam(model.parameters(), lr=c.lr)
        for _ in range(5):
            mel_out, linear_out, align, stop_tokens = model.forward(
                input_dummy, input_lengths, mel_spec, mel_lengths, speaker_ids)
            optimizer.zero_grad()
            loss = criterion(mel_out, mel_spec, mel_lengths)
            stop_loss = criterion_st(stop_tokens, stop_targets)
            loss = loss + criterion(linear_out, linear_spec,
                                    mel_lengths) + stop_loss
            loss.backward()
            optimizer.step()
        # check parameter changes
        count = 0
        for param, param_ref in zip(model.parameters(),
                                    model_ref.parameters()):
            # ignore pre-higway layer since it works conditional
            # if count not in [145, 59]:
            assert (param != param_ref).any(
            ), "param {} with shape {} not updated!! \n{}\n{}".format(
                count, param.shape, param, param_ref)
            count += 1
コード例 #2
0
ファイル: test_tacotron_model.py プロジェクト: gerazov/TTS
    def test_train_step():
        config = config_global.copy()
        config.use_d_vector_file = True

        config.use_gst = True
        config.gst = GSTConfig()

        input_dummy = torch.randint(0, 24, (8, 128)).long().to(device)
        input_lengths = torch.randint(100, 129, (8,)).long().to(device)
        input_lengths[-1] = 128
        mel_spec = torch.rand(8, 30, config.audio["num_mels"]).to(device)
        linear_spec = torch.rand(8, 30, config.audio["fft_size"] // 2 + 1).to(device)
        mel_lengths = torch.randint(20, 30, (8,)).long().to(device)
        mel_lengths[-1] = mel_spec.size(1)
        stop_targets = torch.zeros(8, 30, 1).float().to(device)
        speaker_embeddings = torch.rand(8, 55).to(device)

        for idx in mel_lengths:
            stop_targets[:, int(idx.item()) :, 0] = 1.0

        stop_targets = stop_targets.view(input_dummy.shape[0], stop_targets.size(1) // config.r, -1)
        stop_targets = (stop_targets.sum(2) > 0.0).unsqueeze(2).float().squeeze()

        criterion = L1LossMasked(seq_len_norm=False).to(device)
        criterion_st = nn.BCEWithLogitsLoss().to(device)
        config.d_vector_dim = 55
        model = Tacotron(config).to(device)  # FIXME: missing num_speakers parameter to Tacotron ctor
        model.train()
        print(" > Num parameters for Tacotron model:%s" % (count_parameters(model)))
        model_ref = copy.deepcopy(model)
        count = 0
        for param, param_ref in zip(model.parameters(), model_ref.parameters()):
            assert (param - param_ref).sum() == 0, param
            count += 1
        optimizer = optim.Adam(model.parameters(), lr=config.lr)
        for _ in range(5):
            outputs = model.forward(
                input_dummy, input_lengths, mel_spec, mel_lengths, aux_input={"d_vectors": speaker_embeddings}
            )
            optimizer.zero_grad()
            loss = criterion(outputs["decoder_outputs"], mel_spec, mel_lengths)
            stop_loss = criterion_st(outputs["stop_tokens"], stop_targets)
            loss = loss + criterion(outputs["model_outputs"], linear_spec, mel_lengths) + stop_loss
            loss.backward()
            optimizer.step()
        # check parameter changes
        count = 0
        for name_param, param_ref in zip(model.named_parameters(), model_ref.parameters()):
            # ignore pre-higway layer since it works conditional
            # if count not in [145, 59]:
            name, param = name_param
            if name == "gst_layer.encoder.recurrence.weight_hh_l0":
                continue
            assert (param != param_ref).any(), "param {} with shape {} not updated!! \n{}\n{}".format(
                count, param.shape, param, param_ref
            )
            count += 1
コード例 #3
0
ファイル: test_tacotron_model.py プロジェクト: stjordanis/TTS
    def test_train_step():
        config = config_global.copy()
        config.use_speaker_embedding = True
        config.num_speakers = 10
        config.use_gst = True
        config.gst = GSTConfig()
        # with random gst mel style
        input_dummy = torch.randint(0, 24, (8, 128)).long().to(device)
        input_lengths = torch.randint(100, 129, (8, )).long().to(device)
        input_lengths[-1] = 128
        mel_spec = torch.rand(8, 120, config.audio["num_mels"]).to(device)
        linear_spec = torch.rand(8, 120,
                                 config.audio["fft_size"] // 2 + 1).to(device)
        mel_lengths = torch.randint(20, 120, (8, )).long().to(device)
        mel_lengths[-1] = 120
        stop_targets = torch.zeros(8, 120, 1).float().to(device)
        speaker_ids = torch.randint(0, 5, (8, )).long().to(device)

        for idx in mel_lengths:
            stop_targets[:, int(idx.item()):, 0] = 1.0

        stop_targets = stop_targets.view(input_dummy.shape[0],
                                         stop_targets.size(1) // config.r, -1)
        stop_targets = (stop_targets.sum(2) >
                        0.0).unsqueeze(2).float().squeeze()

        criterion = L1LossMasked(seq_len_norm=False).to(device)
        criterion_st = nn.BCEWithLogitsLoss().to(device)
        config.use_gst = True
        config.gst = GSTConfig()
        model = Tacotron(config).to(
            device)  # FIXME: missing num_speakers parameter to Tacotron ctor
        model.train()
        # print(model)
        print(" > Num parameters for Tacotron GST model:%s" %
              (count_parameters(model)))
        model_ref = copy.deepcopy(model)
        count = 0
        for param, param_ref in zip(model.parameters(),
                                    model_ref.parameters()):
            assert (param - param_ref).sum() == 0, param
            count += 1
        optimizer = optim.Adam(model.parameters(), lr=config.lr)
        for _ in range(10):
            outputs = model.forward(input_dummy,
                                    input_lengths,
                                    mel_spec,
                                    mel_lengths,
                                    aux_input={"speaker_ids": speaker_ids})
            optimizer.zero_grad()
            loss = criterion(outputs["decoder_outputs"], mel_spec, mel_lengths)
            stop_loss = criterion_st(outputs["stop_tokens"], stop_targets)
            loss = loss + criterion(outputs["model_outputs"], linear_spec,
                                    mel_lengths) + stop_loss
            loss.backward()
            optimizer.step()
        # check parameter changes
        count = 0
        for param, param_ref in zip(model.parameters(),
                                    model_ref.parameters()):
            # ignore pre-higway layer since it works conditional
            assert (param != param_ref).any(
            ), "param {} with shape {} not updated!! \n{}\n{}".format(
                count, param.shape, param, param_ref)
            count += 1

        # with file gst style
        mel_spec = (torch.FloatTensor(ap.melspectrogram(
            ap.load_wav(WAV_FILE)))[:, :120].unsqueeze(0).transpose(
                1, 2).to(device))
        mel_spec = mel_spec.repeat(8, 1, 1)

        input_dummy = torch.randint(0, 24, (8, 128)).long().to(device)
        input_lengths = torch.randint(100, 129, (8, )).long().to(device)
        input_lengths[-1] = 128
        linear_spec = torch.rand(8, mel_spec.size(1),
                                 config.audio["fft_size"] // 2 + 1).to(device)
        mel_lengths = torch.randint(20, mel_spec.size(1),
                                    (8, )).long().to(device)
        mel_lengths[-1] = mel_spec.size(1)
        stop_targets = torch.zeros(8, mel_spec.size(1), 1).float().to(device)
        speaker_ids = torch.randint(0, 5, (8, )).long().to(device)

        for idx in mel_lengths:
            stop_targets[:, int(idx.item()):, 0] = 1.0

        stop_targets = stop_targets.view(input_dummy.shape[0],
                                         stop_targets.size(1) // config.r, -1)
        stop_targets = (stop_targets.sum(2) >
                        0.0).unsqueeze(2).float().squeeze()

        criterion = L1LossMasked(seq_len_norm=False).to(device)
        criterion_st = nn.BCEWithLogitsLoss().to(device)
        model = Tacotron(config).to(
            device)  # FIXME: missing num_speakers parameter to Tacotron ctor
        model.train()
        # print(model)
        print(" > Num parameters for Tacotron GST model:%s" %
              (count_parameters(model)))
        model_ref = copy.deepcopy(model)
        count = 0
        for param, param_ref in zip(model.parameters(),
                                    model_ref.parameters()):
            assert (param - param_ref).sum() == 0, param
            count += 1
        optimizer = optim.Adam(model.parameters(), lr=config.lr)
        for _ in range(10):
            outputs = model.forward(input_dummy,
                                    input_lengths,
                                    mel_spec,
                                    mel_lengths,
                                    aux_input={"speaker_ids": speaker_ids})
            optimizer.zero_grad()
            loss = criterion(outputs["decoder_outputs"], mel_spec, mel_lengths)
            stop_loss = criterion_st(outputs["stop_tokens"], stop_targets)
            loss = loss + criterion(outputs["model_outputs"], linear_spec,
                                    mel_lengths) + stop_loss
            loss.backward()
            optimizer.step()
        # check parameter changes
        count = 0
        for param, param_ref in zip(model.parameters(),
                                    model_ref.parameters()):
            # ignore pre-higway layer since it works conditional
            assert (param != param_ref).any(
            ), "param {} with shape {} not updated!! \n{}\n{}".format(
                count, param.shape, param, param_ref)
            count += 1
コード例 #4
0
    def test_train_step():
        config = TacotronConfig(
            num_chars=32,
            num_speakers=10,
            use_speaker_embedding=True,
            out_channels=513,
            decoder_output_dim=80,
            use_capacitron_vae=True,
            capacitron_vae=CapacitronVAEConfig(),
            optimizer="CapacitronOptimizer",
            optimizer_params={
                "RAdam": {
                    "betas": [0.9, 0.998],
                    "weight_decay": 1e-6
                },
                "SGD": {
                    "lr": 1e-5,
                    "momentum": 0.9
                },
            },
        )

        batch = dict({})
        batch["text_input"] = torch.randint(0, 24, (8, 128)).long().to(device)
        batch["text_lengths"] = torch.randint(100, 129,
                                              (8, )).long().to(device)
        batch["text_lengths"] = torch.sort(batch["text_lengths"],
                                           descending=True)[0]
        batch["text_lengths"][0] = 128
        batch["linear_input"] = torch.rand(
            8, 120, config.audio["fft_size"] // 2 + 1).to(device)
        batch["mel_input"] = torch.rand(8, 120,
                                        config.audio["num_mels"]).to(device)
        batch["mel_lengths"] = torch.randint(20, 120, (8, )).long().to(device)
        batch["mel_lengths"] = torch.sort(batch["mel_lengths"],
                                          descending=True)[0]
        batch["mel_lengths"][0] = 120
        batch["stop_targets"] = torch.zeros(8, 120, 1).float().to(device)
        batch["stop_target_lengths"] = torch.randint(0, 120, (8, )).to(device)
        batch["speaker_ids"] = torch.randint(0, 5, (8, )).long().to(device)
        batch["d_vectors"] = None

        for idx in batch["mel_lengths"]:
            batch["stop_targets"][:, int(idx.item()):, 0] = 1.0

        batch["stop_targets"] = batch["stop_targets"].view(
            batch["text_input"].shape[0],
            batch["stop_targets"].size(1) // config.r, -1)
        batch["stop_targets"] = (batch["stop_targets"].sum(2) >
                                 0.0).unsqueeze(2).float().squeeze()

        model = Tacotron(config).to(device)
        criterion = model.get_criterion()
        optimizer = model.get_optimizer()
        model.train()
        print(" > Num parameters for Tacotron with Capacitron VAE model:%s" %
              (count_parameters(model)))
        model_ref = copy.deepcopy(model)
        count = 0
        for param, param_ref in zip(model.parameters(),
                                    model_ref.parameters()):
            assert (param - param_ref).sum() == 0, param
            count += 1
        for _ in range(10):
            _, loss_dict = model.train_step(batch, criterion)
            optimizer.zero_grad()
            loss_dict["capacitron_vae_beta_loss"].backward()
            optimizer.first_step()
            loss_dict["loss"].backward()
            optimizer.step()
        # check parameter changes
        count = 0
        for param, param_ref in zip(model.parameters(),
                                    model_ref.parameters()):
            # ignore pre-higway layer since it works conditional
            assert (param != param_ref).any(
            ), "param {} with shape {} not updated!! \n{}\n{}".format(
                count, param.shape, param, param_ref)
            count += 1
コード例 #5
0
    scheduler_after_epoch=False,  # scheduler doesn't work without this flag
    # Need to experiment with these below for capacitron
    loss_masking=False,
    decoder_loss_alpha=1.0,
    postnet_loss_alpha=1.0,
    postnet_diff_spec_alpha=0.0,
    decoder_diff_spec_alpha=0.0,
    decoder_ssim_alpha=0.0,
    postnet_ssim_alpha=0.0,
)

ap = AudioProcessor(**config.audio.to_dict())

tokenizer, config = TTSTokenizer.init_from_config(config)

train_samples, eval_samples = load_tts_samples(dataset_config, eval_split=True)

model = Tacotron(config, ap, tokenizer, speaker_manager=None)

trainer = Trainer(
    TrainerArgs(),
    config,
    output_path,
    model=model,
    train_samples=train_samples,
    eval_samples=eval_samples,
)

# 🚀
trainer.fit()
コード例 #6
0
    output_path=output_path,
    datasets=[dataset_config],
    use_speaker_embedding=True,  # set this to enable multi-sepeaker training
)

# init audio processor
ap = AudioProcessor(**config.audio.to_dict())

# load training samples
train_samples, eval_samples = load_tts_samples(dataset_config, eval_split=True)

# init speaker manager for multi-speaker training
# it mainly handles speaker-id to speaker-name for the model and the data-loader
speaker_manager = SpeakerManager()
speaker_manager.set_speaker_ids_from_data(train_samples + eval_samples)

# init model
model = Tacotron(config, speaker_manager)

# init the trainer and 🚀
trainer = Trainer(
    TrainingArgs(),
    config,
    output_path,
    model=model,
    train_samples=train_samples,
    eval_samples=eval_samples,
    training_assets={"audio_processor": ap},
)
trainer.fit()