def train(run_id: str, metadata_fpath: str, models_dir: str, save_every: int,
backup_every: int, force_restart:bool, hparams):
models_dir = Path(models_dir)
models_dir.mkdir(exist_ok=True)
model_dir = models_dir.joinpath(run_id)
plot_dir = model_dir.joinpath("plots")
wav_dir = model_dir.joinpath("wavs")
mel_output_dir = model_dir.joinpath("mel-spectrograms")
meta_folder = model_dir.joinpath("metas")
model_dir.mkdir(exist_ok=True)
plot_dir.mkdir(exist_ok=True)
wav_dir.mkdir(exist_ok=True)
mel_output_dir.mkdir(exist_ok=True)
meta_folder.mkdir(exist_ok=True)
weights_fpath = model_dir.joinpath(run_id).with_suffix(".pt")
print("Checkpoint path: {}".format(weights_fpath))
print("Loading training data from: {}".format(metadata_fpath))
print("Using model: Tacotron")
# return
# Book keeping
step = 0
time_window = ValueWindow(100)
loss_window = ValueWindow(100)
# From WaveRNN/train_tacotron.py
if torch.cuda.is_available():
device = torch.device("cuda")
for session in hparams.tts_schedule:
_, _, _, batch_size = session
if batch_size % torch.cuda.device_count() != 0:
raise ValueError("`batch_size` must be evenly divisible by n_gpus!")
else:
device = torch.device("cpu")
print("Using device:", device)
# Instantiate Tacotron Model
print("\nInitialising Tacotron Model...\n")
model = Tacotron(embed_dims=hparams.tts_embed_dims,
num_chars=len(symbols),
encoder_dims=hparams.tts_encoder_dims,
decoder_dims=hparams.tts_decoder_dims,
n_mels=hparams.num_mels,
fft_bins=hparams.num_mels,
postnet_dims=hparams.tts_postnet_dims,
encoder_K=hparams.tts_encoder_K,
lstm_dims=hparams.tts_lstm_dims,
postnet_K=hparams.tts_postnet_K,
num_highways=hparams.tts_num_highways,
dropout=hparams.tts_dropout,
stop_threshold=hparams.tts_stop_threshold,
speaker_embedding_size=hparams.speaker_embedding_size).to(device)
# Initialize the optimizer
optimizer = optim.Adam(model.parameters())
# Load the weights
if force_restart or not weights_fpath.exists():
print("\nStarting the training of Tacotron from scratch\n")
model.save(weights_fpath)
# Embeddings metadata
char_embedding_fpath = meta_folder.joinpath("CharacterEmbeddings.tsv")
with open(char_embedding_fpath, "w", encoding="utf-8") as f:
for symbol in symbols:
if symbol == " ":
symbol = "\\s" # For visual purposes, swap space with \s
f.write("{}\n".format(symbol))
else:
print("\nLoading weights at %s" % weights_fpath)
model.load(weights_fpath, optimizer)
print("Tacotron weights loaded from step %d" % model.step)
# Initialize the dataset
dataset = SynthesizerDataset(metadata_fpath, hparams)
# test_loader = DataLoader(dataset,
# batch_size=1,
# shuffle=True,
# pin_memory=True)
for i, session in enumerate(hparams.tts_schedule):
current_step = model.get_step()
r, lr, max_step, batch_size = session
training_steps = max_step - current_step
# Do we need to change to the next session?
if current_step >= max_step:
# Are there no further sessions than the current one?
if i == len(hparams.tts_schedule) - 1:
# We have completed training. Save the model and exit
model.save(weights_fpath, optimizer)
break
else:
# There is a following session, go to it
continue
model.r = r
# Begin the training
simple_table([(f"Steps with r={r}", str(training_steps // 1000) + "k Steps"),
("Batch Size", batch_size),
("Learning Rate", lr),
("Outputs/Step (r)", model.r)])
for p in optimizer.param_groups:
p["lr"] = lr
data_loader = DataLoader(dataset,
collate_fn=lambda batch: collate_synthesizer(batch, r, hparams),
batch_size=batch_size,
num_workers=2,
shuffle=True,
pin_memory=True)
total_iters = len(dataset)
steps_per_epoch = np.ceil(total_iters / batch_size).astype(np.int32)
epochs = np.ceil(training_steps / steps_per_epoch).astype(np.int32)
for epoch in range(1, epochs+1):
for i, (texts, mels, embeds, idx) in enumerate(data_loader, 1):
start_time = time.time()
start = time.perf_counter()
# Generate stop tokens for training
stop = torch.ones(mels.shape[0], mels.shape[2])
for j, k in enumerate(idx):
stop[j, :int(dataset.metadata[k][3])-1] = 0
texts = texts.to(device)
mels = mels.to(device)
embeds = embeds.to(device)
stop = stop.to(device)
# print('texts', texts.shape)
# print(mels.shape)
# print(embeds.shape)
# print(stop.shape)
# Forward pass
# Parallelize model onto GPUS using workaround due to python bug
if device.type == "cuda" and torch.cuda.device_count() > 1:
m1_hat, m2_hat, attention, stop_pred = data_parallel_workaround(model, texts,
mels, embeds)
else:
m1_hat, m2_hat, attention, stop_pred = model(texts, mels, embeds)
# Backward pass
m1_loss = F.mse_loss(m1_hat, mels) + F.l1_loss(m1_hat, mels)
m2_loss = F.mse_loss(m2_hat, mels)
stop_loss = F.binary_cross_entropy(stop_pred, stop)
loss = m1_loss + m2_loss + stop_loss
optimizer.zero_grad()
loss.backward()
# if hparams.tts_clip_grad_norm is not None:
grad_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), hparams.tts_clip_grad_norm)
if np.isnan(grad_norm.cpu()):
print("grad_norm was NaN!")
optimizer.step()
time_window.append(time.time() - start_time)
loss_window.append(loss.item())
step = model.get_step()
k = step // 1000
msg = f"| Epoch: {epoch}/{epochs} ({i}/{steps_per_epoch}) | Loss: {loss_window.average:#.4} | {1./time_window.average:#.2} steps/s | Step: {k}k | "
stream(msg)
if step % 10 == 0:
good_logger.log_training(reduced_loss=loss.item(),
reduced_mel_loss=loss.item() - stop_loss.item(),
reduced_gate_loss=stop_loss.item(),
grad_norm=grad_norm,
learning_rate=optimizer.param_groups[0]['lr'],
duration=time.perf_counter() - start,
iteration=step)
# Backup or save model as appropriate
if backup_every != 0 and step % backup_every == 0 :
backup_fpath = Path("{}/{}_{}k.pt".format(str(weights_fpath.parent), run_id, k))
model.save(backup_fpath, optimizer)
if save_every != 0 and step % save_every == 0 :
# Must save latest optimizer state to ensure that resuming training
# doesn't produce artifacts
model.save(weights_fpath, optimizer)
# Evaluate model to generate samples
epoch_eval = hparams.tts_eval_interval == -1 and i == steps_per_epoch # If epoch is done
step_eval = hparams.tts_eval_interval > 0 and step % hparams.tts_eval_interval == 0 # Every N steps
if epoch_eval or step_eval:
for sample_idx in range(hparams.tts_eval_num_samples):
# At most, generate samples equal to number in the batch
if sample_idx + 1 <= len(texts):
# Remove padding from mels using frame length in metadata
mel_length = int(dataset.metadata[idx[sample_idx]][3])
mel_prediction = np_now(m2_hat[sample_idx]).T[:mel_length]
target_spectrogram = np_now(mels[sample_idx]).T[:mel_length]
attention_len = mel_length // model.r
eval_model(attention=np_now(attention[sample_idx][:, :attention_len]),
mel_prediction=mel_prediction,
target_spectrogram=target_spectrogram,
input_seq=np_now(texts[sample_idx]),
step=step,
plot_dir=plot_dir,
mel_output_dir=mel_output_dir,
wav_dir=wav_dir,
sample_num=sample_idx + 1,
loss=loss,
hparams=hparams)
# Break out of loop to update training schedule
if step >= max_step:
break
# Add line break after every epoch
print("")
def run_synthesis(in_dir, out_dir, model_dir, hparams):
# This generates ground truth-aligned mels for vocoder training
synth_dir = Path(out_dir).joinpath("mels_gta")
synth_dir.mkdir(exist_ok=True)
print(hparams_debug_string(hparams))
# Check for GPU
if torch.cuda.is_available():
device = torch.device("cuda")
if hparams.synthesis_batch_size % torch.cuda.device_count() != 0:
raise ValueError("`hparams.synthesis_batch_size` must be evenly divisible by n_gpus!")
else:
device = torch.device("cpu")
print("Synthesizer using device:", device)
# Instantiate Tacotron model
model = Tacotron(embed_dims=hparams.tts_embed_dims,
num_chars=len(symbols),
encoder_dims=hparams.tts_encoder_dims,
decoder_dims=hparams.tts_decoder_dims,
n_mels=hparams.num_mels,
fft_bins=hparams.num_mels,
postnet_dims=hparams.tts_postnet_dims,
encoder_K=hparams.tts_encoder_K,
lstm_dims=hparams.tts_lstm_dims,
postnet_K=hparams.tts_postnet_K,
num_highways=hparams.tts_num_highways,
dropout=0., # Use zero dropout for gta mels
stop_threshold=hparams.tts_stop_threshold,
speaker_embedding_size=hparams.speaker_embedding_size).to(device)
# Load the weights
model_dir = Path(model_dir)
model_fpath = model_dir.joinpath(model_dir.stem).with_suffix(".pt")
print("\nLoading weights at %s" % model_fpath)
model.load(model_fpath)
print("Tacotron weights loaded from step %d" % model.step)
# Synthesize using same reduction factor as the model is currently trained
r = np.int32(model.r)
# Set model to eval mode (disable gradient and zoneout)
model.eval()
# Initialize the dataset
in_dir = Path(in_dir)
metadata_fpath = in_dir.joinpath("train.txt")
mel_dir = in_dir.joinpath("mels")
embed_dir = in_dir.joinpath("embeds")
dataset = SynthesizerDataset(metadata_fpath, mel_dir, embed_dir, hparams)
data_loader = DataLoader(dataset,
collate_fn=lambda batch: collate_synthesizer(batch, r,hparams),
batch_size=hparams.synthesis_batch_size,
num_workers=0, #Having an error(Can't pickle local object 'run_synthesis.<locals>.<lambda>') when training in Windows unless you set num_workers=0
shuffle=False,
pin_memory=True)
# Generate GTA mels
meta_out_fpath = Path(out_dir).joinpath("synthesized.txt")
with open(meta_out_fpath, "w") as file:
for i, (texts, mels, embeds, idx) in tqdm(enumerate(data_loader), total=len(data_loader)):
texts = texts.to(device)
mels = mels.to(device)
embeds = embeds.to(device)
# Parallelize model onto GPUS using workaround due to python bug
if device.type == "cuda" and torch.cuda.device_count() > 1:
_, mels_out,_,_ = data_parallel_workaround(model, texts, mels, embeds)
else:
_,mels_out, _,_ = model(texts, mels, embeds)
for j, k in enumerate(idx):
# Note: outputs mel-spectrogram files and target ones have same names, just different folders
mel_filename = Path(synth_dir).joinpath(dataset.metadata[k][1])
mel_out = mels_out[j].detach().cpu().numpy().T
# Use the length of the ground truth mel to remove padding from the generated mels
mel_out = mel_out[:int(dataset.metadata[k][4])]
# Write the spectrogram to disk
np.save(mel_filename, mel_out, allow_pickle=False)
# Write metadata into the synthesized file
file.write("|".join(dataset.metadata[k]))
class Synthesizer:
sample_rate = hparams.sample_rate
hparams = hparams
def __init__(self, model_fpath: Path, verbose=True):
"""
The model isn't instantiated and loaded in memory until needed or until load() is called.
:param model_fpath: path to the trained model file
:param verbose: if False, prints less information when using the model
"""
self.model_fpath = model_fpath
self.verbose = verbose
# Check for GPU
if torch.cuda.is_available():
self.device = torch.device("cuda")
else:
self.device = torch.device("cpu")
if self.verbose:
print("Synthesizer using device:", self.device)
# Tacotron model will be instantiated later on first use.
self._model = None
def is_loaded(self):
"""
Whether the model is loaded in memory.
"""
return self._model is not None
def load(self):
"""
Instantiates and loads the model given the weights file that was passed in the constructor.
"""
self._model = Tacotron(
embed_dims=hparams.tts_embed_dims,
num_chars=len(symbols),
encoder_dims=hparams.tts_encoder_dims,
decoder_dims=hparams.tts_decoder_dims,
n_mels=hparams.num_mels,
fft_bins=hparams.num_mels,
postnet_dims=hparams.tts_postnet_dims,
encoder_K=hparams.tts_encoder_K,
lstm_dims=hparams.tts_lstm_dims,
postnet_K=hparams.tts_postnet_K,
num_highways=hparams.tts_num_highways,
dropout=hparams.tts_dropout,
stop_threshold=hparams.tts_stop_threshold,
speaker_embedding_size=hparams.speaker_embedding_size).to(
self.device)
self._model.load(self.model_fpath)
self._model.eval()
if self.verbose:
print("Loaded synthesizer \"%s\" trained to step %d" %
(self.model_fpath.name, self._model.state_dict()["step"]))
def synthesize_spectrograms(self,
texts: List[str],
embeddings: Union[np.ndarray,
List[np.ndarray]],
return_alignments=False):
"""
Synthesizes mel spectrograms from texts and speaker embeddings.
:param texts: a list of N text prompts to be synthesized
:param embeddings: a numpy array or list of speaker embeddings of shape (N, 256)
:param return_alignments: if True, a matrix representing the alignments between the
characters
and each decoder output step will be returned for each spectrogram
:return: a list of N melspectrograms as numpy arrays of shape (80, Mi), where Mi is the
sequence length of spectrogram i, and possibly the alignments.
"""
# Load the model on the first request.
if not self.is_loaded():
self.load()
# Print some info about the model when it is loaded
tts_k = self._model.get_step() // 1000
simple_table([("Tacotron", str(tts_k) + "k"),
("r", self._model.r)])
# Preprocess text inputs
inputs = [
text_to_sequence(text.strip(), hparams.tts_cleaner_names)
for text in texts
]
if not isinstance(embeddings, list):
embeddings = [embeddings]
# Batch inputs
batched_inputs = [
inputs[i:i + hparams.synthesis_batch_size]
for i in range(0, len(inputs), hparams.synthesis_batch_size)
]
batched_embeds = [
embeddings[i:i + hparams.synthesis_batch_size]
for i in range(0, len(embeddings), hparams.synthesis_batch_size)
]
specs = []
for i, batch in enumerate(batched_inputs, 1):
if self.verbose:
print(f"\n| Generating {i}/{len(batched_inputs)}")
# Pad texts so they are all the same length
text_lens = [len(text) for text in batch]
max_text_len = max(text_lens)
chars = [pad1d(text, max_text_len) for text in batch]
chars = np.stack(chars)
# Stack speaker embeddings into 2D array for batch processing
speaker_embeds = np.stack(batched_embeds[i - 1])
# Convert to tensor
chars = torch.tensor(chars).long().to(self.device)
speaker_embeddings = torch.tensor(speaker_embeds).float().to(
self.device)
# Inference
_, mels, alignments = self._model.generate(chars,
speaker_embeddings)
mels = mels.detach().cpu().numpy()
for m in mels:
# Trim silence from end of each spectrogram
while np.max(m[:, -1]) < hparams.tts_stop_threshold:
m = m[:, :-1]
specs.append(m)
if self.verbose:
print("\n\nDone.\n")
return (specs, alignments) if return_alignments else specs
@staticmethod
def load_preprocess_wav(fpath):
"""
Loads and preprocesses an audio file under the same conditions the audio files were used to
train the synthesizer.
"""
wav = librosa.load(str(fpath), hparams.sample_rate)[0]
if hparams.rescale:
wav = wav / np.abs(wav).max() * hparams.rescaling_max
return wav
@staticmethod
def make_spectrogram(fpath_or_wav: Union[str, Path, np.ndarray]):
"""
Creates a mel spectrogram from an audio file in the same manner as the mel spectrograms that
were fed to the synthesizer when training.
"""
if isinstance(fpath_or_wav, str) or isinstance(fpath_or_wav, Path):
wav = Synthesizer.load_preprocess_wav(fpath_or_wav)
else:
wav = fpath_or_wav
mel_spectrogram = audio.melspectrogram(wav, hparams).astype(np.float32)
return mel_spectrogram
@staticmethod
def griffin_lim(mel):
"""
Inverts a mel spectrogram using Griffin-Lim. The mel spectrogram is expected to have been built
with the same parameters present in hparams.py.
"""
return audio.inv_mel_spectrogram(mel, hparams)