def eval_model(attention, mel_prediction, target_spectrogram, input_seq, step,
plot_dir, mel_output_dir, wav_dir, sample_num, loss, hparams):
# Save some results for evaluation
attention_path = str(
plot_dir.joinpath("attention_step_{}_sample_{}".format(
step, sample_num)))
save_attention(attention, attention_path)
# save predicted mel spectrogram to disk (debug)
mel_output_fpath = mel_output_dir.joinpath(
"mel-prediction-step-{}_sample_{}.npy".format(step, sample_num))
np.save(str(mel_output_fpath), mel_prediction, allow_pickle=False)
# save griffin lim inverted wav for debug (mel -> wav)
wav = audio.inv_mel_spectrogram(mel_prediction.T, hparams)
wav_fpath = wav_dir.joinpath("step-{}-wave-from-mel_sample_{}.wav".format(
step, sample_num))
audio.save_wav(wav, str(wav_fpath), sr=hparams.sample_rate)
# save real and predicted mel-spectrogram plot to disk (control purposes)
spec_fpath = plot_dir.joinpath(
"step-{}-mel-spectrogram_sample_{}.png".format(step, sample_num))
title_str = "{}, {}, step={}, loss={:.5f}".format("Tacotron",
time_string(), step,
loss)
plot_spectrogram(mel_prediction,
str(spec_fpath),
title=title_str,
target_spectrogram=target_spectrogram,
max_len=target_spectrogram.size // hparams.num_mels)
print("Input at step {}: {}".format(step, sequence_to_text(input_seq)))
def run_mel_strip():
import numpy as np
from tools.spec_processor import find_endpoint, find_silences
from synthesizer.audio import inv_mel_spectrogram, save_wav
from synthesizer.hparams import hparams
from matplotlib import pyplot as plt
inpath = Path(
r'E:\lab\zhrtvc\zhrtvc\toolbox\saved_files\mels\wavs-P00173I-001_20170001P00173I0068.wav_1567509749_我家朵朵是世界上最漂亮的朵朵。。知道自己是什么样的人。要做什么。无需活在别人非议或期待里。你勤奋.npy'
)
data = np.load(inpath)
data = data.T
print(data.shape)
end_idx = find_silences(data, min_silence_sec=0.5, hop_silence_sec=0.2)
print(end_idx, len(data))
out_dir = Path(r'data/syns')
for i, pair in enumerate(zip(end_idx[:-1], end_idx[1:]), 1):
a, b = pair
wav = inv_mel_spectrogram(data[a[-1]:b[0]].T, hparams)
save_wav(wav, out_dir.joinpath(f'sil-{i:02d}.wav'),
hparams.sample_rate)
plt.imshow(data.T)
plt.colorbar()
plt.show()
def train(log_dir, args, hparams):
save_dir = os.path.join(log_dir, "taco_pretrained")
plot_dir = os.path.join(log_dir, "plots")
wav_dir = os.path.join(log_dir, "wavs")
mel_dir = os.path.join(log_dir, "mel-spectrograms")
eval_dir = os.path.join(log_dir, "eval-dir")
eval_plot_dir = os.path.join(eval_dir, "plots")
eval_wav_dir = os.path.join(eval_dir, "wavs")
tensorboard_dir = os.path.join(log_dir, "tacotron_events")
meta_folder = os.path.join(log_dir, "metas")
os.makedirs(save_dir, exist_ok=True)
os.makedirs(plot_dir, exist_ok=True)
os.makedirs(wav_dir, exist_ok=True)
os.makedirs(mel_dir, exist_ok=True)
os.makedirs(eval_dir, exist_ok=True)
os.makedirs(eval_plot_dir, exist_ok=True)
os.makedirs(eval_wav_dir, exist_ok=True)
os.makedirs(tensorboard_dir, exist_ok=True)
os.makedirs(meta_folder, exist_ok=True)
checkpoint_fpath = os.path.join(save_dir, "tacotron_model.ckpt")
if hparams.if_use_speaker_classifier:
metadat_fpath = os.path.join(args.synthesizer_root,
"train_augment_speaker.txt")
else:
metadat_fpath = os.path.join(args.synthesizer_root, "train.txt")
log("Checkpoint path: {}".format(checkpoint_fpath))
log("Loading training data from: {}".format(metadat_fpath))
log("Using model: Tacotron")
log(hparams_debug_string())
# Start by setting a seed for repeatability
tf.set_random_seed(hparams.tacotron_random_seed)
# Set up data feeder
coord = tf.train.Coordinator()
with tf.variable_scope("datafeeder") as scope:
feeder = Feeder(coord, metadat_fpath, hparams)
# Set up model:
global_step = tf.Variable(0, name="global_step", trainable=False)
model, stats = model_train_mode(args, feeder, hparams, global_step)
eval_model = model_test_mode(args, feeder, hparams, global_step)
# Embeddings metadata
char_embedding_meta = os.path.join(meta_folder, "CharacterEmbeddings.tsv")
if not os.path.isfile(char_embedding_meta):
with open(char_embedding_meta, "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))
char_embedding_meta = char_embedding_meta.replace(log_dir, "..")
# Book keeping
step = 0
time_window = ValueWindow(100)
loss_window = ValueWindow(100)
saver = tf.train.Saver(max_to_keep=5)
log("Tacotron training set to a maximum of {} steps".format(
args.tacotron_train_steps))
# Memory allocation on the GPU as needed
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.allow_soft_placement = True
# Train
with tf.Session(config=config) as sess:
try:
summary_writer = tf.summary.FileWriter(tensorboard_dir, sess.graph)
sess.run(tf.global_variables_initializer())
# saved model restoring
if args.restore:
# Restore saved model if the user requested it, default = True
try:
checkpoint_state = tf.train.get_checkpoint_state(save_dir)
if checkpoint_state and checkpoint_state.model_checkpoint_path:
log("Loading checkpoint {}".format(
checkpoint_state.model_checkpoint_path),
slack=True)
saver.restore(sess,
checkpoint_state.model_checkpoint_path)
else:
log("No model to load at {}".format(save_dir),
slack=True)
saver.save(sess,
checkpoint_fpath,
global_step=global_step)
except tf.errors.OutOfRangeError as e:
log("Cannot restore checkpoint: {}".format(e), slack=True)
else:
log("Starting new training!", slack=True)
saver.save(sess, checkpoint_fpath, global_step=global_step)
# initializing feeder
feeder.start_threads(sess)
# Training loop
while not coord.should_stop() and step < args.tacotron_train_steps:
start_time = time.time()
step, loss, adversial_loss, opt = sess.run([
global_step, model.loss, model.adversial_loss,
model.optimize
])
loss -= adversial_loss
time_window.append(time.time() - start_time)
loss_window.append(loss)
message = "Step {:7d} [{:.3f} sec/step, loss={:.5f}, avg_loss={:.5f}, adv_loss={:.5f}]".format(
step, time_window.average, loss, loss_window.average,
adversial_loss)
log(message,
end="\r",
slack=(step % args.checkpoint_interval == 0))
print(message)
if loss > 100 or np.isnan(loss):
log("Loss exploded to {:.5f} at step {}".format(
loss, step))
raise Exception("Loss exploded")
if step % args.summary_interval == 0:
log("\nWriting summary at step {}".format(step))
summary_writer.add_summary(sess.run(stats), step)
if step % args.eval_interval == 0:
# Run eval and save eval stats
log("\nRunning evaluation at step {}".format(step))
eval_losses = []
before_losses = []
after_losses = []
stop_token_losses = []
linear_losses = []
linear_loss = None
adversial_losses = []
if hparams.predict_linear:
for i in tqdm(range(feeder.test_steps)):
eloss, before_loss, after_loss, stop_token_loss, linear_loss, mel_p, \
mel_t, t_len, align, lin_p, lin_t = sess.run(
[
eval_model.tower_loss[0], eval_model.tower_before_loss[0],
eval_model.tower_after_loss[0],
eval_model.tower_stop_token_loss[0],
eval_model.tower_linear_loss[0],
eval_model.tower_mel_outputs[0][0],
eval_model.tower_mel_targets[0][0],
eval_model.tower_targets_lengths[0][0],
eval_model.tower_alignments[0][0],
eval_model.tower_linear_outputs[0][0],
eval_model.tower_linear_targets[0][0],
])
eval_losses.append(eloss)
before_losses.append(before_loss)
after_losses.append(after_loss)
stop_token_losses.append(stop_token_loss)
linear_losses.append(linear_loss)
linear_loss = sum(linear_losses) / len(linear_losses)
wav = audio.inv_linear_spectrogram(lin_p.T, hparams)
audio.save_wav(
wav,
os.path.join(
eval_wav_dir,
"step-{}-eval-wave-from-linear.wav".format(
step)),
sr=hparams.sample_rate)
else:
for i in tqdm(range(feeder.test_steps)):
eloss, before_loss, after_loss, stop_token_loss, adversial_loss, mel_p, mel_t, t_len,\
align = sess.run(
[
eval_model.tower_loss[0], eval_model.tower_before_loss[0],
eval_model.tower_after_loss[0],
eval_model.tower_stop_token_loss[0],
eval_model.tower_adversial_loss[0],
eval_model.tower_mel_outputs[0][0],
eval_model.tower_mel_targets[0][0],
eval_model.tower_targets_lengths[0][0],
eval_model.tower_alignments[0][0]
])
eval_losses.append(eloss)
before_losses.append(before_loss)
after_losses.append(after_loss)
stop_token_losses.append(stop_token_loss)
adversial_losses.append(adversial_loss)
eval_loss = sum(eval_losses) / len(eval_losses)
before_loss = sum(before_losses) / len(before_losses)
after_loss = sum(after_losses) / len(after_losses)
stop_token_loss = sum(stop_token_losses) / len(
stop_token_losses)
adversial_loss = sum(adversial_losses) / len(
adversial_losses)
log("Saving eval log to {}..".format(eval_dir))
# Save some log to monitor model improvement on same unseen sequence
wav = audio.inv_mel_spectrogram(mel_p.T, hparams)
audio.save_wav(
wav,
os.path.join(
eval_wav_dir,
"step-{}-eval-wave-from-mel.wav".format(step)),
sr=hparams.sample_rate)
plot.plot_alignment(
align,
os.path.join(eval_plot_dir,
"step-{}-eval-align.png".format(step)),
title="{}, {}, step={}, loss={:.5f}".format(
"Tacotron", time_string(), step, eval_loss),
max_len=t_len // hparams.outputs_per_step)
plot.plot_spectrogram(
mel_p,
os.path.join(
eval_plot_dir, "step-{"
"}-eval-mel-spectrogram.png".format(step)),
title="{}, {}, step={}, loss={:.5f}".format(
"Tacotron", time_string(), step, eval_loss),
target_spectrogram=mel_t,
max_len=t_len)
if hparams.predict_linear:
plot.plot_spectrogram(
lin_p,
os.path.join(
eval_plot_dir,
"step-{}-eval-linear-spectrogram.png".format(
step)),
title="{}, {}, step={}, loss={:.5f}".format(
"Tacotron", time_string(), step, eval_loss),
target_spectrogram=lin_t,
max_len=t_len,
auto_aspect=True)
log("Eval loss for global step {}: {:.3f}".format(
step, eval_loss))
log("Writing eval summary!")
add_eval_stats(summary_writer, step, linear_loss,
before_loss, after_loss, stop_token_loss,
adversial_loss, eval_loss)
if step % args.checkpoint_interval == 0 or step == args.tacotron_train_steps or \
step == 300:
# Save model and current global step
saver.save(sess, checkpoint_fpath, global_step=global_step)
log("\nSaving alignment, Mel-Spectrograms and griffin-lim inverted waveform.."
)
input_seq, mel_prediction, alignment, target, target_length = sess.run(
[
model.tower_inputs[0][0],
model.tower_mel_outputs[0][0],
model.tower_alignments[0][0],
model.tower_mel_targets[0][0],
model.tower_targets_lengths[0][0],
])
# save predicted mel spectrogram to disk (debug)
mel_filename = "mel-prediction-step-{}.npy".format(step)
np.save(os.path.join(mel_dir, mel_filename),
mel_prediction.T,
allow_pickle=False)
# save griffin lim inverted wav for debug (mel -> wav)
wav = audio.inv_mel_spectrogram(mel_prediction.T, hparams)
audio.save_wav(
wav,
os.path.join(wav_dir,
"step-{}-wave-from-mel.wav".format(step)),
sr=hparams.sample_rate)
# save alignment plot to disk (control purposes)
plot.plot_alignment(
alignment,
os.path.join(plot_dir,
"step-{}-align.png".format(step)),
title="{}, {}, step={}, loss={:.5f}".format(
"Tacotron", time_string(), step, loss),
max_len=target_length // hparams.outputs_per_step)
# save real and predicted mel-spectrogram plot to disk (control purposes)
plot.plot_spectrogram(
mel_prediction,
os.path.join(
plot_dir,
"step-{}-mel-spectrogram.png".format(step)),
title="{}, {}, step={}, loss={:.5f}".format(
"Tacotron", time_string(), step, loss),
target_spectrogram=target,
max_len=target_length)
#log("Input at step {}: {}".format(step, sequence_to_text(input_seq)))
if step % args.embedding_interval == 0 or step == args.tacotron_train_steps or step == 1:
# Get current checkpoint state
checkpoint_state = tf.train.get_checkpoint_state(save_dir)
# Update Projector
#log("\nSaving Model Character Embeddings visualization..")
#add_embedding_stats(summary_writer, [model.embedding_table.name],
# [char_embedding_meta],
# checkpoint_state.model_checkpoint_path)
#log("Tacotron Character embeddings have been updated on tensorboard!")
log("Tacotron training complete after {} global steps!".format(
args.tacotron_train_steps),
slack=True)
return save_dir
except Exception as e:
log("Exiting due to exception: {}".format(e), slack=True)
traceback.print_exc()
coord.request_stop(e)
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)
def synthesize(self, texts, basenames, out_dir, log_dir, mel_filenames,
embed_filenames):
hparams = self._hparams
cleaner_names = [x.strip() for x in hparams.cleaners.split(",")]
assert 0 == len(texts) % self._hparams.tacotron_num_gpus
seqs = [
np.asarray(text_to_sequence(text, cleaner_names)) for text in texts
]
input_lengths = [len(seq) for seq in seqs]
size_per_device = len(seqs) // self._hparams.tacotron_num_gpus
#Pad inputs according to each GPU max length
input_seqs = None
split_infos = []
for i in range(self._hparams.tacotron_num_gpus):
device_input = seqs[size_per_device * i:size_per_device * (i + 1)]
device_input, max_seq_len = self._prepare_inputs(device_input)
input_seqs = np.concatenate(
(input_seqs, device_input),
axis=1) if input_seqs is not None else device_input
split_infos.append([max_seq_len, 0, 0, 0])
feed_dict = {
self.inputs: input_seqs,
self.input_lengths: np.asarray(input_lengths, dtype=np.int32),
}
if self.gta:
np_targets = [
np.load(mel_filename) for mel_filename in mel_filenames
]
target_lengths = [len(np_target) for np_target in np_targets]
#pad targets according to each GPU max length
target_seqs = None
for i in range(self._hparams.tacotron_num_gpus):
device_target = np_targets[size_per_device *
i:size_per_device * (i + 1)]
device_target, max_target_len = self._prepare_targets(
device_target, self._hparams.outputs_per_step)
target_seqs = np.concatenate(
(target_seqs, device_target),
axis=1) if target_seqs is not None else device_target
split_infos[i][
1] = max_target_len #Not really used but setting it in case for future development maybe?
feed_dict[self.targets] = target_seqs
assert len(np_targets) == len(texts)
feed_dict[self.split_infos] = np.asarray(split_infos, dtype=np.int32)
feed_dict[self.speaker_embeddings] = [
np.load(f) for f in embed_filenames
]
if self.gta or not hparams.predict_linear:
mels, alignments, stop_tokens = self.session.run(
[
self.mel_outputs, self.alignments,
self.stop_token_prediction
],
feed_dict=feed_dict)
#Linearize outputs (1D arrays)
mels = [mel for gpu_mels in mels for mel in gpu_mels]
alignments = [
align for gpu_aligns in alignments for align in gpu_aligns
]
stop_tokens = [
token for gpu_token in stop_tokens for token in gpu_token
]
if not self.gta:
#Natural batch synthesis
#Get Mel lengths for the entire batch from stop_tokens predictions
target_lengths = self._get_output_lengths(stop_tokens)
#Take off the batch wise padding
mels = [
mel[:target_length, :]
for mel, target_length in zip(mels, target_lengths)
]
assert len(mels) == len(texts)
else:
linears, mels, alignments, stop_tokens = self.session.run(
[
self.linear_outputs, self.mel_outputs, self.alignments,
self.stop_token_prediction
],
feed_dict=feed_dict)
#Linearize outputs (1D arrays)
linears = [
linear for gpu_linear in linears for linear in gpu_linear
]
mels = [mel for gpu_mels in mels for mel in gpu_mels]
alignments = [
align for gpu_aligns in alignments for align in gpu_aligns
]
stop_tokens = [
token for gpu_token in stop_tokens for token in gpu_token
]
#Natural batch synthesis
#Get Mel/Linear lengths for the entire batch from stop_tokens predictions
# target_lengths = self._get_output_lengths(stop_tokens)
target_lengths = [9999]
#Take off the batch wise padding
mels = [
mel[:target_length, :]
for mel, target_length in zip(mels, target_lengths)
]
linears = [
linear[:target_length, :]
for linear, target_length in zip(linears, target_lengths)
]
assert len(mels) == len(linears) == len(texts)
if basenames is None:
raise NotImplemented()
saved_mels_paths = []
for i, mel in enumerate(mels):
# Write the spectrogram to disk
# Note: outputs mel-spectrogram files and target ones have same names, just different folders
mel_filename = os.path.join(out_dir,
"mel-{}.npy".format(basenames[i]))
np.save(mel_filename, mel, allow_pickle=False)
saved_mels_paths.append(mel_filename)
if log_dir is not None:
#save wav (mel -> wav)
wav = audio.inv_mel_spectrogram(mel.T, hparams)
audio.save_wav(wav,
os.path.join(
log_dir,
"wavs/wav-{}-mel.wav".format(basenames[i])),
sr=hparams.sample_rate)
#save alignments
plot.plot_alignment(alignments[i],
os.path.join(
log_dir,
"plots/alignment-{}.png".format(
basenames[i])),
title="{}".format(texts[i]),
split_title=True,
max_len=target_lengths[i])
#save mel spectrogram plot
plot.plot_spectrogram(
mel,
os.path.join(log_dir,
"plots/mel-{}.png".format(basenames[i])),
title="{}".format(texts[i]),
split_title=True)
if hparams.predict_linear:
#save wav (linear -> wav)
wav = audio.inv_linear_spectrogram(linears[i].T, hparams)
audio.save_wav(wav,
os.path.join(
log_dir,
"wavs/wav-{}-linear.wav".format(
basenames[i])),
sr=hparams.sample_rate)
#save linear spectrogram plot
plot.plot_spectrogram(linears[i],
os.path.join(
log_dir,
"plots/linear-{}.png".format(
basenames[i])),
title="{}".format(texts[i]),
split_title=True,
auto_aspect=True)
return saved_mels_paths
# The vocoder synthesizes one waveform at a time, but it's more efficient for long ones. We
# can concatenate the mel spectrograms to a single one.
mel = np.concatenate(mels, axis=1)
# The vocoder can take a callback function to display the generation. More on that later. For
# now we'll simply hide it like this:
no_action = lambda *args: None
print("\tTesting the vocoder...")
# For the sake of making this test short, we'll pass a short target length. The target length
# is the length of the wav segments that are processed in parallel. E.g. for audio sampled
# at 16000 Hertz, a target length of 8000 means that the target audio will be cut in chunks of
# 0.5 seconds which will all be generated together. The parameters here are absurdly short, and
# that has a detrimental effect on the quality of the audio. The default parameters are
# recommended in general.
# vocoder.infer_waveform(mel, target=200, overlap=50, progress_callback=no_action)
generated_wav = audio.inv_mel_spectrogram(mel, hparams.hparams)
print("All test passed! You can now synthesize speech.\n\n")
## Interactive speech generation
print(
"This is a GUI-less example of interface to SV2TTS. The purpose of this script is to "
"show how you can interface this project easily with your own. See the source code for "
"an explanation of what is happening.\n")
print("Interactive generation loop")
num_generated = 0
while True:
try:
# Get the reference audio filepath
message = "Reference voice: enter an audio filepath of a voice to be cloned (mp3, " \
"wav, m4a, flac, ...):\n"
def train(log_dir, args, hparams):
save_dir = os.path.join(log_dir, "taco_pretrained")
plot_dir = os.path.join(log_dir, "plots")
wav_dir = os.path.join(log_dir, "wavs")
mel_dir = os.path.join(log_dir, "mel-spectrograms")
eval_dir = os.path.join(log_dir, "eval-dir")
eval_plot_dir = os.path.join(eval_dir, "plots")
eval_wav_dir = os.path.join(eval_dir, "wavs")
tensorboard_dir = os.path.join(log_dir, "tacotron_events")
os.makedirs(save_dir, exist_ok=True)
os.makedirs(plot_dir, exist_ok=True)
os.makedirs(wav_dir, exist_ok=True)
os.makedirs(mel_dir, exist_ok=True)
os.makedirs(eval_dir, exist_ok=True)
os.makedirs(eval_plot_dir, exist_ok=True)
os.makedirs(eval_wav_dir, exist_ok=True)
os.makedirs(tensorboard_dir, exist_ok=True)
checkpoint_fpath = os.path.join(save_dir, "tacotron_model.ckpt")
log("Checkpoint path: {}".format(checkpoint_fpath))
log("Using model: Tacotron")
log(hparams_debug_string())
# Start by setting a seed for repeatability
tf.set_random_seed(hparams.tacotron_random_seed)
# Set up data feeder
coord = tf.train.Coordinator()
with tf.variable_scope("datafeeder") as scope:
feeder = Feeder(coord, hparams)
# Set up model:
global_step = tf.Variable(0, name="global_step", trainable=False)
model, stats = model_train_mode(args, feeder, hparams, global_step)
#eval_model = model_test_mode(args, feeder, hparams, global_step)
# Book keeping
step = 0
time_window = ValueWindow(100)
loss_window = ValueWindow(100)
saver = tf.train.Saver(max_to_keep=2)
log("Tacotron training set to a maximum of {} steps".format(
args.tacotron_train_steps))
# Memory allocation on the GPU as needed
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.allow_soft_placement = True
# Train
with tf.Session(config=config) as sess:
try:
summary_writer = tf.summary.FileWriter(tensorboard_dir, sess.graph)
sess.run(tf.global_variables_initializer())
# saved model restoring
if args.restore:
# Restore saved model if the user requested it, default = True
try:
checkpoint_state = tf.train.get_checkpoint_state(save_dir)
if checkpoint_state and checkpoint_state.model_checkpoint_path:
log("Loading checkpoint {}".format(
checkpoint_state.model_checkpoint_path),
slack=True)
saver.restore(sess,
checkpoint_state.model_checkpoint_path)
else:
log("No model to load at {}".format(save_dir),
slack=True)
saver.save(sess,
checkpoint_fpath,
global_step=global_step)
except tf.errors.OutOfRangeError as e:
log("Cannot restore checkpoint: {}".format(e), slack=True)
else:
log("Starting new training!", slack=True)
saver.save(sess, checkpoint_fpath, global_step=global_step)
# initializing feeder
feeder.start_threads(sess)
print("Feeder is intialized and model is ready to train.......")
# Training loop
while not coord.should_stop() and step < args.tacotron_train_steps:
start_time = time.time()
step, loss, opt = sess.run(
[global_step, model.loss, model.optimize])
time_window.append(time.time() - start_time)
loss_window.append(loss)
message = "Step {:7d} [{:.3f} sec/step, loss={:.5f}, avg_loss={:.5f}]".format(
step, time_window.average, loss, loss_window.average)
log(message,
end="\r",
slack=(step % args.checkpoint_interval == 0))
print(message)
if loss > 100 or np.isnan(loss):
log("Loss exploded to {:.5f} at step {}".format(
loss, step))
raise Exception("Loss exploded")
if step % args.summary_interval == 0:
log("\nWriting summary at step {}".format(step))
summary_writer.add_summary(sess.run(stats), step)
if step % args.eval_interval == 0:
pass
if step % args.checkpoint_interval == 0 or step == args.tacotron_train_steps or \
step == 300:
# Save model and current global step
saver.save(sess, checkpoint_fpath, global_step=global_step)
log("\nSaving alignment, Mel-Spectrograms and griffin-lim inverted waveform.."
)
input_seq, mel_prediction, alignment, target, target_length = sess.run(
[
model.tower_inputs[0][0],
model.tower_mel_outputs[0][0],
model.tower_alignments[0][0],
model.tower_mel_targets[0][0],
model.tower_targets_lengths[0][0],
])
# save predicted mel spectrogram to disk (debug)
mel_filename = "mel-prediction-step-{}.npy".format(step)
np.save(os.path.join(mel_dir, mel_filename),
mel_prediction.T,
allow_pickle=False)
# save griffin lim inverted wav for debug (mel -> wav)
wav = audio.inv_mel_spectrogram(mel_prediction.T, hparams)
audio.save_wav(
wav,
os.path.join(wav_dir,
"step-{}-wave-from-mel.wav".format(step)),
sr=hparams.sample_rate)
# save alignment plot to disk (control purposes)
plot.plot_alignment(
alignment,
os.path.join(plot_dir,
"step-{}-align.png".format(step)),
title="{}, {}, step={}, loss={:.5f}".format(
"Tacotron", time_string(), step, loss),
max_len=target_length // hparams.outputs_per_step)
# save real and predicted mel-spectrogram plot to disk (control purposes)
plot.plot_spectrogram(
mel_prediction,
os.path.join(
plot_dir,
"step-{}-mel-spectrogram.png".format(step)),
title="{}, {}, step={}, loss={:.5f}".format(
"Tacotron", time_string(), step, loss),
target_spectrogram=target,
max_len=target_length)
if step % args.embedding_interval == 0 or step == args.tacotron_train_steps or step == 1:
# Get current checkpoint state
checkpoint_state = tf.train.get_checkpoint_state(save_dir)
log("Tacotron training complete after {} global steps!".format(
args.tacotron_train_steps),
slack=True)
return save_dir
except Exception as e:
log("Exiting due to exception: {}".format(e), slack=True)
traceback.print_exc()
coord.request_stop(e)
def griffin_lim(mel):
return audio.inv_mel_spectrogram(mel, hparams)
def train(run_id: str, syn_dir: Path, voc_dir: Path, models_dir: Path,
ground_truth: bool, save_every: int, backup_every: int,
force_restart: bool):
# Check to make sure the hop length is correctly factorised
# assert np.cumprod(hp.voc_upsample_factors)[-1] == hp.hop_length
# Instantiate the model
print("Initializing the model...")
# model = WaveRNN(
# rnn_dims=hp.voc_rnn_dims,
# fc_dims=hp.voc_fc_dims,
# bits=hp.bits,
# pad=hp.voc_pad,
# upsample_factors=hp.voc_upsample_factors,
# feat_dims=hp.num_mels,
# compute_dims=hp.voc_compute_dims,
# res_out_dims=hp.voc_res_out_dims,
# res_blocks=hp.voc_res_blocks,
# hop_length=hp.hop_length,
# sample_rate=hp.sample_rate,
# mode=hp.voc_mode
# ).cuda()
model = model_VC(32, 256, 512, 32).cuda()
# Initialize the optimizer
optimizer = optim.Adam(model.parameters())
for p in optimizer.param_groups:
p["lr"] = hp.voc_lr
loss_recon = nn.MSELoss()
loss_content = nn.L1Loss()
# Load the weights
model_dir = models_dir.joinpath(run_id)
model_dir.mkdir(exist_ok=True)
weights_fpath = model_dir.joinpath(run_id + ".pt")
if force_restart or not weights_fpath.exists():
print("\nStarting the training of AutoVC from scratch\n")
model.save(weights_fpath, optimizer)
else:
print("\nLoading weights at %s" % weights_fpath)
model.load(weights_fpath, optimizer)
print("AutoVC weights loaded from step %d" % model.step)
# Initialize the dataset
metadata_fpath = syn_dir.joinpath("train.txt") if ground_truth else \
voc_dir.joinpath("synthesized.txt")
mel_dir = syn_dir.joinpath("mels") if ground_truth else voc_dir.joinpath(
"mels_gta")
wav_dir = syn_dir.joinpath("audio")
#2019.11.26
embed_dir = syn_dir.joinpath("embeds")
dataset = VocoderDataset(metadata_fpath, mel_dir, wav_dir, embed_dir)
test_loader = DataLoader(dataset,
batch_size=1,
shuffle=True,
pin_memory=True)
# Begin the training
simple_table([('Batch size', hp.voc_batch_size), ('LR', hp.voc_lr),
('Sequence Len', hp.voc_seq_len)])
for epoch in range(1, 350):
model.train()
data_loader = DataLoader(dataset,
collate_fn=collate_vocoder,
batch_size=hp.voc_batch_size,
num_workers=2,
shuffle=True,
pin_memory=True)
start = time.time()
running_loss = 0.
for i, (m, e, _) in enumerate(data_loader, 1):
#print("e:",e.shape)
#print("m:",m.shape)
model.train()
m, e = m.cuda(), e.cuda()
# Forward pass
C, X_C, X_before, X_after, _ = model(m, e, e)
#c_org shape: torch.Size([100, 256, 1])
#x shape: torch.Size([100, 80, 544])
#c_org_expand shape torch.Size([100, 256, 544])
#encoder_outputs shape: torch.Size([100, 544, 320])
#C shape: torch.Size([100, 544, 64])
#X shape: torch.Size([100, 1, 544, 80])
X_after = X_after.squeeze(1).permute(0, 2, 1)
X_before = X_before.squeeze(1).permute(0, 2, 1)
#print("C shape:",C.shape)
#if X_C:
# print("X_C shape:",X_C.shape)
#print("X shape:",X.shape)
# Backward pass
loss_rec_before = loss_recon(X_before, m)
loss_rec_after = loss_recon(X_after, m)
loss_c = loss_content(C, X_C)
loss = loss_rec_before + loss_rec_after + loss_c
#print("recon loss:",loss1)
#print("content loss:",loss2)
optimizer.zero_grad()
loss.backward()
optimizer.step()
#print("loss:",loss.item())
running_loss += loss.item()
#print("running loss:",running_loss)
speed = i / (time.time() - start)
avg_loss = running_loss / i
#print("avg_loss:",avg_loss)
step = model.get_step()
if hp.decay_learning_rate == True:
p["lr"] = _learning_rate_decay(p["lr"], step)
k = step // 1000
if step % 100 == 0 and step != 0:
model.eval()
plt.figure(1)
C, X_C, X_before, X_after, _ = model(m, e, e)
X_after = X_after.squeeze(1).permute(0, 2, 1)
mel_out = torch.tensor(X_after).clone().detach().cpu().numpy()
from synthesizer import audio
from synthesizer.hparams import hparams
wav = audio.inv_mel_spectrogram(mel_out[0, :, :], hparams)
librosa.output.write_wav("out.wav", np.float32(wav),
hparams.sample_rate)
mel_out = mel_out[0, :, :].transpose(1, 0)
plt.imshow(mel_out.T, interpolation='nearest', aspect='auto')
plt.title("Generate Spectrogram")
save_path = model_dir
p_path = save_path.joinpath("generate.png")
plt.savefig(p_path)
plt.figure(2)
m_out = m.squeeze(1).permute(0, 2, 1)
m_out = torch.tensor(m).clone().detach().cpu().numpy()
m_out = m_out[0, :, :].transpose(1, 0)
plt.imshow(m_out.T, interpolation='nearest', aspect='auto')
plt.title("Orignal Spectrogram")
o_path = save_path.joinpath("orignal.png")
plt.savefig(o_path)
if backup_every != 0 and step % backup_every == 0:
model.checkpoint(model_dir, optimizer)
if save_every != 0 and step % save_every == 0:
model.save(weights_fpath, optimizer)
torch.save(model, "model_ttsdb_48_48.pkl")
msg = f"| Epoch: {epoch} ({i}/{len(data_loader)}) | " \
f"Loss: {avg_loss:.4f} | {speed:.1f} " \
f"steps/s | Step: {k}k | "
stream(msg)
# gen_testset(model, test_loader, hp.voc_gen_at_checkpoint, hp.voc_gen_batched,hp.voc_target,model_dir)
print("")
mel = torch.from_numpy(mel[None, ...])
embedding_tr = embedding_tr[np.newaxis, :, np.newaxis]
embedding_tr =torch.tensor(embedding_tr)
mel,e1,embedding_tr = mel.cuda(),e1.cuda(),embedding_tr.cuda()
#print("mel shape:",mel.shape)
#print("e1 shape:",e1.shape)
#print("e2 shape:",e2.shape)
C,X_C,X_before,X_after,_ = model(mel, e1, embedding_tr)
mel_out = torch.tensor(X_after).clone().detach().cpu().numpy()
#print("mel_out shape:",mel_out.shape)
if use_wavrnn:
wav = vocoder_wavrnn.infer_waveform(mel_out[0,0,:,:].T)
else:
wav = audio.inv_mel_spectrogram(mel_out[0,0,:,:].T, hparams)
wav = librosa.resample(wav,16000,24000)
out_dir="/data/VCTK/out_v5/vcc2020-teams:00004/"
if not os.path.exists(out_dir):
os.mkdir(out_dir)
fname = t +"_"+ s +"_"+name[:-4]+".wav"
out_dir_fpath = out_dir+"/"+fname
librosa.output.write_wav(out_dir_fpath, wav.astype(np.float32),24000)
print("write:{}".format(fname))
