def load_pd(self, pd_file_path):
# pd_file_path: os.path.join(input_dir,'optimized_frozen_tacotron_o.pb')
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.allow_soft_placement = True
self.sess = tf.Session(config=config)
self.sess.run(tf.global_variables_initializer())
graph_def = tf.GraphDef()
with gfile.FastGFile(pd_file_path, 'rb') as fin:
graph_def.ParseFromString(fin.read())
tf.import_graph_def(graph_def, name='')
self._hp = hparams
self._pad = 0
if self._hp.GL_on_GPU:
# self.GLGPU_mel_inputs = tf.placeholder(tf.float32, (None, self._hp.num_mels), name='GLGPU_mel_inputs')
self.GLGPU_lin_inputs = tf.placeholder(tf.float32,
(None, self._hp.num_freq),
name='GLGPU_lin_inputs')
# self.GLGPU_mel_outputs = audio.inv_mel_spectrogram_tensorflow(self.GLGPU_mel_inputs, self._hp)
self.GLGPU_lin_outputs = audio.inv_linear_spectrogram_tensorflow(
self.GLGPU_lin_inputs, self._hp)
output_node_name = 'Tacotron_model/inference/cbhg_linear_specs_projection/projection_cbhg_linear_specs_projection/BiasAdd:0'
self.linear_outputs = self.sess.graph.get_tensor_by_name(
output_node_name)
self.inputs = self.sess.graph.get_tensor_by_name('inputs:0')
self.input_lengths = self.sess.graph.get_tensor_by_name(
'input_lengths:0')
self.split_infos = self.sess.graph.get_tensor_by_name('split_infos:0')
self.p_inputs = tf.py_func(split_func,
[self.inputs, self.split_infos[:, 0]],
tf.int32)
def load(self, checkpoint_path, hparams, gta=False, model_name='Tacotron'):
log('Constructing model: %s' % model_name)
#Force the batch size to be known in order to use attention masking in batch synthesis
inputs = tf.placeholder(tf.int32, (None, None), name='inputs')
input_lengths = tf.placeholder(tf.int32, (None), name='input_lengths')
targets = tf.placeholder(tf.float32, (None, None, hparams.num_mels), name='mel_targets')
split_infos = tf.placeholder(tf.int32, shape=(hparams.tacotron_num_gpus, None), name='split_infos')
with tf.variable_scope('Tacotron_model', reuse=tf.AUTO_REUSE) as scope:
self.model = create_model(model_name, hparams)
if gta:
self.model.initialize(inputs, input_lengths, targets, gta=gta, split_infos=split_infos)
else:
self.model.initialize(inputs, input_lengths, split_infos=split_infos)
self.mel_outputs = self.model.tower_mel_outputs
self.linear_outputs = self.model.tower_linear_outputs if (hparams.predict_linear and not gta) else None
self.alignments = self.model.tower_alignments
self.stop_token_prediction = self.model.tower_stop_token_prediction
self.targets = targets
if hparams.GL_on_GPU:
self.GLGPU_mel_inputs = tf.placeholder(tf.float32, (None, hparams.num_mels), name='GLGPU_mel_inputs')
self.GLGPU_lin_inputs = tf.placeholder(tf.float32, (None, hparams.num_freq), name='GLGPU_lin_inputs')
self.GLGPU_mel_outputs = audio.inv_mel_spectrogram_tensorflow(self.GLGPU_mel_inputs, hparams)
self.GLGPU_lin_outputs = audio.inv_linear_spectrogram_tensorflow(self.GLGPU_lin_inputs, hparams)
self.gta = gta
self._hparams = hparams
#pad input sequences with the <pad_token> 0 ( _ )
self._pad = 0
#explicitely setting the padding to a value that doesn't originally exist in the spectogram
#to avoid any possible conflicts, without affecting the output range of the model too much
if hparams.symmetric_mels:
self._target_pad = -hparams.max_abs_value
else:
self._target_pad = 0.
self.inputs = inputs
self.input_lengths = input_lengths
self.targets = targets
self.split_infos = split_infos
log('Loading checkpoint: %s' % checkpoint_path)
#Memory allocation on the GPUs as needed
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.allow_soft_placement = True
self.session = tf.Session(config=config)
self.session.run(tf.global_variables_initializer())
saver = tf.train.Saver()
saver.restore(self.session, checkpoint_path)
def load(self, checkpoint_path, hparams, gta=False, model_name='Tacotron2', freezer=False):
log('Constructing model: %s' % model_name)
if freezer:
try:
checkpoint_path = tf.train.get_checkpoint_state(checkpoint_path).model_checkpoint_path
except:
raise RuntimeError('Failed to load checkpoint at {}'.format(checkpoint_path))
# Force the batch size to be known in order to use attention masking in batch synthesis
inputs = tf.placeholder(tf.int32, (None, None), name='inputs')
input_lengths = tf.placeholder(tf.int32, (None,), name='input_lengths')
targets = tf.placeholder(tf.float32, (None, None, hparams.num_mels), name='mel_targets')
with tf.variable_scope('model', reuse=tf.AUTO_REUSE):
self.model = create_model(hparams)
if gta:
self.model.initialize(inputs, input_lengths, targets, gta=gta)
else:
self.model.initialize(inputs, input_lengths)
self.mel_outputs = self.model.mel_outputs
self.linear_outputs = self.model.linear_outputs if (hparams.predict_linear and not gta) else None
if freezer:
self.alignments = tf.identity(self.model.alignments, name="alignments")[0]
self.linear_outputs = inv_linear_spectrogram_tensorflow(self.model.linear_outputs[0], hparams=hparams)
else:
self.alignments = tf.identity(self.model.alignments, name="alignments")
self.stop_token_prediction = self.model.stop_token_prediction
self.targets = targets
self.gta = gta
self._hparams = hparams
# pad input sequences with the <pad_token> 0 ( _ )
self._pad = 0
# explicitely setting the padding to a value that doesn't originally exist in the spectogram
# to avoid any possible conflicts, without affecting the output range of the model too much
if hparams.symmetric_mels:
self._target_pad = -hparams.max_abs_value
else:
self._target_pad = 0.
self.inputs = inputs
self.input_lengths = input_lengths
self.targets = targets
log('Loading checkpoint: %s' % checkpoint_path)
# Memory allocation on the GPUs as needed
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.allow_soft_placement = True
self.session = tf.Session(config=config)
self.session.run(tf.global_variables_initializer())
saver = tf.train.Saver()
saver.restore(self.session, checkpoint_path)
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_path = os.path.join(save_dir, 'tacotron_model.ckpt')
input_path = os.path.join(args.base_dir, args.tacotron_input)
if hparams.predict_linear:
linear_dir = os.path.join(log_dir, 'linear-spectrograms')
os.makedirs(linear_dir, exist_ok=True)
log('Checkpoint path: {}'.format(checkpoint_path))
log('Loading training data from: {}'.format(input_path))
log('Using model: {}'.format(args.model))
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, input_path, 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, '..')
#Potential Griffin-Lim GPU setup
if hparams.GL_on_GPU:
GLGPU_mel_inputs = tf.placeholder(tf.float32, (None, hparams.num_mels),
name='GLGPU_mel_inputs')
GLGPU_lin_inputs = tf.placeholder(tf.float32, (None, hparams.num_freq),
name='GLGPU_lin_inputs')
GLGPU_mel_outputs = audio.inv_mel_spectrogram_tensorflow(
GLGPU_mel_inputs, hparams)
GLGPU_lin_outputs = audio.inv_linear_spectrogram_tensorflow(
GLGPU_lin_inputs, hparams)
#Book keeping
step = 0
time_window = ValueWindow(100)
loss_window = ValueWindow(100)
saver = tf.train.Saver(max_to_keep=20)
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)
ckpt = tf.train.load_checkpoint(
checkpoint_state.model_checkpoint_path)
variables = list(
ckpt.get_variable_to_shape_map().keys())
#print('=====================PRINTING VARS===============================')
#print(variables)
#drop_source_layers = ['Tacotron_model/inference/inputs_embedding','Tacotron_model/Tacotron_model/inference/inputs_embedding/Adam_1','Tacotron_model/Tacotron_model/inference/inputs_embedding/Adam']
#for v in tf.global_variables():
# if not any(layer in v.op.name for layer in drop_source_layers):
# print('Loading', v.op.name)
# v.load(ckpt.get_tensor(v.op.name), session=sess)
# Initialize all variables needed for DS, but not loaded from ckpt
#init_op = tf.variables_initializer([v for v in tf.global_variables() if any(layer in v.op.name for layer in drop_source_layers)])
#sess.run(init_op)
saver.restore(sess,
checkpoint_state.model_checkpoint_path)
else:
log('No model to load at {}'.format(save_dir),
slack=True)
saver.save(sess,
checkpoint_path,
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_path, 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, 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))
if 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
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)
if hparams.GL_on_GPU:
wav = sess.run(GLGPU_lin_outputs,
feed_dict={GLGPU_lin_inputs: lin_p})
wav = audio.inv_preemphasis(
wav, hparams.preemphasis, hparams.preemphasize)
else:
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, 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_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)
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)
log('Saving eval log to {}..'.format(eval_dir))
#Save some log to monitor model improvement on same unseen sequence
if hparams.GL_on_GPU:
wav = sess.run(GLGPU_mel_outputs,
feed_dict={GLGPU_mel_inputs: mel_p})
wav = audio.inv_preemphasis(wav, hparams.preemphasis,
hparams.preemphasize)
else:
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(
args.model, 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(
args.model, 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(
args.model, 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,
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_path, global_step=global_step)
log('\nSaving alignment, Mel-Spectrograms and griffin-lim inverted waveform..'
)
if hparams.predict_linear:
input_seq, mel_prediction, linear_prediction, alignment, target, target_length, linear_target = sess.run(
[
model.tower_inputs[0][0],
model.tower_mel_outputs[0][0],
model.tower_linear_outputs[0][0],
model.tower_alignments[0][0],
model.tower_mel_targets[0][0],
model.tower_targets_lengths[0][0],
model.tower_linear_targets[0][0],
])
#save predicted linear spectrogram to disk (debug)
linear_filename = 'linear-prediction-step-{}.npy'.format(
step)
np.save(os.path.join(linear_dir, linear_filename),
linear_prediction.T,
allow_pickle=False)
#save griffin lim inverted wav for debug (linear -> wav)
if hparams.GL_on_GPU:
wav = sess.run(GLGPU_lin_outputs,
feed_dict={
GLGPU_lin_inputs:
linear_prediction
})
wav = audio.inv_preemphasis(
wav, hparams.preemphasis, hparams.preemphasize)
else:
wav = audio.inv_linear_spectrogram(
linear_prediction.T, hparams)
audio.save_wav(
wav,
os.path.join(
wav_dir,
'step-{}-wave-from-linear.wav'.format(step)),
sr=hparams.sample_rate)
#Save real and predicted linear-spectrogram plot to disk (control purposes)
plot.plot_spectrogram(
linear_prediction,
os.path.join(
plot_dir,
'step-{}-linear-spectrogram.png'.format(step)),
title='{}, {}, step={}, loss={:.5f}'.format(
args.model, time_string(), step, loss),
target_spectrogram=linear_target,
max_len=target_length,
auto_aspect=True)
else:
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)
if hparams.GL_on_GPU:
wav = sess.run(
GLGPU_mel_outputs,
feed_dict={GLGPU_mel_inputs: mel_prediction})
wav = audio.inv_preemphasis(wav, hparams.preemphasis,
hparams.preemphasize)
else:
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(
args.model, 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(
args.model, 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 load(self, checkpoint_path, hparams, gta=False, vae_code_mode='auto', model_name='Tacotron'):
log('Constructing model: %s' % model_name)
#Force the batch size to be known in order to use attention masking in batch synthesis
inputs = tf.placeholder(tf.int32, (None, None), name='inputs')
input_lengths = tf.placeholder(tf.int32, (None), name='input_lengths')
targets = tf.placeholder(tf.float32, (None, None, hparams.num_mels), name='mel_targets')
lengths = tf.placeholder(tf.float32, (None), name='target_lengths')
mel_references = tf.placeholder(tf.float32, (None, None, hparams.num_mels), name='mel_references')
references_lengths = tf.placeholder(tf.float32, (None), name='reference_lengths')
vae_codes = tf.placeholder(tf.float32, (None, hparams.vae_dim), name='vae_codes')
split_infos = tf.placeholder(tf.int32, shape=(hparams.tacotron_num_gpus, None), name='split_infos')
with tf.variable_scope('Tacotron_model', reuse=tf.AUTO_REUSE) as scope:
self.model = create_model(model_name, hparams)
if gta:
if hparams.use_vae:
#Generate vae_code by Gaussian sampling given the mean and variance, which are generated by the VAE network given the mel_targets. Used in GTA synthesis mode.
self.model.initialize(inputs, input_lengths, mel_targets=targets, targets_lengths=lengths, gta=gta, use_vae=True, split_infos=split_infos)
else:
self.model.initialize(inputs, input_lengths, mel_targets=targets, gta=gta, split_infos=split_infos)
else:
if hparams.use_vae:
if vae_code_mode == 'auto':
#To generate vae_code by Gaussian sampling given the mean and variance, which are generated by the VAE network given the mel_references. Used in natural synthesis mode without args.modify_vae_dim specified.
self.model.initialize(inputs, input_lengths, mel_references=mel_references, references_lengths=references_lengths, gta=gta, use_vae=True, split_infos=split_infos)
elif vae_code_mode == 'feed':
#Directly feed in specified vae_code into the Tacotron decoder network while the VAE network are not used. Used in eval mode when mel_reference is not given, no matter args.modify_vae_dim is specified or not.
self.model.initialize(inputs, input_lengths, vae_codes=vae_codes, gta=gta, use_vae=True, split_infos=split_infos)
elif vae_code_mode == 'modify':
#Directly use the mean generated by the VAE network as vae_code, but with some modification according to the variance. Used in natural synthesis mode with args.modify_vae_dim specified.
self.model.initialize(inputs, input_lengths, mel_references=mel_references, references_lengths=references_lengths, vae_codes=vae_codes, gta=gta, use_vae=True, split_infos=split_infos)
elif vae_code_mode == 'inference':
#To get the vae_code(mean) generated by the VAE given the mel_references. Useful when you wish to check the quality of your VAE latent embedding
self.model.initialize(mel_references=mel_references, references_lengths=references_lengths, gta=gta, use_vae=True, split_infos=split_infos)
else:
self.model.initialize(inputs, input_lengths, gta=gta, split_infos=split_infos)
self.mu = self.model.tower_mu
self.log_var = self.model.tower_log_var
self.mel_outputs = self.model.tower_mel_outputs
self.linear_outputs = self.model.tower_linear_outputs if (hparams.predict_linear) else None
self.alignments = self.model.tower_alignments
self.stop_token_prediction = self.model.tower_stop_token_prediction
if hparams.GL_on_GPU:
self.GLGPU_mel_inputs = tf.placeholder(tf.float32, (None, hparams.num_mels), name='GLGPU_mel_inputs')
self.GLGPU_lin_inputs = tf.placeholder(tf.float32, (None, hparams.num_freq), name='GLGPU_lin_inputs')
self.GLGPU_mel_outputs = audio.inv_mel_spectrogram_tensorflow(self.GLGPU_mel_inputs, hparams)
self.GLGPU_lin_outputs = audio.inv_linear_spectrogram_tensorflow(self.GLGPU_lin_inputs, hparams)
self.gta = gta
#force feeding vae codes into the tacotron decoder(for eval mode) or generating the vae codes from the reference mel spectrograms
self.vae_code_mode = vae_code_mode
self._hparams = hparams
#pad input sequences with the <pad_token> 0 ( _ )
self._pad = 0
#explicitely setting the padding to a value that doesn't originally exist in the spectogram
#to avoid any possible conflicts, without affecting the output range of the model too much
if hparams.symmetric_mels:
self._target_pad = -hparams.max_abs_value
else:
self._target_pad = 0.
self.inputs = inputs
self.input_lengths = input_lengths
self.targets = targets
self.lengths = lengths
self.vae_codes = vae_codes
self.mel_references = mel_references
self.references_lengths = references_lengths
self.split_infos = split_infos
log('Loading checkpoint: %s' % checkpoint_path)
#Memory allocation on the GPUs as needed
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.allow_soft_placement = True
self.session = tf.Session(config=config)
self.session.run(tf.global_variables_initializer())
saver = tf.train.Saver()
saver.restore(self.session, checkpoint_path)
def load(self, checkpoint_path, hparams, gta=False, model_name='Tacotron'):
log('Constructing model: %s' % model_name)
#Force the batch size to be known in order to use attention masking in batch synthesis
inputs = tf.placeholder(tf.int32, (None, None), name='inputs')
input_lengths = tf.placeholder(tf.int32, (None), name='input_lengths')
targets = tf.placeholder(tf.float32, (None, None, hparams.num_mels),
name='mel_targets')
split_infos = tf.placeholder(tf.int32,
shape=(hparams.tacotron_num_gpus, None),
name='split_infos')
with tf.variable_scope('Tacotron_model', reuse=tf.AUTO_REUSE) as scope:
self.model = create_model(model_name, hparams)
if gta:
self.model.initialize(inputs,
input_lengths,
targets,
gta=gta,
split_infos=split_infos)
else:
self.model.initialize(inputs,
input_lengths,
split_infos=split_infos)
self.mel_outputs = self.model.tower_mel_outputs
self.linear_outputs = self.model.tower_linear_outputs if (
hparams.predict_linear and not gta) else None
self.alignments = self.model.tower_alignments
self.stop_token_prediction = self.model.tower_stop_token_prediction
self.targets = targets
if hparams.GL_on_GPU:
self.GLGPU_mel_inputs = tf.placeholder(tf.float32,
(None, hparams.num_mels),
name='GLGPU_mel_inputs')
self.GLGPU_lin_inputs = tf.placeholder(tf.float32,
(None, hparams.num_freq),
name='GLGPU_lin_inputs')
self.GLGPU_mel_outputs = audio.inv_mel_spectrogram_tensorflow(
self.GLGPU_mel_inputs, hparams)
self.GLGPU_lin_outputs = audio.inv_linear_spectrogram_tensorflow(
self.GLGPU_lin_inputs, hparams)
self.gta = gta
self._hparams = hparams
#pad input sequences with the <pad_token> 0 ( _ )
self._pad = 0
#explicitely setting the padding to a value that doesn't originally exist in the spectogram
#to avoid any possible conflicts, without affecting the output range of the model too much
if hparams.symmetric_mels:
self._target_pad = -hparams.max_abs_value
else:
self._target_pad = 0.
self.inputs = inputs
self.input_lengths = input_lengths
self.targets = targets
self.split_infos = split_infos
OLD_CHECKPOINT_FILE = checkpoint_path
NEW_CHECKPOINT_FILE = 'logs-Tacotron/taco_pretrained/new_model.ckpt-189500'
log('Loading checkpoint: %s' % checkpoint_path)
#UPDATE CHECKPOINT FILE VARS
vars_to_rename = {
"Tacotron_model/inference/CBHG_postnet/CBHG_postnet_highwaynet_1/H/biases":
"Tacotron_model/inference/CBHG_postnet/CBHG_postnet_highwaynet_1/H/bias",
}
new_checkpoint_vars = {}
reader = tf.train.NewCheckpointReader(OLD_CHECKPOINT_FILE)
for old_name in reader.get_variable_to_shape_map():
if old_name in vars_to_rename:
new_name = vars_to_rename[old_name]
else:
new_name = old_name
new_checkpoint_vars[new_name] = tf.Variable(
reader.get_tensor(old_name))
#Memory allocation on the GPUs as needed
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.allow_soft_placement = True
self.session = tf.Session(config=config)
self.session.run(tf.global_variables_initializer())
saver = tf.train.Saver(new_checkpoint_vars)
saver.restore(self.session, checkpoint_path)
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_path = os.path.join(save_dir, 'tacotron_model.ckpt')
input_path = os.path.join(args.base_dir, args.tacotron_input)
if hparams.predict_linear:
linear_dir = os.path.join(log_dir, 'linear-spectrograms')
os.makedirs(linear_dir, exist_ok=True)
log('Checkpoint path: {}'.format(checkpoint_path))
log('Loading training data from: {}'.format(input_path))
log('Using model: {}'.format(args.model))
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, input_path, hparams, args)
#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, hparams, model)
# if args.TEST:
# for v in tf.global_variables():
# print(v)
#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, '..')
#Potential Griffin-Lim GPU setup
if hparams.GL_on_GPU:
GLGPU_mel_inputs = tf.placeholder(tf.float32, (None, hparams.num_mels),
name='GLGPU_mel_inputs')
GLGPU_lin_inputs = tf.placeholder(tf.float32, (None, hparams.num_freq),
name='GLGPU_lin_inputs')
GLGPU_mel_outputs = audio.inv_mel_spectrogram_tensorflow(
GLGPU_mel_inputs, hparams)
GLGPU_lin_outputs = audio.inv_linear_spectrogram_tensorflow(
GLGPU_lin_inputs, hparams)
#Book keeping
step = 0
time_window = ValueWindow(100)
loss_window = ValueWindow(100)
loss_bef_window = ValueWindow(100)
loss_aft_window = ValueWindow(100)
loss_stop_window = ValueWindow(100)
loss_reg_window = ValueWindow(100)
loss_emt_window = ValueWindow(100)
loss_spk_window = ValueWindow(100)
loss_orthog_window = ValueWindow(100)
loss_up_emt_window = ValueWindow(100)
loss_up_spk_window = ValueWindow(100)
loss_mo_up_emt_window = ValueWindow(100)
loss_mo_up_spk_window = ValueWindow(100)
if args.nat_gan:
d_loss_t_window = ValueWindow(100)
d_loss_p_window = ValueWindow(100)
d_loss_up_window = ValueWindow(100)
g_loss_p_window = ValueWindow(100)
g_loss_up_window = ValueWindow(100)
saver = tf.train.Saver(max_to_keep=args.max_to_keep)
if args.opt_ref_no_mo and not (args.restart_optimizer_r):
print(
"WILL ATTEMPT TO RESTORE OPTIMIZER R - SET ARGS.RESTART_OPTIMIZER_R IF RETRAINING A MODEL THAT DIDN'T HAVE THE OPTIMIZER R"
)
assert (not (args.restart_nat_gan_d and args.restore_nat_gan_d_sep))
var_list = tf.global_variables()
var_list = [v for v in var_list if not ('pretrained' in v.name)]
var_list = [
v for v in var_list
if not ('nat_gan' in v.name or 'optimizer_n' in v.name)
] if (args.restart_nat_gan_d or args.restore_nat_gan_d_sep) else var_list
var_list = [
v for v in var_list
if not ('optimizer_r' in v.name or 'optimizer_3' in v.name)
] if args.restart_optimizer_r else var_list
saver_restore = tf.train.Saver(var_list=var_list)
if args.unpaired and args.pretrained_emb_disc:
saver_restore_emt_disc = tf.train.Saver(var_list=[
v for v in tf.global_variables()
if ('pretrained_ref_enc_emt' in v.name)
])
saver_restore_spk_disc = tf.train.Saver(var_list=[
v for v in tf.global_variables()
if ('pretrained_ref_enc_spk' in v.name)
])
elif args.unpaired and args.pretrained_emb_disc_all:
saver_restore_emt_disc = tf.train.Saver(var_list=[
v for v in tf.global_variables() if ('refnet_emt' in v.name)
])
saver_restore_spk_disc = tf.train.Saver(var_list=[
v for v in tf.global_variables() if ('refnet_spk' in v.name)
])
if args.nat_gan:
saver_nat_gan = tf.train.Saver(var_list=[
v for v in tf.global_variables()
if ('nat_gan' in v.name or 'optimizer_n' in v.name)
])
save_dir_nat_gan = r'nat_gan/pretrained_model'
log('Tacotron training set to a maximum of {} steps'.format(
args.tacotron_train_steps))
if hparams.tacotron_fine_tuning:
print('FINE TUNING SET TO TRUE - MAKE SURE THIS IS WHAT YOU WANT!')
#Memory allocation on the GPU as needed
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.allow_soft_placement = True
eval_feed_dict, emt_labels, spk_labels, \
basenames, basenames_refs = get_eval_feed_dict(hparams, args.synth_metadata_filename,
eval_model, args.input_dir, args.flip_spk_emt)
#Train
with tf.Session(config=config) as sess:
try:
summary_writer = tf.summary.FileWriter(tensorboard_dir, sess.graph)
# for x in tf.global_variables():
# print(x)
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.restore(
sess, checkpoint_state.model_checkpoint_path)
else:
raise ValueError(
'No model to load at {}'.format(save_dir))
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_path, global_step=global_step)
if args.unpaired and (args.pretrained_emb_disc
or args.pretrained_emb_disc_all):
save_dir_emt = r'spk_disc/pretrained_model_emt_disc'
checkpoint_state_emt = tf.train.get_checkpoint_state(
save_dir_emt)
saver_restore_emt_disc.restore(
sess, checkpoint_state_emt.model_checkpoint_path)
log('Loaded Emotion Discriminator from checkpoint {}'.format(
checkpoint_state_emt.model_checkpoint_path),
slack=True)
save_dir_spk = r'spk_disc/pretrained_model_spk_disc'
checkpoint_state_spk = tf.train.get_checkpoint_state(
save_dir_spk)
saver_restore_spk_disc.restore(
sess, checkpoint_state_spk.model_checkpoint_path)
log('Loaded Speaker Discriminator from checkpoint {}'.format(
checkpoint_state_spk.model_checkpoint_path),
slack=True)
if args.nat_gan and args.restore_nat_gan_d_sep:
checkpoint_state_nat_gan = tf.train.get_checkpoint_state(
save_dir_nat_gan)
saver_nat_gan.restore(
sess, checkpoint_state_nat_gan.model_checkpoint_path)
log('Loaded Nat Gan Discriminator from checkpoint {}'.format(
checkpoint_state_nat_gan.model_checkpoint_path),
slack=True)
#initializing feeder
feeder.start_threads(sess)
#Training loop
while not coord.should_stop() and step < args.tacotron_train_steps:
start_time = time.time()
# vars = [global_step, model.loss, model.optimize,model.before_loss, model.after_loss,model.stop_token_loss,
# model.regularization_loss,model.style_emb_loss_emt, model.style_emb_loss_spk, model.style_emb_orthog_loss]
# out = [step, loss, opt, bef, aft, stop, reg, loss_emt, loss_spk, loss_orthog]
# message = 'Step {:7d} {:.3f} sec/step, loss={:.5f}, avg_loss={:.5f}, bef={:.5f}, aft={:.5f}, stop={:.5f},' \
# 'reg={:.5f}, emt={:.5f}, spk={:.5f}, orthog={:.5f}'.format(step, time_window.average, loss, loss_window.average,
# loss_bef_window.average, loss_aft_window.average,
# loss_stop_window.average, loss_reg_window.average,
# loss_emt_window.average, loss_spk_window.average,
# loss_orthog_window.average)
# if args.unpaired:
# vars += [model.style_emb_loss_up_emt, model.style_emb_loss_up_spk,model.style_emb_loss_mel_out_up_emt, model.style_emb_loss_mel_out_up_spk]
# out += [loss_up_emt, loss_up_spk, loss_mo_up_emt, loss_mo_up_spk]
# message += ' up_emt={:.5f}, up_spk={:.5f}, mo_up_emt={:.5f}, mo_up_spk={:.5f}]'.format(loss_up_emt_window.average,
# loss_up_spk_window.average,
# loss_mo_up_emt_window.average,
# loss_mo_up_spk_window.average)
# if False:
# vars += [model.tower_style_emb_logit_emt[0], model.tower_emt_labels[0],model.tower_style_emb_logit_up_emt[0],
# model.tower_emt_up_labels[0],model.tower_spk_labels[0]]
# out += [emt_logit, emt_labels, emt_up_logit, emt_up_labels, spk_labels]
#
# out = sess.run([vars])
if args.nat_gan and (args.restart_nat_gan_d
or not (args.restore)) and step == 0:
log("Will start with Training Nat GAN Discriminator",
end='\r')
disc_epochs = 300 if args.unpaired else 200
disc_epochs = 0 if args.TEST else disc_epochs
for i in range(disc_epochs + 1):
d_loss_t, d_loss_p, d_loss_up,\
d_loss_t_emt, d_loss_p_emt, d_loss_up_emt, \
d_loss_t_spk, d_loss_p_spk, d_loss_up_spk, \
opt_n = sess.run([model.d_loss_targ, model.d_loss_p, model.d_loss_up,
model.d_loss_targ_emt, model.d_loss_p_emt, model.d_loss_up_emt,
model.d_loss_targ_spk, model.d_loss_p_spk, model.d_loss_up_spk,
model.optimize_n])
message = 'step: {}, d_loss_t={:.5f}, d_loss_p ={:.5f}, d_loss_up ={:.5f},' \
' d_loss_t_emt={:.5f}, d_loss_p_emt ={:.5f}, d_loss_up_emt ={:.5f},' \
' d_loss_t_spk={:.5f}, d_loss_p_spk ={:.5f}, d_loss_up_spk ={:.5f}'.format(i, d_loss_t, d_loss_p, d_loss_up,
d_loss_t_emt, d_loss_p_emt, d_loss_up_emt,
d_loss_t_spk, d_loss_p_spk, d_loss_up_spk)
log(message, end='\r')
os.makedirs(r'nat_gan', exist_ok=True)
os.makedirs(r'nat_gan/pretrained_model', exist_ok=True)
checkpoint_path_nat_gan = os.path.join(
save_dir_nat_gan, 'nat_gan_model.ckpt')
saver_nat_gan.save(sess,
checkpoint_path_nat_gan,
global_step=i)
if args.nat_gan:
d_loss_t, d_loss_p, d_loss_up, opt_n = sess.run([
model.d_loss_targ, model.d_loss_p, model.d_loss_up,
model.optimize_n
])
if args.unpaired:
step, tfr, loss, opt, bef, aft, stop, reg, loss_emt, loss_spk, loss_orthog, \
loss_up_emt, loss_up_spk, loss_mo_up_emt, loss_mo_up_spk, g_loss_p, g_loss_up, mels, opt_r\
= sess.run([global_step, model.ratio, model.loss, model.optimize,model.before_loss, model.after_loss,model.stop_token_loss,
model.regularization_loss, model.style_emb_loss_emt, model.style_emb_loss_spk, model.style_emb_orthog_loss,
model.style_emb_loss_up_emt, model.style_emb_loss_up_spk,model.style_emb_loss_mel_out_up_emt,
model.style_emb_loss_mel_out_up_spk,model.g_loss_p, model.g_loss_up, model.tower_mel_outputs[0], model.optimize_r])
else:
step, tfr, loss, opt, bef, aft, stop, reg, loss_emt, loss_spk, loss_orthog, \
loss_up_emt, loss_up_spk, loss_mo_up_emt, loss_mo_up_spk, g_loss_p, g_loss_up, mels,dec_out,opt_r = sess.run([global_step, model.helper._ratio, model.loss,
model.optimize, model.before_loss, model.after_loss, model.stop_token_loss,
model.regularization_loss, model.style_emb_loss_emt, model.style_emb_loss_spk, model.style_emb_orthog_loss,
model.style_emb_loss_up_emt, model.style_emb_loss_up_spk,model.style_emb_loss_mel_out_up_emt,
model.style_emb_loss_mel_out_up_spk, model.g_loss_p, model.g_loss_up, model.tower_mel_outputs[0],model.tower_decoder_output[0],model.optimize_r])
# step, loss, opt, bef, aft, stop, reg, loss_emt, loss_spk, loss_orthog, \
# loss_up_emt, loss_up_spk, loss_mo_up_emt, loss_mo_up_spk, g_loss_p, g_loss_up, mels,ref_emt,ref_spk,ref_up_emt,ref_up_spk,emb,enc_out,enc_out_up,\
# stop_pred, targ, inp, inp_len,targ_len,stop_targ,mels_up,dec_out,dec_out_up,opt_r\
# = sess.run([global_step, model.loss, model.optimize,model.before_loss, model.after_loss,model.stop_token_loss,
# model.regularization_loss, model.style_emb_loss_emt, model.style_emb_loss_spk, model.style_emb_orthog_loss,
# model.style_emb_loss_up_emt, model.style_emb_loss_up_spk,model.style_emb_loss_mel_out_up_emt,
# model.style_emb_loss_mel_out_up_spk,model.g_loss_p, model.g_loss_up, model.tower_mel_outputs[0],
# model.tower_refnet_out_emt[0],model.tower_refnet_out_spk[0],model.tower_refnet_out_up_emt[0],model.tower_refnet_out_up_spk[0],
# model.tower_embedded_inputs[0], model.tower_encoder_outputs[0],model.tower_encoder_outputs_up[0],model.tower_stop_token_prediction[0],
# model.tower_mel_targets[0],model.tower_inputs[0],model.tower_input_lengths[0],model.tower_targets_lengths[0],
# model.tower_stop_token_targets[0],model.tower_mel_outputs_up[0],model.tower_decoder_output[0],model.tower_decoder_output_up[0],model.optimize_r])
#
# if args.save_output_vars:
# import pandas as pd
# pd.DataFrame(emb[:, 0, :]).to_csv(r'C:\Users\t-mawhit\Documents\code\Tacotron-2\eval\mels_save\emb.csv')
# pd.DataFrame(enc_out[:, 0, :]).to_csv(r'C:\Users\t-mawhit\Documents\code\Tacotron-2\eval\mels_save\enc_out.csv')
# pd.DataFrame(enc_out_up[:, 0, :]).to_csv(r'C:\Users\t-mawhit\Documents\code\Tacotron-2\eval\mels_save\enc_out_up.csv')
# pd.DataFrame(stop_pred[:, :]).to_csv(r'C:\Users\t-mawhit\Documents\code\Tacotron-2\eval\mels_save\stop.csv')
# pd.DataFrame(targ[:, 0, :]).to_csv(r'C:\Users\t-mawhit\Documents\code\Tacotron-2\eval\mels_save\targ.csv')
# pd.DataFrame(inp[:, :]).to_csv(r'C:\Users\t-mawhit\Documents\code\Tacotron-2\eval\mels_save\inp.csv')
# pd.DataFrame(inp_len[:]).to_csv(r'C:\Users\t-mawhit\Documents\code\Tacotron-2\eval\mels_save\inp_len.csv')
# pd.DataFrame(targ_len[:]).to_csv(r'C:\Users\t-mawhit\Documents\code\Tacotron-2\eval\mels_save\targ_len.csv')
# pd.DataFrame(stop_targ[:, :]).to_csv(r'C:\Users\t-mawhit\Documents\code\Tacotron-2\eval\mels_save\stop_targ.csv')
# pd.DataFrame(mels_up[:, 0, 0:5]).to_csv(r'C:\Users\t-mawhit\Documents\code\Tacotron-2\eval\mels_save\mels_up.csv')
# pd.DataFrame(dec_out_up[:, 0, 0:5]).to_csv(r'C:\Users\t-mawhit\Documents\code\Tacotron-2\eval\mels_save\dec_out_up.csv')
if args.save_output_vars:
import pandas as pd
pd.DataFrame(mels[:, 0, 0:5]).to_csv(
r'C:\Users\t-mawhit\Documents\code\Tacotron-2\eval\mels_save\mels.csv'
)
pd.DataFrame(dec_out[:, 0, 0:5]).to_csv(
r'C:\Users\t-mawhit\Documents\code\Tacotron-2\eval\mels_save\dec_out.csv'
)
# import pandas as pd
# print(emt_logit.shape, emt_labels.shape)
# if len(emt_logit.shape)>2:
# emt_logit = emt_logit.squeeze(1)
# emt_up_logit = emt_up_logit.squeeze(1)
# emt_labels = emt_labels.reshape(-1,1)
# emt_up_labels = emt_up_labels.reshape(-1, 1)
# spk_labels = spk_labels.reshape(-1, 1)
# df = np.concatenate((emt_logit,emt_labels,spk_labels,emt_up_logit,emt_up_labels),axis=1)
# print(emt_labels)
# print(emt_logit)
# print(emt_up_labels)
# print(emt_up_logit)
#
# pd.DataFrame(df).to_csv(r'C:\Users\t-mawhit\Documents\code\Tacotron-2\eval\mels_save\emt_logit_.001_up_10k.csv')
# raise
time_window.append(time.time() - start_time)
loss_window.append(loss)
loss_bef_window.append(bef)
loss_aft_window.append(aft)
loss_stop_window.append(stop)
loss_reg_window.append(reg)
loss_emt_window.append(loss_emt)
loss_spk_window.append(loss_spk)
loss_orthog_window.append(loss_orthog)
loss_up_emt_window.append(loss_up_emt)
loss_up_spk_window.append(loss_up_spk)
loss_mo_up_emt_window.append(loss_mo_up_emt)
loss_mo_up_spk_window.append(loss_mo_up_spk)
if args.nat_gan:
d_loss_t_window.append(d_loss_t)
d_loss_p_window.append(d_loss_p)
d_loss_up_window.append(d_loss_up)
g_loss_p_window.append(g_loss_p)
g_loss_up_window.append(g_loss_up)
message = 'Step {:7d} {:.3f} sec/step, tfr={:.3f}, loss={:.5f}, avg_loss={:.5f}, bef={:.5f}, aft={:.5f}, stop={:.5f}, reg={:.5f}'.format(
step, time_window.average, tfr, loss, loss_window.average,
loss_bef_window.average, loss_aft_window.average,
loss_stop_window.average, loss_reg_window.average)
if args.emt_attn:
message += ' emt={:.5f}, spk={:.5f}, spk_l2={:.5f}'.format(
loss_emt_window.average, loss_spk_window.average,
loss_orthog_window.average)
else:
message += ' emt={:.5f}, spk={:.5f}, orthog={:.5f},'.format(
loss_emt_window.average, loss_spk_window.average,
loss_orthog_window.average)
if args.unpaired:
message += ' up_emt={:.5f}, up_spk={:.5f}, mo_up_emt={:.5f}, mo_up_spk={:.5f}'.format(
loss_up_emt_window.average, loss_up_spk_window.average,
loss_mo_up_emt_window.average,
loss_mo_up_spk_window.average)
if args.nat_gan:
message += ' d_loss_t={:.5f}, d_loss_p ={:.5f}, d_loss_up ={:.5f}, g_loss_p ={:.5f}, g_loss_up ={:.5f}'.format(
d_loss_t_window.average, d_loss_p_window.average,
d_loss_up_window.average, g_loss_p_window.average,
g_loss_up_window.average)
log(message,
end='\r',
slack=(step % args.checkpoint_interval == 0))
if np.isnan(loss) or loss > 100.:
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 and saving model at step {}'.format(step))
# saver.save(sess, checkpoint_path, global_step=global_step)
#
# eval_losses = []
# before_losses = []
# after_losses = []
# stop_token_losses = []
# linear_losses = []
# linear_loss = None
#
# 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)
#
# if hparams.GL_on_GPU:
# wav = sess.run(GLGPU_lin_outputs, feed_dict={GLGPU_lin_inputs: lin_p})
# wav = audio.inv_preemphasis(wav, hparams.preemphasis, hparams.preemphasize)
# else:
# 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, input_seq, 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_inputs[0][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)
#
# 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)
#
# # log('Saving eval log to {}..'.format(eval_dir))
# #Save some log to monitor model improvement on same unseen sequence
# if hparams.GL_on_GPU:
# wav = sess.run(GLGPU_mel_outputs, feed_dict={GLGPU_mel_inputs: mel_p})
# wav = audio.inv_preemphasis(wav, hparams.preemphasis, hparams.preemphasize)
# else:
# 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)
#
# input_seq = sequence_to_text(input_seq)
# plot.plot_alignment(align, os.path.join(eval_plot_dir, 'step-{}-eval-align.png'.format(step)),
# title='{}, {}, step={}, loss={:.5f}\n{}'.format(args.model, time_string(), step, eval_loss, input_seq),
# 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}\n{}'.format(args.model, time_string(), step, eval_loss,input_seq), 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(args.model, time_string(), step, eval_loss), target_spectrogram=lin_t,
# max_len=t_len, auto_aspect=True)
#
# log('Step {:7d} [eval loss: {:.3f}, before loss: {:.3f}, after loss: {:.3f}, stop loss: {:.3f}]'.format(step, eval_loss, before_loss, after_loss, stop_token_loss))
# # log('Writing eval summary!')
# add_eval_stats(summary_writer, step, linear_loss, before_loss, after_loss, stop_token_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_path, global_step=global_step)
log('\nSaved model at step {}'.format(step))
if step % args.eval_interval == 0:
if hparams.predict_linear:
raise ValueError('predict linear not implemented')
# input_seq, mel_prediction, linear_prediction, alignment, target, target_length, linear_target = sess.run([
# model.tower_inputs[0][0],
# model.tower_mel_outputs[0][0],
# model.tower_linear_outputs[0][0],
# model.tower_alignments[0][0],
# model.tower_mel_targets[0][0],
# model.tower_targets_lengths[0][0],
# model.tower_linear_targets[0][0],
# ])
#
# #save predicted linear spectrogram to disk (debug)
# linear_filename = 'linear-prediction-step-{}.npy'.format(step)
# np.save(os.path.join(linear_dir, linear_filename), linear_prediction.T, allow_pickle=False)
#
# #save griffin lim inverted wav for debug (linear -> wav)
# if hparams.GL_on_GPU:
# wav = sess.run(GLGPU_lin_outputs, feed_dict={GLGPU_lin_inputs: linear_prediction})
# wav = audio.inv_preemphasis(wav, hparams.preemphasis, hparams.preemphasize)
# else:
# wav = audio.inv_linear_spectrogram(linear_prediction.T, hparams)
# audio.save_wav(wav, os.path.join(wav_dir, 'step-{}-wave-from-linear.wav'.format(step)), sr=hparams.sample_rate)
#
# #Save real and predicted linear-spectrogram plot to disk (control purposes)
# plot.plot_spectrogram(linear_prediction, os.path.join(plot_dir, 'step-{}-linear-spectrogram.png'.format(step)),
# title='{}, {}, step={}, loss={:.5f}'.format(args.model, time_string(), step, loss), target_spectrogram=linear_target,
# max_len=target_length, auto_aspect=True)
else:
input_seqs, mels, alignments,\
stop_tokens = sess.run([eval_model.tower_inputs,
eval_model.tower_mel_outputs,
eval_model.tower_alignments,
eval_model.tower_stop_token_prediction],
feed_dict=eval_feed_dict)
# num_evals = len(input_seqs) if False else 1
# for i in range(num_evals):
# input_seq = input_seqs[i]
# mel_prediction = mel_predictions[i]
# alignment = alignments[i]
# target = targets[i]
# target_length = target_lengths[i]
# emt = emts[i]
# spk = spks[i]
# if args.emt_attn and args.attn=='simple':
# alignment_emt = alignments_emt[0][i]
# Linearize outputs (n_gpus -> 1D)
inp = [
inp for gpu_inp in input_seqs for inp in gpu_inp
]
mels = [mel for gpu_mels in mels for mel in gpu_mels]
# targets = [target for gpu_targets in targets for target in gpu_targets]
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
]
try:
target_lengths = get_output_lengths(stop_tokens)
# Take off the batch wise padding
mels = [
mel[:target_length, :]
for mel, target_length in zip(
mels, target_lengths)
]
T2_output_range = (
-hparams.max_abs_value, hparams.max_abs_value
) if hparams.symmetric_mels else (
0, hparams.max_abs_value)
mels = [
np.clip(m, T2_output_range[0],
T2_output_range[1]) for m in mels
]
folder_bucket = 'step_{}'.format(step // 500)
folder_wavs_save = os.path.join(
wav_dir, folder_bucket)
folder_plot_save = os.path.join(
plot_dir, folder_bucket)
os.makedirs(folder_wavs_save, exist_ok=True)
os.makedirs(folder_plot_save, exist_ok=True)
for i, (mel, align, basename,
basename_ref) in enumerate(
zip(mels, alignments, basenames,
basenames_refs)):
#save griffin lim inverted wav for debug (mel -> wav)
if hparams.GL_on_GPU:
wav = sess.run(
GLGPU_mel_outputs,
feed_dict={GLGPU_mel_inputs: mel})
wav = audio.inv_preemphasis(
wav, hparams.preemphasis,
hparams.preemphasize)
else:
wav = audio.inv_mel_spectrogram(
mel.T, hparams)
audio.save_wav(
wav,
os.path.join(
folder_wavs_save,
'step_{}_wav_{}_{}_{}.wav'.format(
step, i, basename, basename_ref)),
sr=hparams.sample_rate)
input_seq = sequence_to_text(inp[i])
#save alignment plot to disk (control purposes)
try:
plot.plot_alignment(
align,
os.path.join(
folder_plot_save,
'step_{}_wav_{}_{}_{}_align.png'.
format(step, i, basename,
basename_ref)),
title='{}, {}, step={}\n{}'.format(
args.model, time_string(), step,
input_seq),
max_len=target_lengths[i] //
hparams.outputs_per_step)
except:
print("failed to plot alignment")
try:
#save real and predicted mel-spectrogram plot to disk (control purposes)
plot.plot_spectrogram(
mel,
os.path.join(
folder_plot_save,
'step-{}-{}-mel-spectrogram.png'.
format(step, i)),
title='{}, {}, step={}\n{}'.format(
args.model, time_string(), step,
input_seq))
# target_spectrogram=targets[i],
# max_len=target_lengths[i])
except:
print("failed to plot spectrogram")
log('Saved synthesized samples for step {}'.format(
step),
end='\r')
except:
print("Couldn't synthesize samples")
# log('Input at step {}: {}'.format(step, input_seq), end='\r')
# 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)
