def log_validation(self, reduced_loss, model, y, y_pred, iteration, val_teacher_force_till, val_p_teacher_forcing, diagonality, avg_prob):
self.add_scalar("validation.loss", reduced_loss, iteration)
self.add_scalar("validation.attention_alignment_diagonality", diagonality, iteration)
self.add_scalar("validation.average_max_attention_weight", avg_prob, iteration)
self.add_scalar("validation.p_teacher_forcing", val_p_teacher_forcing, iteration)
self.add_scalar("validation.teacher_force_till", val_teacher_force_till, iteration)
_, mel_outputs, gate_outputs, alignments = y_pred
mel_targets, gate_targets, *_ = y
# plot distribution of parameters
for tag, value in model.named_parameters():
tag = tag.replace('.', '/')
self.add_histogram(tag, value.data.cpu().numpy(), iteration)
# plot alignment, mel target and predicted, gate target and predicted
idx = 0 # plot longest audio file
self.add_image(
"alignment",
plot_alignment_to_numpy(alignments[idx].data.cpu().numpy().T),
iteration, dataformats='HWC')
idx = 1 # and randomly pick a second one to plot
self.add_image(
"alignment2",
plot_alignment_to_numpy(alignments[idx].data.cpu().numpy().T),
iteration, dataformats='HWC')
def log_validation(self, reduced_loss, model, y, y_pred, iteration):
self.add_scalar("validation.loss", reduced_loss, iteration)
z, alignments, pred_output_lengths, log_s_sum, logdet_w_sum = y_pred
mel_targets, *_ = y
# plot distribution of parameters
for tag, value in model.named_parameters():
tag = tag.replace('.', '/')
self.add_histogram(tag, value.data.cpu().numpy(), iteration)
# plot alignment, mel target and predicted, gate target and predicted
if len(alignments.shape) == 4:
for head_i in range(alignments.shape[1]):
idx = 0 # plot longest audio file
self.add_image(
f"alignment1/h{head_i}",
plot_alignment_to_numpy(alignments[idx][head_i].data.cpu().numpy().T),
iteration, dataformats='HWC')
if alignments.shape[0] > 1: # if batch_size > 1...
idx = 1 # pick a second plot
self.add_image(
f"alignment2/h{head_i}",
plot_alignment_to_numpy(alignments[idx][head_i].data.cpu().numpy().T),
iteration, dataformats='HWC')
else:
for idx in range(2): # plot longest audio file
self.add_image(
f"alignment/{idx}",
plot_alignment_to_numpy(alignments[idx].data.cpu().numpy().T),
iteration, dataformats='HWC')
def log_alignment(self, model, enc_slf_attn, dec_enc_attn, out_mel, target,
iteration):
# plot distribution of parameters
for tag, value in model.named_parameters():
tag = tag.replace('.', '/')
self.add_histogram(tag, value.data.cpu().numpy(), iteration)
# plot alignment, mel target and predicted, gate target and predicted
#idx = random.randint(0, enc_slf_attn.size(0) - 1)
idx = 0
self.add_image(
"encoder_self_alignment",
plot_alignment_to_numpy(enc_slf_attn[idx].data.cpu().numpy().T),
iteration)
self.add_image(
"decoder_encoder_alignment",
plot_alignment_to_numpy(dec_enc_attn[idx].data.cpu().numpy().T),
iteration)
self.add_image(
"mel_target",
plot_spectrogram_to_numpy(target[idx].data.cpu().numpy()),
iteration)
self.add_image(
"mel_predicted",
plot_spectrogram_to_numpy(out_mel[idx].data.cpu().numpy()),
iteration)
def log_validation(self, reduced_loss_main, reduced_loss_join,
reduced_loss_class, reduced_loss, model, y, y_pred,
iteration):
self.add_scalar("validation.loss_mian", reduced_loss_main, iteration)
self.add_scalar("validation.loss_join", reduced_loss_join, iteration)
self.add_scalar("validation.loss_class", reduced_loss_class, iteration)
self.add_scalar("validation.loss", reduced_loss, iteration)
_, mel_outputs, alignment_outputs, acoustics_of_phone, join_outs, text_alignment = y_pred
mel_targets, alignment_targets, alignments_weights, text_alignment_padded = y
# plot distribution of parameters
for tag, value in model.named_parameters():
tag = tag.replace('.', '/')
self.add_histogram(tag, value.data.cpu().numpy(), iteration)
# plot alignment, mel target and predicted, gate target and predicted
idx = random.randint(0, mel_targets.size(0) - 1)
self.add_image(
"alignment_target",
plot_alignment_to_numpy(
alignment_targets[idx].data.cpu().numpy().T), iteration)
self.add_image(
"alignment_output",
plot_alignment_to_numpy(
alignment_outputs[idx].data.cpu().numpy().T), iteration)
self.add_image(
"mel_target",
plot_spectrogram_to_numpy(mel_targets[idx].data.cpu().numpy()),
iteration)
self.add_image(
"mel_predicted",
plot_spectrogram_to_numpy(mel_outputs[idx].data.cpu().numpy()),
iteration)
self.add_image(
"alignments_weights",
plot_weight_outputs_to_numpy(
alignments_weights[idx].data.cpu().numpy()), iteration)
self.add_image(
"acoustic_of_phone_reference",
plot_alignment_to_numpy(
acoustics_of_phone[idx].data.cpu().numpy().T,
x_label='Encoder timestep',
y_label='Phoneme Acoustic'), iteration)
self.add_image(
"acoustic_of_phone_predicted",
plot_alignment_to_numpy(join_outs[idx].data.cpu().numpy().T,
x_label='Encoder timestep',
y_label='Phoneme Acoustic'), iteration)
self.add_image(
"phone_level_acoustic_text_alignment_output",
plot_alignment_to_numpy(text_alignment[idx].data.cpu().numpy().T,
figsize=(8, 6),
x_label='Encoder timestep',
y_label='Encoder timestep'), iteration)
def log_validation(self, reduced_loss, model, y, y_pred, iteration):
log_dict = {
"loss/val": reduced_loss,
}
_, mel_outputs, gate_outputs, alignments = y_pred
mel_targets, gate_targets = y
# plot alignment, mel target and predicted, gate target and predicted
idx = random.randint(0, alignments.size(0) - 1)
align = Image.fromarray(
plot_alignment_to_numpy(alignments[idx].data.cpu().numpy().T))
align.save(os.path.join(self.outdir, f'align_{iteration:08}.png'))
target = Image.fromarray(
plot_spectrogram_to_numpy(mel_targets[idx].data.cpu().numpy()))
target.save(os.path.join(self.outdir, f'target_{iteration:08}.png'))
output = Image.fromarray(
plot_spectrogram_to_numpy(mel_outputs[idx].data.cpu().numpy()))
output.save(os.path.join(self.outdir, f'output_{iteration:08}.png'))
gate = Image.fromarray(
plot_gate_outputs_to_numpy(
gate_targets[idx].data.cpu().numpy(),
torch.sigmoid(gate_outputs[idx]).data.cpu().numpy()))
gate.save(os.path.join(self.outdir, f'gate_{iteration:08}.png'))
log_dict.update({
"alignment":
wandb.Image(Image.fromarray(
plot_alignment_to_numpy(alignments[idx].data.cpu().numpy().T)),
caption='att'),
"mel_target":
wandb.Image(Image.fromarray(
plot_spectrogram_to_numpy(
mel_targets[idx].data.cpu().numpy())),
caption='att'),
"mel_predicted":
wandb.Image(Image.fromarray(
plot_spectrogram_to_numpy(
mel_outputs[idx].data.cpu().numpy())),
caption='att'),
"gate":
wandb.Image(Image.fromarray(
plot_gate_outputs_to_numpy(
gate_targets[idx].data.cpu().numpy(),
torch.sigmoid(gate_outputs[idx]).data.cpu().numpy())),
caption='att'),
})
wandb.log(log_dict, step=iteration)
def log_teacher_forced_validation(self, reduced_loss, model, y, y_pred, iteration, val_teacher_force_till, val_p_teacher_forcing, diagonality, avg_prob):
self.add_scalar("teacher_forced_validation.loss", reduced_loss, iteration)
self.add_scalar("teacher_forced_validation.attention_alignment_diagonality", diagonality, iteration)
self.add_scalar("teacher_forced_validation.average_max_attention_weight", avg_prob, iteration)
_, mel_outputs, gate_outputs, alignments = y_pred
mel_targets, gate_targets, *_ = y
# plot distribution of parameters
for tag, value in model.named_parameters():
tag = tag.replace('.', '/')
self.add_histogram(tag, value.data.cpu().numpy(), iteration)
# plot alignment, mel target and predicted, gate target and predicted
idx = 0 # plot longest audio file
self.add_image(
"teacher_forced_alignment",
plot_alignment_to_numpy(alignments[idx].data.cpu().numpy().T),
iteration, dataformats='HWC')
self.add_image(
"mel_target",
plot_spectrogram_to_numpy(mel_targets[idx].data.cpu().numpy()),
iteration, dataformats='HWC')
self.add_image(
"mel_predicted",
plot_spectrogram_to_numpy(mel_outputs[idx].data.cpu().numpy()),
iteration, dataformats='HWC')
self.add_image(
"gate",
plot_gate_outputs_to_numpy(
gate_targets[idx].data.cpu().numpy(),
torch.sigmoid(gate_outputs[idx]).data.cpu().numpy()),
iteration, dataformats='HWC')
idx = 1 # and plot 2nd longest audio file
self.add_image(
"teacher_forced_alignment2",
plot_alignment_to_numpy(alignments[idx].data.cpu().numpy().T),
iteration, dataformats='HWC')
self.add_image(
"mel_target2",
plot_spectrogram_to_numpy(mel_targets[idx].data.cpu().numpy()),
iteration, dataformats='HWC')
self.add_image(
"mel_predicted2",
plot_spectrogram_to_numpy(mel_outputs[idx].data.cpu().numpy()),
iteration, dataformats='HWC')
self.add_image(
"gate2",
plot_gate_outputs_to_numpy(
gate_targets[idx].data.cpu().numpy(),
torch.sigmoid(gate_outputs[idx]).data.cpu().numpy()),
iteration, dataformats='HWC')
def log_validation(self, reduced_loss, model, y, y_pred, gst_scores,
iteration):
self.add_scalar("validation.loss", reduced_loss, iteration)
_, mel_outputs, gate_outputs, alignments, _ = y_pred
mel_targets, gate_targets = y
# plot distribution of parameters
for tag, value in model.named_parameters():
tag = tag.replace('.', '/')
self.add_histogram(tag, value.data.cpu().numpy(), iteration)
# plot alignment, mel target and predicted, gate target and predicted
idx = random.randint(0, alignments.size(0) - 1)
align_idx = alignments[idx].data.cpu().numpy().T
gst_scores = gst_scores.data.cpu().numpy().T
# print("Validation GST scores before plotting to tensorboard: {}".format(gst_scores.shape))
meltarg_idx = mel_targets[idx].data.cpu().numpy()
melout_idx = mel_outputs[idx].data.cpu().numpy()
self.add_image("alignment", plot_alignment_to_numpy(align_idx),
iteration)
self.add_image("gst_scores", plot_gst_scores_to_numpy(gst_scores),
iteration)
self.add_image("mel_target", plot_spectrogram_to_numpy(meltarg_idx),
iteration)
self.add_image("mel_predicted", plot_spectrogram_to_numpy(melout_idx),
iteration)
self.add_image(
"gate",
plot_gate_outputs_to_numpy(
gate_targets[idx].data.cpu().numpy(),
F.sigmoid(gate_outputs[idx]).data.cpu().numpy()), iteration)
def log_infer(self, reduced_loss_dict, reduced_bestval_loss_dict, model, y, y_pred, iteration, val_teacher_force_till, val_p_teacher_forcing):
prepend = 'inference'
# plot datapoints/graphs
self.plot_loss_dict(reduced_loss_dict, iteration, f'{prepend}')
self.plot_loss_dict(reduced_bestval_loss_dict, iteration, f'{prepend}_best')
# plot spects / imgs
n_items = min(self.n_items, y['gt_mel'].shape[0])
for idx in range(n_items):# plot target spectrogram of longest audio file(s)
self.add_image(
f"{prepend}_{idx}/alignment",
plot_alignment_to_numpy(y_pred['alignments'][idx].data.cpu().numpy().T),
iteration, dataformats='HWC')
self.add_image(
f"{prepend}_{idx}/mel_pred",
plot_spectrogram_to_numpy(y_pred['pred_mel_postnet'][idx].data.cpu().numpy()),
iteration, dataformats='HWC')
if self.plotted_targets_inf < 10:
self.add_image(
f"{prepend}_{idx}/mel_gt",
plot_spectrogram_to_numpy(y['gt_mel'][idx].data.cpu().numpy()),
iteration, dataformats='HWC')
self.plotted_targets_inf +=1 # target spect doesn't change so only needs to be plotted ~~once~~ a couple times.
def log_validation(self, reduced_loss_dict, reduced_bestval_loss_dict, model, y, y_pred, iteration, val_teacher_force_till, val_p_teacher_forcing):
prepend = 'validation'
# plot datapoints/graphs
self.plot_loss_dict(reduced_loss_dict, iteration, f'{prepend}')
self.plot_loss_dict(reduced_bestval_loss_dict, iteration, f'{prepend}_best')
# plot spects / imgs
n_items = min(self.n_items, y['gt_mel'].shape[0])
mel_L1_map = torch.nn.L1Loss(reduction='none')(y_pred['pred_mel_postnet'], y['gt_mel'])
mel_L1_map[:, -1, -1] = 5.0 # because otherwise the color map scale is crap
for idx in range(n_items):# plot target spectrogram of longest audio file(s)
self.add_image(
f"{prepend}_{idx}/alignment",
plot_alignment_to_numpy(y_pred['alignments'][idx].data.cpu().numpy().T),
iteration, dataformats='HWC')
self.add_image(
f"{prepend}_{idx}/mel_pred",
plot_spectrogram_to_numpy(y_pred['pred_mel_postnet'][idx].data.cpu().numpy()),
iteration, dataformats='HWC')
self.add_image(
f"{prepend}_{idx}/mel_SE",
plot_spectrogram_to_numpy(mel_L1_map[idx].data.cpu().numpy()),
iteration, dataformats='HWC')
if self.plotted_targets_val < 2:
self.add_image(
f"{prepend}_{idx}/mel_gt",
plot_spectrogram_to_numpy(y['gt_mel'][idx].data.cpu().numpy()),
iteration, dataformats='HWC')
self.plotted_targets_val +=1 # target spect doesn't change so only needs to be plotted once.
def log_validation(self, reduced_loss, model, y, y_pred, iteration):
self.add_scalar("validation.loss", reduced_loss, iteration)
_, mel_outputs, gate_outputs, alignments = y_pred
mel_targets, gate_targets, alignment_targets = y
# plot distribution of parameters
for tag, value in model.named_parameters():
tag = tag.replace('.', '/')
self.add_histogram(tag, value.data.cpu().numpy(), iteration)
# plot alignment, mel target and predicted, gate target and predicted
idx = random.randint(0, alignments.size(0) - 1)
self.add_image(
"alignment",
plot_alignment_to_numpy(alignments[idx].data.cpu().numpy().T),
iteration, dataformats='HWC')
self.add_image(
"mel_target",
plot_spectrogram_to_numpy(mel_targets[idx].data.cpu().numpy()),
iteration, dataformats='HWC')
self.add_image(
"mel_predicted",
plot_spectrogram_to_numpy(mel_outputs[idx].data.cpu().numpy()),
iteration, dataformats='HWC')
self.add_image(
"gate",
plot_gate_outputs_to_numpy(
gate_targets[idx].data.cpu().numpy(),
torch.sigmoid(gate_outputs[idx]).data.cpu().numpy()),
iteration, dataformats='HWC')
def log_validation(self, reduced_loss, model, y, y_pred, iteration):
# self.add_scalar("validation.loss", reduced_loss, iteration)
_, mel_outputs, gate_outputs, alignments = y_pred
mel_targets, gate_targets = y
# plot distribution of parameters
# for tag, value in model.named_parameters():
# tag = tag.replace('.', '/')
# self.add_histogram(tag, value.data.cpu().numpy(), iteration)
# plot alignment, mel target and predicted, gate target and predicted
idx = random.randint(0, alignments.size(0) - 1)
align = plot_alignment_to_numpy(alignments[idx].data.cpu().numpy().T)
spec = plot_spectrogram_to_numpy(mel_targets[idx].data.cpu().numpy())
mel = plot_spectrogram_to_numpy(mel_outputs[idx].data.cpu().numpy())
gate = plot_gate_outputs_to_numpy(
gate_targets[idx].data.cpu().numpy(),
torch.sigmoid(gate_outputs[idx]).data.cpu().numpy())
wandb = self.wandb
wandb.log({
"validation loss": reduced_loss,
"alignment": wandb.Image(align),
"spectrogram": wandb.Image(spec),
"mel_spec": wandb.Image(mel),
"gate": wandb.Image(gate),
})
def log_infer(self, reduced_loss, model, y, y_pred, iteration, val_teacher_force_till, val_p_teacher_forcing, diagonality, avg_prob):
self.add_scalar("infer.loss", reduced_loss, iteration)
self.add_scalar("infer.attention_alignment_diagonality", diagonality, iteration)
self.add_scalar("infer.average_max_attention_weight", avg_prob, iteration)
_, mel_outputs, gate_outputs, alignments = y_pred
mel_targets, gate_targets, *_ = y
# plot alignment, mel target and predicted, gate target and predicted
idx = 0 # plot longest audio file
self.add_image(
"infer_alignment",
plot_alignment_to_numpy(alignments[idx].data.cpu().numpy().T),
iteration, dataformats='HWC')
self.add_image(
"infer_mel_target",
plot_spectrogram_to_numpy(mel_targets[idx].data.cpu().numpy()),
iteration, dataformats='HWC')
self.add_image(
"infer_mel_predicted",
plot_spectrogram_to_numpy(mel_outputs[idx].data.cpu().numpy()),
iteration, dataformats='HWC')
self.add_image(
"infer_gate",
plot_gate_outputs_to_numpy(
gate_targets[idx].data.cpu().numpy(),
torch.sigmoid(gate_outputs[idx]).data.cpu().numpy()),
iteration, dataformats='HWC')
idx = 1 # and plot 2nd longest audio file
self.add_image(
"infer_alignment2",
plot_alignment_to_numpy(alignments[idx].data.cpu().numpy().T),
iteration, dataformats='HWC')
self.add_image(
"infer_mel_target2",
plot_spectrogram_to_numpy(mel_targets[idx].data.cpu().numpy()),
iteration, dataformats='HWC')
self.add_image(
"infer_mel_predicted2",
plot_spectrogram_to_numpy(mel_outputs[idx].data.cpu().numpy()),
iteration, dataformats='HWC')
self.add_image(
"infer_gate2",
plot_gate_outputs_to_numpy(
gate_targets[idx].data.cpu().numpy(),
torch.sigmoid(gate_outputs[idx]).data.cpu().numpy()),
iteration, dataformats='HWC')
def validate(model, criterion, valset, epoch, batch_iter, batch_size,
world_size, collate_fn, distributed_run, rank, batch_to_gpu,
summary_writer=None):
"""Handles all the validation scoring and printing"""
with evaluating(model), torch.no_grad():
val_sampler = DistributedSampler(valset) if distributed_run else None
val_loader = DataLoader(valset, num_workers=1, shuffle=False,
sampler=val_sampler,
batch_size=batch_size, pin_memory=False,
collate_fn=collate_fn)
val_loss = 0.0
num_iters = 0
val_items_per_sec = 0.0
for i, batch in enumerate(val_loader):
torch.cuda.synchronize()
iter_start_time = time.perf_counter()
x, y, num_items = batch_to_gpu(batch)
y_pred = model(x)
loss = criterion(y_pred, y)
if i % 50 == 0 and summary_writer is not None:
_, mel_outputs_postnet, _, alignments = y_pred
summary_writer.add_image(
f'attention_weights_sample_{i}',
plot_alignment_to_numpy(alignments[0].data.cpu().numpy().T),
epoch,
dataformats='HWC',
)
summary_writer.add_image(
f'mel_outputs_sample_{i}',
plot_mel_to_numpy(mel_outputs_postnet[0].data.cpu().numpy()),
epoch,
dataformats='HWC',
)
if distributed_run:
reduced_val_loss = reduce_tensor(loss.data, world_size).item()
reduced_num_items = reduce_tensor(num_items.data, 1).item()
else: #
reduced_val_loss = loss.item()
reduced_num_items = num_items.item()
val_loss += reduced_val_loss
torch.cuda.synchronize()
iter_stop_time = time.perf_counter()
iter_time = iter_stop_time - iter_start_time
items_per_sec = reduced_num_items/iter_time
DLLogger.log(step=(epoch, batch_iter, i), data={'val_items_per_sec': items_per_sec})
val_items_per_sec += items_per_sec
num_iters += 1
val_loss = val_loss/(i + 1)
DLLogger.log(step=(epoch,), data={'val_loss': val_loss})
DLLogger.log(step=(epoch,), data={'val_items_per_sec':
(val_items_per_sec/num_iters if num_iters > 0 else 0.0)})
return val_loss, val_items_per_sec
def log_alignment(self, model, dec_enc_attn, alignment, mel_padded, mel_predict, test_attn, iteration):
# plot distribution of parameters
for tag, value in model.named_parameters():
tag = tag.replace('.', '/')
self.add_histogram(tag, value.data.cpu().numpy(), iteration)
# plot alignment, mel target and predicted, gate target and predicted
idx = random.randint(0, dec_enc_attn[0].size(0) - 1)
mel_padded = mel_padded.permute(0, 2, 1)
mel_predict = mel_predict.permute(0, 2, 1)
'''self.add_image(
"encoder_self_alignment",
plot_alignment_to_numpy(enc_slf_attn[idx].data.cpu().numpy().T),
iteration)'''
for i in range(len(dec_enc_attn)):
self.add_image(
"decoder_encoder_alignment_{}".format(i),
plot_alignment_to_numpy(dec_enc_attn[i][idx].data.cpu().numpy().T),
iteration)
self.add_image(
"test_alignment_{}".format(i),
plot_alignment_to_numpy(test_attn[len(test_attn)-i-1][idx].data.cpu().numpy().T),
iteration)
self.add_image(
"target_alignment",
plot_alignment_to_numpy(alignment[idx].data.cpu().numpy().T),
iteration)
self.add_image(
"target_mel",
plot_spectrogram_to_numpy(mel_padded[idx].data.cpu().numpy().T),
iteration)
self.add_image(
"predict_mel",
plot_spectrogram_to_numpy(mel_predict[idx].data.cpu().numpy().T),
iteration)
def inference(args):
sentences = get_sentences(args)
model = load_model(hparams)
model.load_state_dict(torch.load(args.checkpoint)['state_dict'])
model.cuda().eval()
test_set = TextMelLoaderEval(sentences, hparams)
test_collate_fn = TextMelCollateEval(hparams)
test_sampler = DistributedSampler(
valset) if hparams.distributed_run else None
test_loader = DataLoader(test_set,
num_workers=0,
sampler=test_sampler,
batch_size=hparams.synth_batch_size,
pin_memory=False,
drop_last=True,
collate_fn=test_collate_fn)
T2_output_range = (-hparams.max_abs_value,
hparams.max_abs_value) if hparams.symmetric_mels else (
0, hparams.max_abs_value)
os.makedirs(args.out_filename, exist_ok=True)
with torch.no_grad():
for i, batch in enumerate(test_loader):
mel_outputs, mel_outputs_postnet, _, alignments = model.inference(
batch)
align_img = Image.fromarray(
plot_alignment_to_numpy(alignments[0].data.cpu().numpy().T))
spec_img = Image.fromarray(
plot_spectrogram_to_numpy(
mel_outputs_postnet[0].data.cpu().numpy()))
align_img.save(
os.path.join(args.out_filename,
'sentence_{}_alignment.jpg'.format(i)))
spec_img.save(
os.path.join(args.out_filename,
'sentence_{}_mel-spectrogram.jpg'.format(i)))
mels = mel_outputs_postnet[0].cpu().numpy()
mel_path = os.path.join(args.out_filename,
'sentence_{}_mel-feats.npy'.format(i))
mels = np.clip(mels, T2_output_range[0], T2_output_range[1])
np.save(mel_path, mels.T, allow_pickle=False)
print('CHECK MEL SHAPE:', mels.T.shape)
def log_validation(self,
reduced_loss,
model,
y,
y_pred,
iteration,
model_name="",
log_embedding=False):
self.add_scalar("validation.loss", reduced_loss, iteration)
_, mel_outputs, gate_outputs, alignments = y_pred
mel_targets, gate_targets = y
# plot distribution of parameters
for tag, value in model.named_parameters():
tag = tag.replace('.', '/')
self.add_histogram(tag, value.data.cpu().numpy(), iteration)
# plot alignment, mel target and predicted, gate target and predicted
idx = random.randint(0, alignments.size(0) - 1)
self.add_image(
"alignment",
np.moveaxis(
plot_alignment_to_numpy(alignments[idx].data.cpu().numpy().T),
2, 0), iteration)
self.add_image(
"mel_target",
np.moveaxis(
plot_spectrogram_to_numpy(mel_targets[idx].data.cpu().numpy()),
2, 0), iteration)
self.add_image(
"mel_predicted",
np.moveaxis(
plot_spectrogram_to_numpy(mel_outputs[idx].data.cpu().numpy()),
2, 0), iteration)
self.add_image(
"gate",
np.moveaxis(
plot_gate_outputs_to_numpy(
gate_targets[idx].data.cpu().numpy(),
torch.sigmoid(gate_outputs[idx]).data.cpu().numpy()), 2,
0), iteration)
if log_embedding:
self.add_embedding(
model.speaker_embedding.weight.detach().cpu().numpy(), [
str(i)
for i in range(model.speaker_embedding.num_embeddings)
],
global_step=iteration,
tag='emb_{}'.format(model_name))
def log_validation(self, reduced_loss, reduced_losses, reduced_acces,
model, y, y_pred, iteration, task):
self.add_scalar('validation.loss.%s' % task, reduced_loss, iteration)
self.add_scalar('validation.acc.%s.texcl' % task, reduced_acces[0],
iteration)
mel_hidden, text_logit_from_mel_hidden, audio_seq2seq_alignments, text_lengths = y_pred
# plot alignment, mel target and predicted, stop target and predicted
idx = random.randint(0, audio_seq2seq_alignments.size(0) - 1)
audio_seq2seq_alignments = audio_seq2seq_alignments.data.cpu().numpy()
self.add_image("%s.audio_seq2seq_alignment" % task,
plot_alignment_to_numpy(
audio_seq2seq_alignments[idx].T),
iteration,
dataformats='HWC')
def log_infer(self, reduced_loss, model, y, y_pred, iteration,
val_teacher_force_till, val_p_teacher_forcing, diagonality,
avg_prob):
self.add_scalar("infer.loss", reduced_loss, iteration)
self.add_scalar("infer.attention_alignment_diagonality", diagonality,
iteration)
self.add_scalar("infer.average_max_attention_weight", avg_prob,
iteration)
mel_outputs, mel_outputs_postnet, gate_outputs, alignments, *_ = y_pred
if mel_outputs_postnet is not None:
mel_outputs = mel_outputs_postnet
mel_outputs_GAN = y_pred[8][0]
mel_targets, gate_targets, *_ = y
mel_outputs = mel_outputs[:, :mel_targets.shape[1], :]
plot_n_files = 5
# plot infer alignment, mel target and predicted, gate predicted
for idx in range(plot_n_files): # plot longest x audio files
str_idx = '' if idx == 0 else idx
self.add_image(f"infer_alignment{str_idx}",
plot_alignment_to_numpy(
alignments[idx].data.cpu().numpy().T),
iteration,
dataformats='HWC')
self.add_image(f"infer_mel_target{str_idx}",
plot_spectrogram_to_numpy(
mel_targets[idx].data.cpu().numpy()),
iteration,
dataformats='HWC')
self.add_image(f"infer_mel_predicted{str_idx}",
plot_spectrogram_to_numpy(
mel_outputs[idx].data.cpu().numpy()),
iteration,
dataformats='HWC')
if mel_outputs_GAN is not None:
self.add_image(f"mel_predicted_GAN{str_idx}",
plot_spectrogram_to_numpy(
mel_outputs_GAN[idx].data.cpu().numpy()),
iteration,
dataformats='HWC')
self.add_image(f"infer_gate{str_idx}",
plot_gate_outputs_to_numpy(
gate_targets[idx].data.cpu().numpy(),
torch.sigmoid(
gate_outputs[idx]).data.cpu().numpy()),
iteration,
dataformats='HWC')
def log_validation(self,
reduced_loss,
model,
y,
y_pred,
iteration,
speaker_acc=0,
augment_acc=0):
self.add_scalar("validation.loss", reduced_loss, iteration)
self.add_scalar("Speaker_classifier_ACC", speaker_acc, iteration)
self.add_scalar("Augment_classifier_ACC", augment_acc, iteration)
_, mel_outputs, gate_outputs, alignments, speaker_output, augmentation_output, _, _ = y_pred
mel_targets, gate_targets, speaker_id, labels = y
# plot distribution of parameters
for tag, value in model.named_parameters():
tag = tag.replace('.', '/')
self.add_histogram(tag, value.data.cpu().numpy(), iteration)
# plot alignment, mel target and predicted, gate target and predicted
idx = random.randint(0, alignments.size(0) - 1)
self.add_image(
"alignment",
torch.from_numpy(
plot_alignment_to_numpy(
alignments[idx].data.cpu().numpy().T)).permute(2, 0, 1),
iteration)
self.add_image(
"mel_target",
torch.from_numpy(
plot_spectrogram_to_numpy(
mel_targets[idx].data.cpu().numpy())).permute(2, 0, 1),
iteration)
self.add_image(
"mel_predicted",
torch.from_numpy(
plot_spectrogram_to_numpy(
mel_outputs[idx].data.cpu().numpy())).permute(2, 0, 1),
iteration)
self.add_image(
"gate",
torch.from_numpy(
plot_gate_outputs_to_numpy(
gate_targets[idx].data.cpu().numpy(),
F.sigmoid(gate_outputs[idx]).data.cpu().numpy())).permute(
2, 0, 1), iteration)
def log_validation(self, reduced_loss, model, y, y_pred, iteration):
self.add_scalar("validation.loss", reduced_loss, iteration)
_, mel_outputs, gate_outputs, alignments = y_pred
mel_targets, gate_targets = y
# plot distribution of parameters
for tag, value in model.named_parameters():
tag = tag.replace('.', '/')
self.add_histogram(tag, value.data.cpu().numpy(), iteration)
# plot alignment, mel target and predicted, gate target and predicted
idx = random.randint(0, alignments.size(0) - 1)
self.add_image("alignment",
plot_alignment_to_numpy(
alignments[idx].data.cpu().numpy().T),
iteration,
dataformats='HWC')
self.add_image("mel_target",
plot_spectrogram_to_numpy(
mel_targets[idx].data.cpu().numpy()),
iteration,
dataformats='HWC')
self.add_image("mel_predicted",
plot_spectrogram_to_numpy(
mel_outputs[idx].data.cpu().numpy()),
iteration,
dataformats='HWC')
self.add_image("gate",
plot_gate_outputs_to_numpy(
gate_targets[idx].data.cpu().numpy(),
torch.sigmoid(
gate_outputs[idx]).data.cpu().numpy()),
iteration,
dataformats='HWC')
mel = mel_outputs.cpu()[0]
if len(mel.shape) == 2:
mel = mel.unsqueeze(0)
audio = self.melgan.inference(mel)
self.add_audio('audio',
audio,
global_step=iteration,
sample_rate=self.sampling_rate,
walltime=None)
def log_validation(self, reduced_loss, model, y, y_pred, iteration):
self.add_scalar("validation.loss", reduced_loss, iteration)
wandb.log({'validation.loss': reduced_loss}, step=iteration)
_, mel_outputs, gate_outputs, alignments = y_pred
mel_targets, gate_targets = y
# plot distribution of parameters
for tag, value in model.named_parameters():
tag = tag.replace('.', '/')
self.add_histogram(tag, value.data.cpu().numpy(), iteration)
# plot alignment, mel target and predicted, gate target and predicted
idx = random.randint(0, alignments.size(0) - 1)
alignment_arr = plot_alignment_to_numpy(
alignments[idx].data.cpu().numpy().T)
self.add_image("alignment", alignment_arr, iteration)
wandb.log(
{"alignment": [wandb.Image(alignment_arr, caption="Alignment")]},
step=iteration)
mel_target = plot_spectrogram_to_numpy(
mel_targets[idx].data.cpu().numpy())
self.add_image("mel_target", mel_target, iteration)
wandb.log(
{"mel_target": [wandb.Image(mel_target, caption="Mel target")]},
step=iteration)
mel_predicted = plot_spectrogram_to_numpy(
mel_outputs[idx].data.cpu().numpy())
self.add_image("mel_predicted", mel_predicted, iteration)
wandb.log(
{
"mel_predicted":
[wandb.Image(mel_predicted, caption="Mel predicted")]
},
step=iteration)
self.add_image(
"gate",
plot_gate_outputs_to_numpy(
gate_targets[idx].data.cpu().numpy(),
torch.sigmoid(gate_outputs[idx]).data.cpu().numpy()),
iteration)
def log_validation(self, reduced_loss, model, y, y_pred, iteration):
self.add_scalar("validation.loss", reduced_loss, iteration)
if self.use_vae:
_, mel_outputs, gate_outputs, alignments, mus, _, _, emotions = y_pred
else:
_, mel_outputs, gate_outputs, alignments = y_pred
mel_targets, gate_targets = y
#print('emotion:\n{}'.format(emotions))
# plot distribution of parameters
for tag, value in model.named_parameters():
tag = tag.replace('.', '/')
self.add_histogram(tag, value.data.cpu().numpy(), iteration)
# plot alignment, mel target and predicted, gate target and predicted
idx = random.randint(0, alignments.size(0) - 1)
self.add_image(
"alignment",
plot_alignment_to_numpy(alignments[idx].data.cpu().numpy().T),
iteration, dataformats=self.dataformat)
self.add_image(
"mel_target",
plot_spectrogram_to_numpy(mel_targets[idx].data.cpu().numpy()),
iteration, dataformats=self.dataformat)
self.add_image(
"mel_predicted",
plot_spectrogram_to_numpy(mel_outputs[idx].data.cpu().numpy()),
iteration, dataformats=self.dataformat)
self.add_image(
"gate",
plot_gate_outputs_to_numpy(
gate_targets[idx].data.cpu().numpy(),
torch.sigmoid(gate_outputs[idx]).data.cpu().numpy()),
iteration, dataformats=self.dataformat)
if self.use_vae:
self.add_image(
"latent_dim (regular)",
plot_scatter(mus, emotions),
iteration, dataformats=self.dataformat)
self.add_image(
"latent_dim (t-sne)",
plot_tsne(mus, emotions),
iteration, dataformats=self.dataformat)
def log_validation(self, reduced_loss, model, x, y, y_pred, iteration,
hparams):
self.add_scalar("validation.loss", reduced_loss, iteration)
_, mel_outputs, gate_outputs, alignments = y_pred
mel_targets, gate_targets = y
# plot distribution of parameters
for tag, value in model.named_parameters():
tag = tag.replace('.', '/')
self.add_histogram(tag, value.data.cpu().numpy(), iteration)
# plot alignment, mel target and predicted, gate target and predicted
idx = random.randint(0, alignments.size(0) - 1)
self.add_image(
"alignment",
plot_alignment_to_numpy(alignments[idx].data.cpu().numpy().T),
iteration)
self.add_image(
"mel_target",
plot_spectrogram_to_numpy(mel_targets[idx].data.cpu().numpy()),
iteration)
self.add_image(
"mel_predicted",
plot_spectrogram_to_numpy(mel_outputs[idx].data.cpu().numpy()),
iteration)
self.add_image(
"gate",
plot_gate_outputs_to_numpy(
gate_targets[idx].data.cpu().numpy(),
F.sigmoid(gate_outputs[idx]).data.cpu().numpy()), iteration)
self.add_audio(
"audio_from_target",
synthesis_griffin_lim(mel_targets[idx].unsqueeze(0), hparams),
iteration, hparams.sampling_rate)
self.add_audio(
"audio_from_predicted",
synthesis_griffin_lim(mel_outputs[idx].unsqueeze(0), hparams),
iteration, hparams.sampling_rate)
self.add_text(
"text", ''.join([
_id_to_symbol[symbol_id]
for symbol_id in x[0][idx].data.cpu().numpy()
]), iteration)
def log_validation(self, params, iteration):
for key, key_params in params.items():
self.add_scalar(f'{key}.loss', key_params['loss'], iteration)
_, mel_outputs, gate_outputs, alignments = key_params['y_pred']
mel_targets, gate_targets = key_params['y']
# plot alignment, mel target and predicted, gate target and predicted
idx = random.randint(0, mel_outputs.size(0) - 1)
self.add_image(
f'{key}.alignment',
plot_alignment_to_numpy(alignments[idx].data.cpu().numpy().T),
iteration, dataformats='HWC')
self.add_image(
f'{key}.mel_target',
plot_spectrogram_to_numpy(mel_targets[idx].data.cpu().numpy()),
iteration, dataformats='HWC')
self.add_image(
f'{key}.mel_predicted',
plot_spectrogram_to_numpy(mel_outputs[idx].data.cpu().numpy()),
iteration, dataformats='HWC')
def log_validation(self, reduced_loss, model, y, y_pred, iteration, x):
self.add_scalar("validation.loss", reduced_loss, iteration)
_, mel_outputs, gate_outputs, alignments = y_pred
mel_targets, gate_targets = y
text_inputs = x[0]
speaker_ids = x[5]
# plot distribution of parameters
for tag, value in model.named_parameters():
tag = tag.replace('.', '/')
self.add_histogram(tag, value.data.cpu().numpy(), iteration)
# plot alignment, mel target and predicted, gate target and predicted
idx = random.randint(0, alignments.size(0) - 1)
self.add_image("alignment",
plot_alignment_to_numpy(
alignments[idx].data.cpu().numpy().T),
iteration,
dataformats='HWC')
self.add_image("mel_predicted",
plot_spectrogram_to_numpy(
mel_outputs[idx].data.cpu().numpy()),
iteration,
dataformats='HWC')
self.add_image("gate",
plot_gate_outputs_to_numpy(
gate_targets[idx].data.cpu().numpy(),
torch.sigmoid(
gate_outputs[idx]).data.cpu().numpy()),
iteration,
dataformats='HWC')
# 记录一下合成的语音效果。
audio_predicted = inv_linearspectrogram(
mel_outputs[idx].data.cpu().numpy())
self.add_audio('audio_predicted',
torch.from_numpy(audio_predicted),
iteration,
sample_rate=default_hparams.sample_rate)
self.add_image("mel_target",
plot_spectrogram_to_numpy(
mel_targets[idx].data.cpu().numpy()),
iteration,
dataformats='HWC')
audio_target = inv_linearspectrogram(
mel_targets[idx].data.cpu().numpy())
self.add_audio('audio_target',
torch.from_numpy(audio_target),
iteration,
sample_rate=default_hparams.sample_rate)
spk = int(speaker_ids[idx].data.cpu().numpy().flatten()[0])
ph_ids = text_inputs[idx].data.cpu().numpy().flatten()
phs_text = sequence_to_text(ph_ids)
phs_size = len(ph_ids)
reduced_loss = float(reduced_loss)
audt_duration = int(
len(audio_target) / (default_hparams.sample_rate / 1000))
audp_duration = int(
len(audio_predicted) / (default_hparams.sample_rate / 1000))
spect_shape = mel_targets[idx].data.cpu().numpy().shape
specp_shape = mel_outputs[idx].data.cpu().numpy().shape
align_shape = alignments[idx].data.cpu().numpy().T.shape
out_text = dict(speaker_id=spk,
phonemes=phs_text,
phonemes_size=phs_size,
validation_loss=reduced_loss,
audio_target_ms=audt_duration,
audio_predicted_ms=audp_duration,
spectrogram_target_shape=str(spect_shape),
spectrogram_predicted_shape=str(specp_shape),
alignment_shape=str(align_shape))
out_text = json.dumps(out_text, indent=4, ensure_ascii=False)
out_text = f'<pre>{out_text}</pre>' # 支持html标签
self.add_text('text', out_text, iteration)
def log_validation(self, reduced_loss, model, x, y, y_pred, iteration,
epoch, sample_rate):
text_padded, input_lengths, mel_padded, max_len, output_lengths = x
#self.add_scalar("validation.loss", reduced_loss, iteration) # Tensorboard log
_, mel_outputs, gate_outputs, alignments = y_pred
mel_targets, gate_targets = y
# plot distribution of parameters
for tag, value in model.named_parameters():
tag = tag.replace('.', '/')
# self.add_histogram(tag, value.data.cpu().numpy(), iteration) # Tensorboard log
wandb.log(
{
tag: wandb.Histogram(value.data.cpu().numpy()),
"epoch": epoch,
"iteration": iteration
},
step=iteration)
# plot alignment, mel target and predicted, gate target and predicted
idx = random.randint(0, alignments.size(0) - 1)
text_len = input_lengths[idx].item()
text_string = sequence_to_text(text_padded[idx].tolist())[:text_len]
mel_len = get_mel_length(alignments, idx, text_len)
mel = mel_outputs[idx:idx + 1, :, :mel_len]
np_wav = self.mel2wav(mel.type('torch.cuda.HalfTensor'))
np_alignment = plot_alignment_to_numpy(
alignments[idx].data.cpu().numpy().T, decoding_len=mel_len)
'''self.add_image(
"alignment",
np_alignment,
iteration, dataformats='HWC')'''
np_mel_target = plot_spectrogram_to_numpy(
mel_targets[idx].data.cpu().numpy())
'''self.add_image(
"mel_target",
np_mel_target,
iteration, dataformats='HWC')'''
np_mel_predicted = plot_spectrogram_to_numpy(
mel_outputs[idx].data.cpu().numpy())
'''self.add_image(
"mel_predicted",
np_mel_predicted,
iteration, dataformats='HWC')'''
np_gate = plot_gate_outputs_to_numpy(
gate_targets[idx].data.cpu().numpy(),
torch.sigmoid(gate_outputs[idx]).data.cpu().numpy())
'''self.add_image(
"gate",
np_gate,
iteration, dataformats='HWC')'''
# wandb log
wandb.log(
{
"val/loss":
reduced_loss,
"val/alignment":
[wandb.Image(np_alignment, caption=text_string)],
"val/audio": [
wandb.Audio(np_wav.astype(np.float32),
caption=text_string,
sample_rate=sample_rate)
],
"val/mel_target": [wandb.Image(np_mel_target)],
"val/mel_predicted": [wandb.Image(np_mel_predicted)],
"val/gate": [wandb.Image(np_gate)],
"epoch":
epoch,
"iteration":
iteration
},
step=iteration)
# foward attention ratio
hop_list = [1]
for hop_size in hop_list:
mean_far, batch_far = forward_attention_ratio(
alignments, input_lengths, hop_size)
log_name = "mean_forward_attention_ratio.val/hop_size={}".format(
hop_size)
wandb.log(
{
log_name: mean_far,
"epoch": epoch,
"iteration": iteration
},
step=iteration)
log_name = "forward_attention_ratio.val/hop_size={}".format(
hop_size)
wandb.log(
{
log_name: wandb.Histogram(batch_far.data.cpu().numpy()),
"epoch": epoch,
"iteration": iteration
},
step=iteration)
def log_teacher_forced_validation(self, reduced_loss, model, y, y_pred,
iteration, val_teacher_force_till,
val_p_teacher_forcing, diagonality,
avg_prob):
self.add_scalar("teacher_forced_validation.loss", reduced_loss,
iteration)
self.add_scalar(
"teacher_forced_validation.attention_alignment_diagonality",
diagonality, iteration)
self.add_scalar(
"teacher_forced_validation.average_max_attention_weight", avg_prob,
iteration)
mel_outputs, mel_outputs_postnet, gate_outputs, alignments, *_ = y_pred
if mel_outputs_postnet is not None:
mel_outputs = mel_outputs_postnet
mel_outputs_GAN = y_pred[8][0]
mel_targets, gate_targets, *_ = y
mel_outputs = mel_outputs[:, :mel_targets.shape[1], :]
mel_MSE_map = torch.nn.MSELoss(reduction='none')(mel_outputs,
mel_targets)
mel_MSE_map[:, -1,
-1] = 20.0 # because otherwise the color map scale is crap
# plot distribution of parameters
for tag, value in model.named_parameters():
tag = tag.replace('.', '/')
self.add_histogram(tag, value.data.cpu().numpy(), iteration)
plot_n_files = 5
# plot alignment, mel target and predicted, gate target and predicted
for idx in range(plot_n_files): # plot longest x audio files
str_idx = '' if idx == 0 else idx
self.add_image(f"teacher_forced_alignment{str_idx}",
plot_alignment_to_numpy(
alignments[idx].data.cpu().numpy().T),
iteration,
dataformats='HWC')
self.add_image(f"mel_target{str_idx}",
plot_spectrogram_to_numpy(
mel_targets[idx].data.cpu().numpy()),
iteration,
dataformats='HWC')
self.add_image(f"mel_predicted{str_idx}",
plot_spectrogram_to_numpy(
mel_outputs[idx].data.cpu().numpy()),
iteration,
dataformats='HWC')
if mel_outputs_GAN is not None:
self.add_image(f"mel_predicted_GAN{str_idx}",
plot_spectrogram_to_numpy(
mel_outputs_GAN[idx].data.cpu().numpy()),
iteration,
dataformats='HWC')
self.add_image(f"mel_squared_error{str_idx}",
plot_spectrogram_to_numpy(
mel_MSE_map[idx].data.cpu().numpy()),
iteration,
dataformats='HWC')
self.add_image(f"gate{str_idx}",
plot_gate_outputs_to_numpy(
gate_targets[idx].data.cpu().numpy(),
torch.sigmoid(
gate_outputs[idx]).data.cpu().numpy()),
iteration,
dataformats='HWC')
def log_validation(self, reduced_loss, model, y, y_pred, iteration):
self.add_scalar("validation.loss", reduced_loss, iteration)
_, mel_outputs, gate_outputs, alignments = y_pred
mel_targets, gate_targets = y
# plot distribution of parameters
for tag, value in model.named_parameters():
tag = tag.replace('.', '/')
self.add_histogram(tag, value.data.cpu().numpy(), iteration)
# plot alignment, mel target and predicted, gate target and predicted
idx = random.randint(0, alignments.size(0) - 1)
self.add_image("alignment",
plot_alignment_to_numpy(
alignments[idx].data.cpu().numpy().T),
iteration,
dataformats='HWC')
self.add_image("mel_target",
plot_spectrogram_to_numpy(
mel_targets[idx].data.cpu().numpy()),
iteration,
dataformats='HWC')
self.add_image("mel_predicted",
plot_spectrogram_to_numpy(
mel_outputs[idx].data.cpu().numpy()),
iteration,
dataformats='HWC')
self.add_image("gate",
plot_gate_outputs_to_numpy(
gate_targets[idx].data.cpu().numpy(),
torch.sigmoid(
gate_outputs[idx]).data.cpu().numpy()),
iteration,
dataformats='HWC')
# import random
# import torch.nn.functional as F
# from tensorboardX import SummaryWriter
# from plotting_utils import plot_alignment_to_numpy, plot_spectrogram_to_numpy
# from plotting_utils import plot_gate_outputs_to_numpy
# class Tacotron2Logger(SummaryWriter):
# def __init__(self, logdir):
# super(Tacotron2Logger, self).__init__(logdir)
# def log_training(self, reduced_loss, grad_norm, learning_rate, duration,
# iteration):
# self.add_scalar("training.loss", reduced_loss, iteration)
# self.add_scalar("grad.norm", grad_norm, iteration)
# self.add_scalar("learning.rate", learning_rate, iteration)
# self.add_scalar("duration", duration, iteration)
# def log_validation(self, reduced_loss, model, y, y_pred, iteration):
# self.add_scalar("validation.loss", reduced_loss, iteration)
# _, mel_outputs, gate_outputs, alignments = y_pred
# mel_targets, gate_targets = y
# # plot distribution of parameters
# for tag, value in model.named_parameters():
# tag = tag.replace('.', '/')
# self.add_histogram(tag, value.data.cpu().numpy(), iteration)
# # plot alignment, mel target and predicted, gate target and predicted
# idx = random.randint(0, alignments.size(0) - 1)
# self.add_image(
# "alignment",
# plot_alignment_to_numpy(alignments[idx].data.cpu().numpy().T),
# iteration)
# self.add_image(
# "mel_target",
# plot_spectrogram_to_numpy(mel_targets[idx].data.cpu().numpy()),
# iteration)
# self.add_image(
# "mel_predicted",
# plot_spectrogram_to_numpy(mel_outputs[idx].data.cpu().numpy()),
# iteration)
# self.add_image(
# "gate",
# plot_gate_outputs_to_numpy(
# gate_targets[idx].data.cpu().numpy(),
# F.sigmoid(gate_outputs[idx]).data.cpu().numpy()),
# iteration)
def log_validation(self, reduced_loss, reduced_losses, reduced_acces, model, y, y_pred, iteration, task):
self.add_scalar('validation.loss.%s'%task, reduced_loss, iteration)
self.add_scalar("validation.loss.%s.recon"%task, reduced_losses[0], iteration)
self.add_scalar("validation.loss.%s.recon_post"%task, reduced_losses[1], iteration)
self.add_scalar("validation.loss.%s.stop"%task, reduced_losses[2], iteration)
self.add_scalar("validation.loss.%s.contr"%task, reduced_losses[3], iteration)
self.add_scalar("validation.loss.%s.consi"%task, reduced_losses[4], iteration)
self.add_scalar("validation.loss.%s.spenc"%task, reduced_losses[5], iteration)
self.add_scalar("validation.loss.%s.spcla"%task, reduced_losses[6], iteration)
self.add_scalar("validation.loss.%s.texcl"%task, reduced_losses[7], iteration)
self.add_scalar("validation.loss.%s.spadv"%task, reduced_losses[8], iteration)
self.add_scalar('validation.acc.%s.spenc'%task, reduced_acces[0], iteration)
self.add_scalar('validation.acc.%s.spcla'%task, reduced_acces[1], iteration)
self.add_scalar('validation.acc.%s.texcl'%task, reduced_acces[2], iteration)
predicted_mel, post_output, predicted_stop, alignments, \
text_hidden, mel_hidden, text_logit_from_mel_hidden, \
audio_seq2seq_alignments, \
speaker_logit_from_mel, speaker_logit_from_mel_hidden, \
text_lengths, mel_lengths, SE_alignments = y_pred
#predicted_mel, post_output, predicted_stop, alignments, \
# text_hidden, mel_hidden, text_logit_from_mel_hidden, \
# audio_seq2seq_alignments, \
# speaker_logit_from_mel_hidden, \
# text_lengths, mel_lengths = y_pred
# text_target, mel_target, spc_target, speaker_target, stop_target = y
text_target, mel_target, speaker_target, stop_target = y
stop_target = stop_target.reshape(stop_target.size(0), -1, int(stop_target.size(1)/predicted_stop.size(1)))
stop_target = stop_target[:,:,0]
#pdb.set_trace()
# plot distribution of parameters
#for tag, value in model.named_parameters():
# tag = tag.replace('.', '/')
# self.add_histogram(tag, value.data.cpu().numpy(), iteration)
# plot alignment, mel target and predicted, stop target and predicted
idx = random.randint(0, alignments.size(0) - 1)
alignments = alignments.data.cpu().numpy()
audio_seq2seq_alignments = audio_seq2seq_alignments.data.cpu().numpy()
SE_alignments = SE_alignments.data.cpu().numpy()
self.add_image(
"%s.alignment"%task,
plot_alignment_to_numpy(alignments[idx].T),
iteration, dataformats='HWC')
# plot more alignments
plot_alignment(alignments[:4], self.ali_path+'/step-%d-%s.pdf'%(iteration, task))
self.add_image(
"%s.audio_seq2seq_alignment"%task,
plot_alignment_to_numpy(audio_seq2seq_alignments[idx].T),
iteration, dataformats='HWC')
self.add_image(
"%s.SE_alignments"%task,
plot_alignment_to_numpy(SE_alignments[idx].T),
iteration, dataformats='HWC')
self.add_image(
"%s.mel_target"%task,
plot_spectrogram_to_numpy(mel_target[idx].data.cpu().numpy()),
iteration, dataformats='HWC')
self.add_image(
"%s.mel_predicted"%task,
plot_spectrogram_to_numpy(predicted_mel[idx].data.cpu().numpy()),
iteration, dataformats='HWC')
# self.add_image(
# "%s.spc_target"%task,
# plot_spectrogram_to_numpy(spc_target[idx].data.cpu().numpy()),
# iteration, dataformats='HWC')
self.add_image(
"%s.post_predicted"%task,
plot_spectrogram_to_numpy(post_output[idx].data.cpu().numpy()),
iteration, dataformats='HWC')
self.add_image(
"%s.stop"%task,
plot_gate_outputs_to_numpy(
stop_target[idx].data.cpu().numpy(),
F.sigmoid(predicted_stop[idx]).data.cpu().numpy()),
iteration, dataformats='HWC')
