TTS/tests/test_tacotron_model.py

150 lines
6.1 KiB
Python
Raw Normal View History

2018-05-10 22:59:05 +00:00
import os
import copy
import torch
import unittest
from torch import optim
from torch import nn
from TTS.utils.io import load_config
from TTS.layers.losses import L1LossMasked
from TTS.models.tacotron import Tacotron
2018-05-10 22:59:05 +00:00
2019-07-19 09:48:12 +00:00
#pylint: disable=unused-variable
2018-05-10 22:59:05 +00:00
torch.manual_seed(1)
use_cuda = torch.cuda.is_available()
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
file_path = os.path.dirname(os.path.realpath(__file__))
c = load_config(os.path.join(file_path, 'test_config.json'))
def count_parameters(model):
r"""Count number of trainable parameters in a network"""
return sum(p.numel() for p in model.parameters() if p.requires_grad)
class TacotronTrainTest(unittest.TestCase):
@staticmethod
def test_train_step():
input_dummy = torch.randint(0, 24, (8, 128)).long().to(device)
input_lengths = torch.randint(100, 129, (8, )).long().to(device)
input_lengths[-1] = 128
mel_spec = torch.rand(8, 30, c.audio['num_mels']).to(device)
linear_spec = torch.rand(8, 30, c.audio['num_freq']).to(device)
mel_lengths = torch.randint(20, 30, (8, )).long().to(device)
stop_targets = torch.zeros(8, 30, 1).float().to(device)
speaker_ids = torch.randint(0, 5, (8, )).long().to(device)
for idx in mel_lengths:
stop_targets[:, int(idx.item()):, 0] = 1.0
stop_targets = stop_targets.view(input_dummy.shape[0],
stop_targets.size(1) // c.r, -1)
stop_targets = (stop_targets.sum(2) >
0.0).unsqueeze(2).float().squeeze()
2020-01-15 22:10:11 +00:00
criterion = L1LossMasked(seq_len_norm=False).to(device)
criterion_st = nn.BCEWithLogitsLoss().to(device)
model = Tacotron(
num_chars=32,
num_speakers=5,
postnet_output_dim=c.audio['num_freq'],
decoder_output_dim=c.audio['num_mels'],
r=c.r,
memory_size=c.memory_size
).to(device) #FIXME: missing num_speakers parameter to Tacotron ctor
model.train()
print(" > Num parameters for Tacotron model:%s" %
(count_parameters(model)))
model_ref = copy.deepcopy(model)
count = 0
for param, param_ref in zip(model.parameters(),
model_ref.parameters()):
assert (param - param_ref).sum() == 0, param
count += 1
optimizer = optim.Adam(model.parameters(), lr=c.lr)
for _ in range(5):
mel_out, linear_out, align, stop_tokens = model.forward(
input_dummy, input_lengths, mel_spec, speaker_ids)
optimizer.zero_grad()
loss = criterion(mel_out, mel_spec, mel_lengths)
stop_loss = criterion_st(stop_tokens, stop_targets)
loss = loss + criterion(linear_out, linear_spec,
mel_lengths) + stop_loss
loss.backward()
optimizer.step()
# check parameter changes
count = 0
for param, param_ref in zip(model.parameters(),
model_ref.parameters()):
# ignore pre-higway layer since it works conditional
# if count not in [145, 59]:
assert (param != param_ref).any(
), "param {} with shape {} not updated!! \n{}\n{}".format(
count, param.shape, param, param_ref)
count += 1
class TacotronGSTTrainTest(unittest.TestCase):
@staticmethod
def test_train_step():
input_dummy = torch.randint(0, 24, (8, 128)).long().to(device)
2019-03-06 12:43:29 +00:00
input_lengths = torch.randint(100, 129, (8, )).long().to(device)
2019-03-12 08:52:01 +00:00
input_lengths[-1] = 128
mel_spec = torch.rand(8, 120, c.audio['num_mels']).to(device)
linear_spec = torch.rand(8, 120, c.audio['num_freq']).to(device)
mel_lengths = torch.randint(20, 120, (8, )).long().to(device)
stop_targets = torch.zeros(8, 120, 1).float().to(device)
2019-07-19 09:12:48 +00:00
speaker_ids = torch.randint(0, 5, (8, )).long().to(device)
2018-08-02 14:34:17 +00:00
2018-05-10 22:59:05 +00:00
for idx in mel_lengths:
stop_targets[:, int(idx.item()):, 0] = 1.0
2018-08-02 14:34:17 +00:00
stop_targets = stop_targets.view(input_dummy.shape[0],
2018-08-02 14:34:17 +00:00
stop_targets.size(1) // c.r, -1)
stop_targets = (stop_targets.sum(2) >
0.0).unsqueeze(2).float().squeeze()
2018-08-02 14:34:17 +00:00
2020-01-15 22:10:11 +00:00
criterion = L1LossMasked(seq_len_norm=False).to(device)
2019-03-25 23:48:35 +00:00
criterion_st = nn.BCEWithLogitsLoss().to(device)
model = Tacotron(
2019-07-19 09:49:03 +00:00
num_chars=32,
num_speakers=5,
gst=True,
postnet_output_dim=c.audio['num_freq'],
decoder_output_dim=c.audio['num_mels'],
r=c.r,
memory_size=c.memory_size
).to(device) #FIXME: missing num_speakers parameter to Tacotron ctor
2018-05-10 22:59:05 +00:00
model.train()
print(model)
print(" > Num parameters for Tacotron GST model:%s" %
(count_parameters(model)))
2018-05-10 22:59:05 +00:00
model_ref = copy.deepcopy(model)
count = 0
2018-08-02 14:34:17 +00:00
for param, param_ref in zip(model.parameters(),
model_ref.parameters()):
2018-05-10 22:59:05 +00:00
assert (param - param_ref).sum() == 0, param
count += 1
optimizer = optim.Adam(model.parameters(), lr=c.lr)
for _ in range(10):
2018-08-02 14:34:17 +00:00
mel_out, linear_out, align, stop_tokens = model.forward(
input_dummy, input_lengths, mel_spec, speaker_ids)
2018-05-10 22:59:05 +00:00
optimizer.zero_grad()
2018-08-02 14:34:17 +00:00
loss = criterion(mel_out, mel_spec, mel_lengths)
2018-04-30 13:01:02 +00:00
stop_loss = criterion_st(stop_tokens, stop_targets)
2018-08-02 14:34:17 +00:00
loss = loss + criterion(linear_out, linear_spec,
mel_lengths) + stop_loss
2018-05-10 22:59:05 +00:00
loss.backward()
optimizer.step()
# check parameter changes
count = 0
2018-08-02 14:34:17 +00:00
for param, param_ref in zip(model.parameters(),
model_ref.parameters()):
# ignore pre-higway layer since it works conditional
assert (param != param_ref).any(
), "param {} with shape {} not updated!! \n{}\n{}".format(
count, param.shape, param, param_ref)
2019-07-19 06:46:23 +00:00
count += 1