-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdtc_stfs_inverse_test.py
225 lines (176 loc) · 8.72 KB
/
dtc_stfs_inverse_test.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
import time
import librosa
import torch
from torch import nn
from torch.nn import functional as F
from model_loader.stft_dataloader import SFTFDataloader
from model_trainer.trainer_specs import ExperimentSpecs
from models.lbfs_mel_Inverse import DTCInverseSTFS
from models.torch_mel_inverse import LibrosaInverseMelScale
def _get_ratio(mat):
return (mat.sum() / mat.numel()).item()
def inverse_test(dataset_name=""):
"""
:return:
"""
trainer_spec = ExperimentSpecs(spec_config='../config.yaml')
model_spec = trainer_spec.get_model_spec().get_spec('spectrogram_layer')
pk_dataset = trainer_spec.get_audio_dataset(dataset_name)
assert 'train_set' in pk_dataset
assert 'validation_set' in pk_dataset
assert 'test_set' in pk_dataset
dataloader = SFTFDataloader(trainer_spec, batch_size=2, verbose=True)
loaders, collate = dataloader.get_loader()
# get all
data_loaders, collate_fn = dataloader.get_all()
_train_loader = data_loaders['train_set']
print(f"Train datasize {dataloader.get_train_dataset_size()}")
n_fft = 1024
n_stft = n_fft // 2 + 1
inverse_mel = LibrosaInverseMelScale(n_stft, f_max=8000.0)
for bidx, batch in enumerate(_train_loader):
text_padded, input_lengths, mel_padded, gate_padded, output_lengths, stft_padded = batch
epsilon = 1e-60
start_time = time.time()
torch_mel_inverse = inverse_mel(mel_padded)
print(f"MEL shape {mel_padded.shape}, "
f"output lengths {output_lengths} "
f"stft target shape {stft_padded.shape} "
f"inverse shape {torch_mel_inverse.shape}")
print(f"Train datasize {dataloader.get_train_dataset_size()}")
print("--- %s load time, seconds ---" % (time.time() - start_time))
torch_mel_inverse = F.pad(torch_mel_inverse, (1, 1), "constant", 0)
mel_padded = F.pad(mel_padded, (1, 1), "constant", 0)
torch_inverse_mel_padded = inverse_mel(mel_padded)
# diff between torch mel and original STFT
torch_relative_diff = torch.abs((torch_mel_inverse - stft_padded) / (torch_mel_inverse + epsilon))
relative_diff_padded = torch.abs(
(torch_inverse_mel_padded - stft_padded) / (torch_inverse_mel_padded + epsilon))
for tol in [1e-1, 1e-3, 1e-5, 1e-10]:
print(f"Ratio of relative diff smaller than {tol:e} is " f"{_get_ratio(torch_relative_diff < tol)}")
for tol in [1e-1, 1e-3, 1e-5, 1e-10]:
print(f"Ratio of relative diff padded than {tol:e} is " f"{_get_ratio(relative_diff_padded < tol)}")
print("-------------------------------------------------------------------------------------------------")
# librosa version numpy
librosa_inverse = librosa.feature.inverse.mel_to_stft(mel_padded.detach().cpu().numpy(), n_fft=1024, sr=22050)
librosa_inverse = torch.from_numpy(librosa_inverse)
diff_librosa = torch.abs((librosa_inverse - stft_padded.numpy()) / (librosa_inverse + epsilon))
for tol in [1e-1, 1e-3, 1e-5, 1e-10]:
print(f"Ratio of relative diff librosa than {tol:e} is " f"{_get_ratio(diff_librosa < tol)}")
print("-------------------------------------------------------------------------------------------------")
# mine
librosa_module = DTCInverseSTFS(1024 // 2 + 1, f_max=8000.0)
x = librosa_module(mel_padded)
assert x.shape == stft_padded.shape == librosa_inverse.shape
my_inverse = torch.abs((x - stft_padded) / (x + epsilon))
for tol in [1e-1, 1e-3, 1e-5, 1e-10]:
print(f"Ratio of relative diff my impl {tol:e} is " f"{_get_ratio(my_inverse < tol)}")
print("my inverse dtype", my_inverse.dtype)
# assert _get_ratio(my_inverse < 1e-1) > 0.2
# assert _get_ratio(my_inverse < 1e-3) > 5e-3
# assert _get_ratio(my_inverse < 1e-5) > 1e-5
# spec = inverse_mel(mel_padded)
break
# # example = train_dataset[1]
# mse = torch.square(melspec - melspec_librosa).mean().item()
# print("Mean Square Difference: ", mse)
def compute_epsilon_deltas(computed_inverse):
for tol in [1e-1, 1e-3, 1e-5, 1e-10]:
print(f"Ratio of relative diff my impl {tol:e} is " f"{_get_ratio(computed_inverse < tol)}")
def inverse_test_combine_error(dataset_name="", epsilon=1e-60, max_iteration=100, verbose=False):
"""
:return:
"""
trainer_spec = ExperimentSpecs(spec_config='../config.yaml')
pk_dataset = trainer_spec.get_audio_dataset(dataset_name)
assert 'train_set' in pk_dataset
assert 'validation_set' in pk_dataset
assert 'test_set' in pk_dataset
dataloader = SFTFDataloader(trainer_spec, batch_size=2, verbose=True)
loaders, collate = dataloader.get_loader()
# get all
data_loaders, collate_fn = dataloader.get_all()
_train_loader = data_loaders['train_set']
print(f"Train datasize {dataloader.get_train_dataset_size()}")
n_fft = 1024
n_stft = n_fft // 2 + 1
inverse_mel = LibrosaInverseMelScale(n_stft, f_max=8000.0)
torch_abs_error = 0
mine_abs_error = 0
librosa_abs_error = 0
torch_abs_error_padded = 0
for bidx, batch in enumerate(_train_loader):
if bidx == max_iteration:
break
text_padded, input_lengths, mel_padded, gate_padded, output_lengths, stft_padded = batch
# start_time = time.time()
torch_mel_inverse = inverse_mel(mel_padded)
torch_mel_inverse = F.pad(torch_mel_inverse, (1, 1), "constant", 0)
mel_padded = F.pad(mel_padded, (1, 1), "constant", 0)
torch_inverse_mel_padded = inverse_mel(mel_padded)
# diff between torch mel and original STFT
torch_relative_diff = torch.abs((torch_mel_inverse - stft_padded) / (torch_mel_inverse + epsilon))
relative_diff_padded = torch.abs(
(torch_inverse_mel_padded - stft_padded) / (torch_inverse_mel_padded + epsilon))
if verbose:
compute_epsilon_deltas(torch_relative_diff)
compute_epsilon_deltas(relative_diff_padded)
torch_abs_error += nn.L1Loss()(torch_mel_inverse, stft_padded).item()
torch_abs_error_padded += nn.L1Loss()(torch_inverse_mel_padded, stft_padded).item()
# librosa version numpy
librosa_inverse = torch.from_numpy(
librosa.feature.inverse.mel_to_stft(
mel_padded.detach().cpu().numpy(), n_fft=1024, sr=22050))
diff_librosa = torch.abs((librosa_inverse - stft_padded.numpy()) / (librosa_inverse + epsilon))
if verbose:
compute_epsilon_deltas(diff_librosa)
librosa_abs_error += nn.L1Loss()(librosa_inverse, stft_padded).item()
# mine
librosa_module = DTCInverseSTFS(1024 // 2 + 1, f_max=8000.0)
x = librosa_module(mel_padded)
assert x.shape == stft_padded.shape == librosa_inverse.shape
if verbose:
compute_epsilon_deltas(x)
mine_abs_error += nn.L1Loss()(x, stft_padded).item()
print(f"torch {torch_abs_error}, torch padded {torch_abs_error_padded}, "
f"librosa {librosa_abs_error}, dtc {mine_abs_error}")
def inverse_test_gpu(dataset_name="", config='config.yaml',
epsilon=1e-60, max_iteration=100,
batch_size=1, verbose=False):
"""
:return:
"""
trainer_spec = ExperimentSpecs(spec_config=config)
pk_dataset = trainer_spec.get_audio_dataset(dataset_name)
assert 'train_set' in pk_dataset
assert 'validation_set' in pk_dataset
assert 'test_set' in pk_dataset
dataloader = SFTFDataloader(trainer_spec, batch_size=batch_size, verbose=True)
data_loaders, collate_fn = dataloader.get_all()
_train_loader = data_loaders['train_set']
n_stft = 1024 // 2 + 1
dts_inverse = DTCInverseSTFS(n_stft, f_max=8000.0).to("cuda")
abs_error = 0
start_time = time.time()
for bidx, batch in enumerate(_train_loader):
if bidx == max_iteration:
break
text_padded, input_lengths, mel_padded, gate_padded, output_lengths, stft_padded = batch
# mine
mel_padded = mel_padded.contiguous().cuda(non_blocking=True)
stft_padded = stft_padded.contiguous().cuda(non_blocking=True)
mel_padded = F.pad(mel_padded, (1, 1), "constant", 0)
x = dts_inverse(mel_padded)
abs_error += nn.L1Loss()(x, stft_padded).item()
print("--- %s load time, seconds ---" % (time.time() - start_time))
print(f"torch {abs_error}")
if __name__ == '__main__':
"""
"""
# test_download()
# test_create_from_numpy_in_memory()
# test_create_from_numpy_and_iterator()
# inverse_test('lj_speech_1k_raw')
# inverse_test_combine_error('lj_speech_1k_raw')
# inverse_test_gpu('lj_speech_1k_raw')
inverse_test_gpu('LJSpeech')