-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_and_eval_siam.py
227 lines (189 loc) · 11.8 KB
/
train_and_eval_siam.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
226
227
import os
import random
import string
import librosa
import matplotlib.pyplot as plt
import noisereduce as nr
import numpy as np
import torch
import torch.nn.functional as F
import torchaudio
import torchsummary
from model import SiameseSiren
from dataset import AudioFileDataset, LibriSpeechDataset
from torch.utils.data import DataLoader
from tqdm import trange
def plot_spectrogram(y, path):
fig = plt.figure()
S = librosa.feature.melspectrogram(y=y, sr=22050)
S_dB = librosa.power_to_db(S, ref=np.max)
img = librosa.display.specshow(S_dB, y_axis='mel', sr=22050)
plt.axis('off')
fig.savefig(path, bbox_inches='tight',transparent=True, pad_inches=0)
def find_value_for_string(string, search_string):
return string[string.find(search_string)+len(search_string):string.find('\n', string.find(search_string))]
def get_model_parameters(ts_result):
total_params = find_value_for_string(ts_result, 'Total params: ')
trainable_params = find_value_for_string(ts_result, 'Trainable params: ')
non_trainable_params = find_value_for_string(ts_result, 'Non-trainable params: ')
params_size = find_value_for_string(ts_result, 'Params size (MB): ')
return total_params, trainable_params, non_trainable_params, params_size
def size_of_model(model):
tmp_name = ''.join(random.choices(string.ascii_uppercase + string.digits, k=9))
np.save(tmp_name+'.npy', np.array(list(model.cpu().state_dict().items())), allow_pickle=True)
pth_size = os.path.getsize(tmp_name+'.npy') * 8 * 0.001
os.remove(tmp_name+'.npy')
# returns size in kilo bits, getsize returns bytes
return pth_size
def save_audio(y, path):
torchaudio.save(path, y, 22050)
def export_to_onnx(torch_model, folder_path):
# Input to the model
x = torch.randn(1, 220500, 1, requires_grad=True)
torch_out = torch_model(x)
# Export the model
torch.onnx.export(torch_model, x, f"{folder_path}/siam.onnx", export_params=True, opset_version=10,
do_constant_folding=True, input_names = ['input'], output_names = ['output'],
dynamic_axes={'input': {0: 'batch_size'}, 'output': {0: 'batch_size'}})
def run_training_and_eval(runs, val_timepoints, val_amplitude, total_steps, result_folder_name, audio_save_interval, img_save_interval, mel_spec_quant):
result_folder_name_img = f'{result_folder_name}_imgs'
folder_path = os.path.join('results', result_folder_name_img)
folder_path_imgs = os.path.join('results', result_folder_name_img)
if not os.path.exists(folder_path):
os.makedirs(folder_path)
if not os.path.exists(folder_path_imgs):
os.makedirs(folder_path_imgs)
plot_spectrogram(val_amplitude.squeeze().detach().cpu().numpy(),
f'{folder_path_imgs}/{result_folder_name}_ground_truth_mel_spectrogram.png')
save_audio(val_amplitude.squeeze(-1).detach().cpu(), f"{folder_path}/{result_folder_name}_ground_truth.wav")
for sweep in range(len(runs)):
hidden_features = runs[sweep]['hidden_features']
num_frq = runs[sweep]['num_frq']
first_omega_0 = runs[sweep]['first_omega_0']
hidden_omega_0 = runs[sweep]['hidden_omega_0']
optimizer = runs[sweep]['optim']
weight_decay = runs[sweep]['weight_decay']
loss_fn = runs[sweep]['loss_fn']
siam_features = runs[sweep]['siam_features']
separate_last_layer = runs[sweep]['separate_last_layer']
audio_siren = SiameseSiren(in_features=1, out_features=1, hidden_features=hidden_features, siam_features=siam_features,
first_omega_0=first_omega_0, hidden_omega_0=hidden_omega_0, outermost_linear=True,
num_frq=num_frq, separate_last_layer=separate_last_layer)
export_to_onnx(audio_siren, folder_path)
audio_siren.cuda()
torchsummary.summary(audio_siren, input_size=(1, 220500, 1))
optim = optimizer(lr=1e-4, params=audio_siren.parameters())
if weight_decay is not None:
optim = optimizer(lr=1e-4, params=audio_siren.parameters(), weight_decay=weight_decay)
best_loss = float('inf')
audio_siren.cuda()
ground_truth = val_amplitude.cuda()
model_input = val_timepoints.cuda()
tr = trange(total_steps, leave=True)
for step in tr:
model_output, coords = audio_siren(model_input)
loss = loss_fn(model_output, ground_truth.repeat(1,1,2))
audio_siren.cuda()
optim.zero_grad()
loss.backward()
optim.step()
if loss < best_loss:
best_loss = loss
torch.save(audio_siren.state_dict(), f"{folder_path}/{result_folder_name}_optimized_siren.pth")
if step+1 in audio_save_interval:
siren = SiameseSiren(in_features=1, out_features=1, hidden_features=hidden_features, siam_features=siam_features,
first_omega_0=first_omega_0, hidden_omega_0=hidden_omega_0, outermost_linear=True,
num_frq=num_frq, separate_last_layer=separate_last_layer)
siren.load_state_dict(torch.load(f"{folder_path}/{result_folder_name}_optimized_siren.pth"))
siren_quant = torch.quantization.quantize_dynamic(siren, {torch.nn.Linear}, dtype=torch.qint8)
siren.cuda()
wave0 = siren(model_input)[0].cpu()[:,:,0]
wave1 = siren(model_input)[0].cpu()[:,:,1]
wave_stereo = siren(model_input)[0].cpu().squeeze()
save_audio(wave0, f"{folder_path}/{result_folder_name}_{step+1}_best_ch0.wav")
save_audio(wave1, f"{folder_path}/{result_folder_name}_{step+1}_best_ch1.wav")
save_audio(wave_stereo, f"{folder_path}/{result_folder_name}_{step+1}_best_stereo.wav")
quant_wav0 = siren_quant(model_input.cpu())[0].cpu()[:,:,0]
quant_wav1 = siren_quant(model_input.cpu())[0].cpu()[:,:,1]
quant_wav_stereo = siren_quant(model_input.cpu())[0].cpu().squeeze()
save_audio(quant_wav0, f"{folder_path}/{result_folder_name}_{step+1}_quant_best_ch0.wav")
save_audio(quant_wav1, f"{folder_path}/{result_folder_name}_{step+1}_quant_best_ch1.wav")
save_audio(quant_wav_stereo, f"{folder_path}/{result_folder_name}_{step+1}_quant_best_stereo.wav")
noise = wave0 - wave1
signal = torch.tensor(nr.reduce_noise(y=wave0.detach().cpu().numpy().squeeze(), sr=22050,
y_noise=noise.detach().cpu().numpy().squeeze(), stationary=True))
save_audio(noise, f"{folder_path}/{result_folder_name}_{step+1}_best_noise.wav")
save_audio(signal.unsqueeze(0), f"{folder_path}/{result_folder_name}_{step+1}_best_nr.wav")
noise = quant_wav0 - quant_wav1
denoised_quant = torch.tensor(nr.reduce_noise(y=quant_wav0.detach().cpu().numpy().squeeze(), sr=22050,
y_noise=noise.detach().cpu().numpy().squeeze(), stationary=True))
save_audio(noise, f"{folder_path}/{result_folder_name}_{step+1}_quant_best_noise.wav")
save_audio(denoised_quant.unsqueeze(0), f"{folder_path}/{result_folder_name}_{step+1}_quant_best_nr.wav")
denoised_stationary_quant = torch.tensor(nr.reduce_noise(y=quant_wav0.detach().cpu().numpy().squeeze(), sr=22050, stationary=True))
denoised_nostationary_quant = torch.tensor(nr.reduce_noise(y=quant_wav0.detach().cpu().numpy().squeeze(), sr=22050))
save_audio(denoised_stationary_quant.unsqueeze(0),
f"{folder_path}/{result_folder_name}_{step+1}_quant_best_nr_stationary.wav")
save_audio(denoised_nostationary_quant.unsqueeze(0),
f"{folder_path}/{result_folder_name}_{step+1}_quant_best_nr_no_stationary.wav")
if step+1 in img_save_interval:
# TODO reusing siren_quant from save_audio, this works as long as save_audio is the same as save_img
if mel_spec_quant:
wave = siren_quant(model_input.cpu())[0].detach().cpu().numpy()
_quant = "_quant"
else:
wave = siren(model_input)[0].detach().cpu().numpy()
_quant = ""
wave0 = wave[:,:,0].squeeze()
wave1 = wave[:,:,1].squeeze()
denoised = nr.reduce_noise(y=wave0, sr=22050, y_noise=wave0-wave1, stationary=True)
plot_spectrogram(wave0-wave1, f'{folder_path_imgs}/{result_folder_name}_{step+1}_noise_estimate.png')
plot_spectrogram(wave0, f'{folder_path_imgs}/{result_folder_name}_{step+1}{_quant}_best_melspec.png')
plot_spectrogram(denoised, f'{folder_path_imgs}/nr_{result_folder_name}_{step+1}{_quant}_best_melspec.png')
plot_spectrogram(denoised_stationary_quant.detach().cpu().numpy(),
f'{folder_path_imgs}/nr_{result_folder_name}_{step+1}{_quant}_best_melspec_stationary_no_estimate.png')
plot_spectrogram(denoised_nostationary_quant.detach().cpu().numpy(),
f'{folder_path_imgs}/nr_{result_folder_name}_{step+1}{_quant}_best_melspec_no_stationary_no_estimate.png')
model_input.cuda()
tr.set_description(f'loss: {loss.item():.8f}')
with open(os.path.join(folder_path, 'configs.txt'), 'w') as f:
f.write("fp32 size "+str(size_of_model(audio_siren))+"\n")
f.write("quant size "+str(size_of_model(siren_quant))+"\n")
total_params, trainable_params, non_trainable_params, params_size = get_model_parameters(torchsummary.summary_string(siren, input_size=(1, 220500, 1))[0])
f.write("total params "+str(total_params)+"\n")
f.write("trainable params "+str(trainable_params)+"\n")
f.write("non trainable params "+str(non_trainable_params)+"\n")
f.write("params size (MB) "+str(params_size)+"\n")
print('total params: ', total_params)
print('trainable params: ', trainable_params)
print('non trainable params: ', non_trainable_params)
print('params size: ', params_size)
if __name__ == '__main__':
runs = []
runs.append(
{
'hidden_features': [256, 256],
'siam_features': [128],
'num_frq': 16,
'first_omega_0': 100,
'hidden_omega_0': 100,
'optim': torch.optim.Adam,
'weight_decay': 1e-5,
'loss_fn': F.mse_loss,
'separate_last_layer': False,
},
)
# how many steps to train and how often to save the results
total_steps = 100000
audio_save_interval = [1000, 5000, 10000, 25000, 50000, 100000]
img_save_interval = [1000, 5000, 10000, 25000, 50000, 100000]
# store spectrogram for quantized model? affects only images, not audio
mel_spec_quant = True
qq = '_quant' if mel_spec_quant else ''
audio_example = 'choice'
audio_path = librosa.ex(audio_example)
result_folder_name = f'{audio_example}_2x256_1x128siam_PE_melspec' + qq
val_loader = DataLoader(AudioFileDataset(audio_path, start_time_sec=0, end_time_sec=10), batch_size=1, pin_memory=True, num_workers=0)
val_sample = next(iter(val_loader))
val_timepoints = val_sample['timepoints'].cuda()
val_amplitude = val_sample['amplitude'].cuda()
run_training_and_eval(runs, val_timepoints, val_amplitude, total_steps, result_folder_name, audio_save_interval, img_save_interval, mel_spec_quant)