-
Notifications
You must be signed in to change notification settings - Fork 8
/
Copy pathbaseline.py
177 lines (153 loc) · 8.26 KB
/
baseline.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
import sys
sys.path.append('..')
from dataset.dataloaders import UnetInput
from flerken import pytorchfw
from flerken.models import UNet
from flerken.framework.pytorchframework import set_training, config, ctx_iter
from flerken.framework import val
from torch.optim.lr_scheduler import ReduceLROnPlateau
from utils.utils import *
from models.wrapper import Wrapper
from tqdm import tqdm
from loss.losses import *
from collections import OrderedDict
from settings import *
class Baseline(pytorchfw):
def __init__(self, model, rootdir, workname, main_device=0, trackgrad=False):
super(Baseline, self).__init__(model, rootdir, workname, main_device, trackgrad)
self.audio_dumps_path=os.path.join(DUMPS_FOLDER, 'audio')
self.visual_dumps_path = os.path.join(DUMPS_FOLDER, 'visuals')
self.audio_dumps_folder = os.path.join(self.audio_dumps_path, TEST_UNET_CONFIG, 'test')
self.visual_dumps_folder = os.path.join(self.visual_dumps_path, TEST_UNET_CONFIG, 'test')
self.main_device = main_device
self.grid_unwarp = torch.from_numpy(
warpgrid(BATCH_SIZE, NFFT // 2 + 1, STFT_WIDTH, warp=False)).to(self.main_device)
self.val_iterations = 0
def print_args(self):
setup_logger('log_info', self.workdir+'/info_file.txt',
FORMAT="[%(asctime)-15s %(filename)s:%(lineno)d %(funcName)s] %(message)s]")
logger = logging.getLogger('log_info')
self.print_info(logger)
logger.info(f'\r\t Spectrogram data dir: {ROOT_DIR}\r'
'TRAINING PARAMETERS: \r\t'
f'Run name: {self.workname}\r\t'
f'Batch size {BATCH_SIZE} \r\t'
f'Optimizer {OPTIMIZER} \r\t'
f'Initializer {INITIALIZER} \r\t'
f'Epochs {EPOCHS} \r\t'
f'LR General: {LR} \r\t'
f'SGD Momentum {MOMENTUM} \r\t'
f'Weight Decay {WEIGHT_DECAY} \r\t'
f'Pre-trained model: {PRETRAINED} \r'
'MODEL PARAMETERS \r\t'
f'Nº instruments (K) {K} \r\t'
f'U-Net activation: {ACTIVATION} \r\t'
f'U-Net Input channels {INPUT_CHANNELS}\r\t'
f'U-Net Batch normalization {USE_BN} \r\t')
def set_optim(self,*args,**kwargs):
if OPTIMIZER == 'adam':
return torch.optim.Adam(*args,**kwargs)
elif OPTIMIZER == 'SGD':
return torch.optim.SGD(*args,**kwargs)
else:
raise Exception('Non considered optimizer. Implement it')
def hyperparameters(self):
self.dataparallel = False
self.initializer = INITIALIZER
self.EPOCHS = EPOCHS
self.optimizer = self.set_optim(self.model.parameters(), momentum=MOMENTUM, lr=LR)
self.LR = LR
self.scheduler = ReduceLROnPlateau(self.optimizer, patience=7, threshold=3e-4)
def set_config(self):
self.batch_size = BATCH_SIZE
self.criterion = SingleSourceDirectLoss(self.main_device)
@config
@set_training
def train(self):
self.print_args()
validation_data = UnetInput('test')
self.val_loader = torch.utils.data.DataLoader(validation_data,
batch_size=BATCH_SIZE,
shuffle=True,
num_workers=10)
for self.epoch in range(self.start_epoch,self.EPOCHS):
with val(self):
self.run_epoch()
break
def validate_epoch(self):
with tqdm(self.val_loader, desc='Validation: [{0}/{1}]'.format(self.epoch, self.EPOCHS)) as pbar, ctx_iter(
self):
for inputs, visualization in pbar:
self.val_iterations += 1
self.loss_.data.update_timed()
inputs = self._allocate_tensor(inputs)
output = self.model(*inputs) if isinstance(inputs, list) else self.model(inputs)
self.loss = self.criterion(output)
self.tensorboard_writer(self.loss, output, None, self.absolute_iter, visualization)
pbar.set_postfix(loss=self.loss)
self.loss = self.loss_.data.update_epoch(self.state)
self.tensorboard_writer(self.loss, output, None, self.absolute_iter, visualization)
def tensorboard_writer(self, loss, output, gt, absolute_iter, visualization):
text = visualization[1]
self.writer.add_text('Filepath', text[-1], self.val_iterations)
phase = visualization[0].detach().cpu().clone().numpy()
gt_mags_sq, pred_mags_sq, gt_mags, mix_mag, gt_masks, pred_masks = output
if len(text) == BATCH_SIZE:
grid_unwarp = self.grid_unwarp
else: # for the last batch, where the number of samples are generally lesser than the batch_size
grid_unwarp = torch.from_numpy(
warpgrid(len(text), NFFT // 2 + 1, STFT_WIDTH, warp=False)).to(self.main_device)
pred_masks_linear = linearize_log_freq_scale(pred_masks, grid_unwarp)
gt_masks_linear = linearize_log_freq_scale(gt_masks, grid_unwarp)
oracle_spec = (mix_mag * gt_masks_linear)
pred_spec = (mix_mag * pred_masks_linear)
j = ISOLATED_SOURCE_ID
source = SOURCES_SUBSET[j]
for i, sample in enumerate(text):
sample_id = os.path.basename(sample)[:-4]
folder_name = os.path.basename(os.path.dirname(sample))
pred_audio_out_folder = os.path.join(self.audio_dumps_folder, folder_name, sample_id)
create_folder(pred_audio_out_folder)
visuals_out_folder = os.path.join(self.visual_dumps_folder, folder_name, sample_id)
create_folder(visuals_out_folder)
gt_audio = torch.from_numpy(
istft_reconstruction(gt_mags.detach().cpu().numpy()[i][j], phase[i][0], HOP_LENGTH))
pred_audio = torch.from_numpy(
istft_reconstruction(pred_spec.detach().cpu().numpy()[i][0], phase[i][0], HOP_LENGTH))
librosa.output.write_wav(os.path.join(pred_audio_out_folder, 'GT_' + source + '.wav'),
gt_audio.cpu().detach().numpy(), TARGET_SAMPLING_RATE)
librosa.output.write_wav(os.path.join(pred_audio_out_folder, 'PR_' + source + '.wav'),
pred_audio.cpu().detach().numpy(), TARGET_SAMPLING_RATE)
### SAVING MAG SPECTROGRAMS ###
save_spectrogram(gt_mags[i][j].unsqueeze(0).detach().cpu(),
os.path.join(visuals_out_folder, source), '_MAG_GT.png')
save_spectrogram(oracle_spec[i][j].unsqueeze(0).detach().cpu(),
os.path.join(visuals_out_folder, source), '_MAG_ORACLE.png')
save_spectrogram(pred_spec[i][0].unsqueeze(0).detach().cpu(),
os.path.join(visuals_out_folder, source), '_MAG_ESTIMATE.png')
### PLOTTING MAG SPECTROGRAMS ON TENSORBOARD ###
plot_spectrogram(self.writer, gt_mags[:, j].detach().cpu().view(-1, 1, 512, 256)[:8],
self.state + '_GT_MAG', self.val_iterations)
plot_spectrogram(self.writer, (pred_masks_linear * mix_mag).detach().cpu().view(-1, 1, 512, 256)[:8],
self.state + '_PRED_MAG', self.val_iterations )
def main():
os.environ['CUDA_VISIBLE_DEVICES'] = '0,1,2'
# SET MODEL
u_net = UNet([32, 64, 128, 256, 512, 1024, 2048], 1, None, dropout=DROPOUT, verbose=False, useBN=True)
if not os.path.exists(ROOT_DIR):
raise Exception('Directory does not exist')
state_dict = torch.load(TEST_UNET_WEIGHTS_PATH, map_location=lambda storage, loc: storage)
if 'checkpoint' in TEST_UNET_WEIGHTS_PATH:
state_dict = state_dict['state_dict']
new_state_dict = OrderedDict()
for k, v in state_dict.items():
name = k.replace('model.', '')
new_state_dict[name] = v
u_net.load_state_dict(new_state_dict, strict=True)
model = Wrapper(u_net, main_device=MAIN_DEVICE)
work = Baseline(model, ROOT_DIR, PRETRAINED, main_device=MAIN_DEVICE, trackgrad=TRACKGRAD)
work.model_version = 'BASELINE_TESTING'
work.train()
if __name__ == '__main__':
main()
# Usage python3 energy_based.py --train/test