-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
381 lines (352 loc) · 16.4 KB
/
train.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
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
import os, sys
import argparse
import json
import yaml
import torch
import pytorch_lightning as pl
#from neptune.new.integrations.pytorch_lightning import NeptuneLogger
from pytorch_lightning.loggers import WandbLogger
import asteroid
from asteroid import engine
import numpy as np
import pandas as pd
import datasets
import models
import losses
import utils
pl.seed_everything(1, workers=True)
# Keys which are not in the conf.yml file can be added here.
# In the hierarchical dictionary created when parsing, the key `key` can be
# found at dic['main_args'][key]
# By default train.py will use all available GPUs. The `id` option in run.sh
# will limit the number of available GPUs for train.py .
parser = argparse.ArgumentParser()
parser.add_argument(
"--exp_dir",
default="tmp",
help="Full path to save experiment results"
)
def main(conf):
# neptune_logger = NeptuneLogger(name=conf["model"]["name"],
# project=conf["neptune"]["project"],
# api_key=conf["neptune"]["api_key"],
# ) # your credentials
wandb_logger = WandbLogger(name=conf["neptune"]["expname"],project=conf["neptune"]["project"])
dataloader_kwargs = (
{"num_workers": conf["training"]["num_workers"], "pin_memory": True} if "gpu" in arg_dic["main_args"]["device"] else {}
)
train_set = datasets.CombineDatasets(
speech_dirs=conf["data"]["speech_train_dir"],
music_dirs=conf["data"]["music_train_dir"],
sound_dirs=conf["data"]["sound_train_dir"] if "sound_train_dir" in conf["data"] else None,
sample_rate=conf["data"]["sample_rate"],
original_sample_rate=conf["data"]["original_sample_rate"],
segment=conf["data"]["segment"],
shuffle_tracks=True,
multi_speakers=conf["training"]["multi_speakers"],
multi_speakers_frequency=conf["training"]["multi_speakers_frequency"],
data_ratio=conf["training"]["data_ratio"] if "data_ratio" in conf["training"] else 1.,
new_data=conf["training"]["new_data"] if "new_data" in conf["training"] else False,
sound_probability=conf["data"]["sound_probability"] if "sound_probability" in conf["data"] else 0.,
mixwithspeech=conf["data"]["mixwithspeech"] if "mixwithspeech" in conf["data"] else True,
)
val_set = datasets.CombineDatasets(
speech_dirs=conf["data"]["speech_valid_dir"],
music_dirs=conf["data"]["music_valid_dir"],
sound_dirs=conf["data"]["sound_valid_dir"] if "sound_valid_dir" in conf["data"] else None,
sample_rate=conf["data"]["sample_rate"],
original_sample_rate=conf["data"]["original_sample_rate"],
segment=conf["data"]["segment"],
shuffle_tracks=False,
multi_speakers=conf["training"]["multi_speakers"],
multi_speakers_frequency=conf["training"]["multi_speakers_frequency"],
data_ratio=1.,
new_data=False,
sound_probability=conf["data"]["sound_probability"] if "sound_probability" in conf["data"] else 0.,
mixwithspeech=conf["data"]["mixwithspeech"] if "mixwithspeech" in conf["data"] else True,
)
train_loader = torch.utils.data.DataLoader(
train_set,
shuffle=True,
batch_size=conf["training"]["batch_size"],
drop_last=True,
**dataloader_kwargs
)
val_loader = torch.utils.data.DataLoader(
val_set,
shuffle=False,
batch_size=conf["training"]["batch_size"],
drop_last=True,
**dataloader_kwargs
)
###Models
if(conf["model"]["name"] == "ConvTasNet"):
conf["masknet"].update({"n_src": conf["data"]["n_src"]})
model = models.ConvTasNetNorm(
conf["filterbank"],
conf["masknet"],
sample_rate=conf["data"]["sample_rate"],
device=arg_dic["main_args"]["device"]
)
#loss_func = losses.LogL2Time()
plugins = pl.plugins.DDPPlugin(find_unused_parameters=False)
elif (conf["model"]["name"] == "UNet"):
# UNet with logl2 time loss and normalization inside model
model = models.UNet(
conf["data"]["sample_rate"],
conf["data"]["fft_size"],
conf["data"]["hop_size"],
conf["data"]["window_size"],
conf["model"]["kernel_size"],
conf["model"]["stride"],
#sources=conf["data"]["sources"],
device=arg_dic["main_args"]["device"],
mask_logit=conf["model"]["mask_logit"]
)
#loss_func = losses.LogL2Time_weighted(weights=[0.1,0.9])
plugins = pl.plugins.DDPPlugin(find_unused_parameters=False)
elif(conf["model"]["name"] == "OUMX"):
scaler_mean, scaler_std = datasets.get_statistics(conf, val_set)
max_bin = utils.bandwidth_to_max_bin(conf["data"]["sample_rate"], conf["model"]["in_chan"], conf["model"]["bandwidth"])
#scaler_mean, scaler_std = np.array([0]*max_bin), np.array([0.5]*max_bin)
model = models.OUMX(
window_length=conf["model"]["window_length"],
input_mean=scaler_mean,
input_scale=scaler_std,
nb_channels=conf["model"]["nb_channels"],
hidden_size=conf["model"]["hidden_size"],
in_chan=conf["model"]["in_chan"],
n_hop=conf["model"]["nhop"],
sources=conf["data"]["sources"],
max_bin=max_bin,
bidirectional=conf["model"]["bidirectional"],
sample_rate=conf["data"]["sample_rate"],
spec_power=conf["model"]["spec_power"],
)
plugins=None
elif(conf["model"]["name"] == "UNetAttn"):
#scaler_mean, scaler_std = datasets.get_statistics(conf, val_set)
max_bins = utils.bandwidth_to_max_bin(conf["data"]["sample_rate"], conf["model"]["in_chan"], conf["model"]["bandwidth"])
#scaler_mean, scaler_std = np.array([0]*max_bin), np.array([0.5]*max_bin)
model = models.UNetAttn(
window_length=conf["model"]["window_length"],
nb_channels=conf["model"]["nb_channels"],
in_chan=conf["model"]["in_chan"],
n_hop=conf["model"]["nhop"],
max_bins=max_bins,
sample_rate=conf["data"]["sample_rate"],
spec_power=conf["model"]["spec_power"],
n_src=conf["data"]["n_src"],
mask_logit=conf["model"]["mask_logit"],
k=conf["model"]["k"],
hidden_size=conf["model"]["hidden_size"],
device=arg_dic["main_args"]["device"],
)
plugins = pl.plugins.DDPPlugin(find_unused_parameters=False)
elif (conf["model"]["name"] == "Waveunet"):
num_features = [conf["model"]["features"]*i for i in range(1, conf["model"]["levels"]+1)] if conf["model"]["feature_growth"]== "add" else \
[conf["model"]["features"]*2**i for i in range(0, conf["model"]["levels"])]
target_outputs = int(conf["data"]["segment"] * conf["data"]["sample_rate"])
model = models.Waveunet(
num_inputs=conf["model"]["channels"],
num_channels=num_features,
num_outputs=conf["model"]["channels"],
instruments=conf["data"]["sources"],
kernel_size=conf["model"]["kernel_size"],
target_output_size=target_outputs,
depth=conf["model"]["depth"],
strides=conf["model"]["stride"],
conv_type=conf["model"]["conv_type"],
res=conf["model"]["res"],
separate=conf["model"]["separate"],
device=arg_dic["main_args"]["device"],
sample_rate=conf["data"]["sample_rate"]
)
elif (conf["model"]["name"] == "Demucsv2"):
model = models.Demucsv2(
sources=conf["data"]["sources"],
audio_channels=conf["model"]["audio_channels"],
channels=conf["model"]["channels"],
depth=conf["model"]["depth"],
rewrite=True,
glu=conf["model"]["glu"],
rescale=conf["model"]["rescale"],
resample=False,
kernel_size=conf["model"]["kernel_size"],
stride=conf["model"]["stride"],
growth=conf["model"]["growth"],
lstm_layers=conf["model"]["lstm_layers"],
context=conf["model"]["context"],
normalize=True,
samplerate=conf["data"]["sample_rate"],
segment_length=len(conf["data"]["sources"]) * conf["data"]["segment"] * conf["data"]["sample_rate"],
device=arg_dic["main_args"]["device"])
###Losses
loss_module = utils.my_import("losses."+conf["training"]['loss'])
if 'weighted' in conf["training"]['loss']:
loss_func = loss_module(weights=np.array(conf["training"]['class_weights'].split(';'),dtype=np.float32))
elif 'MultiDomain' in conf["training"]['loss']:
loss_func = losses.MultiDomainLoss(
window_length=conf["model"]["window_length"],
in_chan=conf["model"]["in_chan"],
n_hop=conf["model"]["nhop"],
spec_power=conf["model"]["spec_power"],
nb_channels=conf["model"]["nb_channels"],
loss_combine_sources=conf["training"]['loss_combine_sources'],
loss_use_multidomain=conf["training"]['loss_use_multidomain'],
mix_coef=conf["training"]['mix_coef'],
)
else:
loss_func = loss_module()
###Optimizer
optimizer = asteroid.engine.optimizers.make_optimizer(model.parameters(), lr=conf["optim"]["lr"], weight_decay=conf["optim"]["weight_decay"])
if (conf["model"]["name"] == "Waveunet"):
scheduler = torch.optim.lr_scheduler.CyclicLR(
optimizer=optimizer, cycle_momentum=False, base_lr=conf["optim"]["min_lr"], max_lr=conf["optim"]["lr"],step_size_up=conf["optim"]["cycles"],
)
elif conf["training"]["half_lr"]:
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
optimizer=optimizer, factor=conf["optim"]["lr_decay_gamma"], patience=conf["optim"]["lr_decay_patience"], cooldown=10
)
# Just after instantiating, save the args. Easy loading in the future.
exp_dir = os.path.join(conf["main_args"]["exp_dir"],conf["main_args"]['config'].split('.yml')[0])
os.makedirs(exp_dir, exist_ok=True)
conf_path = os.path.join(exp_dir, "conf.yml")
with open(conf_path, "w") as outfile:
yaml.safe_dump(conf, outfile)
### delete paths key which raises error
if "speech_train_dir" in conf["data"]: del conf["data"]["speech_train_dir"]
if "speech_valid_dir" in conf["data"]: del conf["data"]["speech_valid_dir"]
if "music_train_dir" in conf["data"]: del conf["data"]["music_train_dir"]
if "music_valid_dir" in conf["data"]: del conf["data"]["music_valid_dir"]
if "sound_train_dir" in conf["data"]: del conf["data"]["sound_train_dir"]
if "sound_valid_dir" in conf["data"]: del conf["data"]["sound_valid_dir"]
if 'sources' in conf["data"]: conf["data"].pop("sources")
###Systems
# if "system" in conf["training"]:
# system_module = utils.my_import("systems."+conf["training"]['system'])
# system = system_module(
# model=model,
# loss_func=loss_func,
# optimizer=optimizer,
# train_loader=train_loader,
# val_loader=val_loader,
# scheduler=scheduler,
# config=conf,
# val_dur=conf["training"]['val_dur'],
# )
# else:
system = asteroid.engine.system.System(
model=model,
loss_func=loss_func,
optimizer=optimizer,
train_loader=train_loader,
val_loader=val_loader,
scheduler=scheduler,
config=conf
)
# Define callbacks
class shuffleData(pl.callbacks.Callback):
def on_train_epoch_start(self, trainer, pl_module):
trainer.train_dataloader.dataset.datasets.data_shuffle()
class saveSNR(pl.callbacks.Callback):
def on_train_epoch_end(self, trainer, pl_module):
list_music_gain = trainer.train_dataloader.dataset.datasets.list_music_gain
list_mix_snr = trainer.train_dataloader.dataset.datasets.list_mix_snr
list_speech_snr = trainer.train_dataloader.dataset.datasets.list_speech_snr
df = pd.DataFrame(list(zip(list_music_gain, list_mix_snr, list_speech_snr)),columns =['music_gain', 'mix_snr', 'speech_snr'])
df.to_csv(os.path.join(exp_dir,'music_gains.csv'), "w")
callbacks = []
checkpoint_dir = os.path.join(exp_dir, "checkpoints")
checkpoint = pl.callbacks.ModelCheckpoint(
checkpoint_dir,
monitor="val_loss",
mode="min",
save_top_k=5,
verbose=True
)
callbacks.append(checkpoint)
if conf["training"]["early_stop"]:
callbacks.append(pl.callbacks.EarlyStopping(
monitor="val_loss",
mode="min",
patience=conf["optim"]["patience"],
verbose=True
))
callbacks.append(saveSNR())
if "data_ratio" in conf["training"] and conf["training"]["data_ratio"]<1:
callbacks.append(shuffleData())
if conf["main_args"]["load"]:
checkpoint_files = [os.path.join(checkpoint_dir, x) for x in os.listdir(checkpoint_dir) if x.endswith(".ckpt")]
if len(checkpoint_files)>0:
newest_checkpoint = max(checkpoint_files, key = os.path.getctime)
else:
newest_checkpoint = None
else:
newest_checkpoint = None
# Don't ask GPU if they are not available.
gpus = -1 if torch.cuda.is_available() and not conf["main_args"]["disable_cuda"] else None
#distributed_backend = "ddp" if torch.cuda.is_available() and not conf["main_args"]["disable_cuda"] else None
distributed_backend = "ddp" if torch.cuda.is_available() and not conf["main_args"]["disable_cuda"] else None
accelerator = "gpu" if torch.cuda.is_available() and not conf["main_args"]["disable_cuda"] else 'cpu'
trainer = pl.Trainer(
max_epochs=conf["training"]["epochs"],
callbacks=callbacks,
default_root_dir=exp_dir,
gpus=gpus,
gradient_clip_val=5.0,
resume_from_checkpoint=newest_checkpoint,
precision=32,
logger= wandb_logger,
#plugins=plugins,
#limit_train_batches=1., # Useful for fast experiment
#strategies=distributed_backend,
#accelerator=distributed_backend,
reload_dataloaders_every_n_epochs=1 if "new_data" in conf["training"] and conf["training"]["new_data"] else 0
)#logger=neptune_logger
trainer.fit(system)
best_k = {k: v.item() for k, v in checkpoint.best_k_models.items()}
with open(os.path.join(exp_dir, "best_k_models.json"), "w") as f:
print(best_k,f)
json.dump(best_k, f, indent=0)
print(checkpoint.best_model_path)
state_dict = torch.load(checkpoint.best_model_path)
system.load_state_dict(state_dict=state_dict["state_dict"])
system.cpu()
to_save = system.model.serialize()
to_save.update(train_set.get_infos())
torch.save(to_save, os.path.join(exp_dir, "best_model.pth"))
#train_set.list_music_gain
#train_set.list_mix_snr
if __name__ == "__main__":
# We start with opening the config file conf.yml as a dictionary from
# which we can create parsers. Each top level key in the dictionary defined
# by the YAML file creates a group in the parser.
parser.add_argument(
"--config", type=str, required=True, help="the config file for the experiments"
)
parser.add_argument('--disable-cuda', action='store_true',help='Disable CUDA')
parser.add_argument(
"--load",
type=bool,
default=False,
help="restore the most recent checkpoint with .ckpt extension"
)
config_model = sys.argv[2]
with open(config_model) as f:
def_conf = yaml.safe_load(f)
parser = asteroid.utils.prepare_parser_from_dict(def_conf, parser=parser)
# Arguments are then parsed into a hierarchical dictionary (instead of
# flat, as returned by argparse) to facilitate calls to the different
# asteroid methods (see in main).
# plain_args is the direct output of parser.parse_args() and contains all
# the attributes in an non-hierarchical structure. It can be useful to also
# have it so we included it here but it is not used.
arg_dic, plain_args = asteroid.utils.parse_args_as_dict(parser, return_plain_args=True)
arg_dic["main_args"]["device"] = {}
if not arg_dic["main_args"]["disable_cuda"] and torch.cuda.is_available():
arg_dic["main_args"]["device"] = 'cuda'
else:
arg_dic["main_args"]["device"] = 'cpu'
main(arg_dic)
## CUDA_VISIBLE_DEVICES=0,1 python train.py --config cfg/UNet_config.yml