-
Notifications
You must be signed in to change notification settings - Fork 0
/
trainer.py
353 lines (296 loc) · 13.3 KB
/
trainer.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
import os
import csv
import torch
import random
import pandas as pd
import pytorch_lightning as pl
from torchaudio.transforms import AmplitudeToDB, MelSpectrogram, Spectrogram
from torchmetrics.classification import Accuracy, AUROC
from utils.scaler import TorchScaler
from utils.metric import batched_preds, compute_test_auc, compute_batch_anomaly_score, represent_extractor, decode_class_label
from utils.data_aug import mixup
class ASDTask(pl.LightningModule):
# 初始化
def __init__(
self,
hparams,
model,
opt,
train_data,
valid_data,
test_data,
scheduler,
fast_dev_run=False,
center_loss=None,
opt_center_loss=None
):
super(ASDTask, self).__init__()
self.hparams.update(hparams)
self.save_hyperparameters(hparams)
try:
self.log_dir = self.logger.log_dir
except Exception as e:
self.log_dir = os.path.join(self.hparams["log_dir"], self.hparams["version"])
os.makedirs(self.log_dir, exist_ok=True)
self.model = model
self.opt = opt
self.scheduler = scheduler
self.train_data = train_data
self.valid_data = valid_data
self.test_data = test_data
self.test_data_label_dict = {v: k for k, v in test_data.label_dict.items()}
self.fast_dev_run = fast_dev_run
if self.fast_dev_run:
self.num_workers = 1
else:
self.num_workers = self.hparams["training"]["num_workers"]
feat_params = self.hparams["feats"]
if self.hparams["feats"]["feature"] == "STFT":
self.feat_transform = Spectrogram(
n_fft=feat_params["n_window"],
win_length=feat_params["n_window"],
hop_length=feat_params["hop_length"],
window_fn=torch.hamming_window,
wkwargs={"periodic": False}
)
elif self.hparams["feats"]["feature"] == "Mel":
self.feat_transform = MelSpectrogram(
sample_rate=feat_params["sample_rate"],
n_fft=feat_params["n_window"],
win_length=feat_params["n_window"],
hop_length=feat_params["hop_length"],
f_min=feat_params["f_min"],
f_max=feat_params["f_max"],
n_mels=feat_params["n_mels"],
window_fn=torch.hamming_window,
wkwargs={"periodic": False},
power=1,
)
else:
raise NotImplementedError
self.scaler = self._init_scaler()
self.supervised_loss = torch.nn.CrossEntropyLoss()
self.center_loss = center_loss
self.opt_center_loss = opt_center_loss
self.embedding_list = []
self.accuracy_calculator = Accuracy()
self.auc_calculator = AUROC()
self.pauc_calculator = AUROC(max_fpr=self.hparams["training"]["max_fpr"])
self.test_buffer = pd.DataFrame()
def _init_scaler(self):
if self.hparams["scaler"]["statistic"] == "instance":
scaler = TorchScaler(
"instance",
self.hparams["scaler"]["normtype"],
self.hparams["scaler"]["dims"],
)
return scaler
elif self.hparams["scaler"]["statistic"] == "dataset":
# we fit the scaler
scaler = TorchScaler(
"dataset",
self.hparams["scaler"]["normtype"],
self.configs["scaler"]["dims"],
)
else:
raise NotImplementedError
if self.hparams["scaler"]["savepath"] is not None:
if os.path.exists(self.hparams["scaler"]["savepath"]):
scaler = torch.load(self.hparams["scaler"]["savepath"])
print(
"Loaded Scaler from previous checkpoint from {}".format(
self.hparams["scaler"]["savepath"]
)
)
return scaler
self.train_loader = self.train_dataloader()
scaler.fit(
self.train_loader,
transform_func=lambda x: self.take_log(self.feat_transform(x[0])),
)
if self.hparams["scaler"]["savepath"] is not None:
torch.save(scaler, self.hparams["scaler"]["savepath"])
print(
"Saving Scaler from previous checkpoint at {}".format(
self.hparams["scaler"]["savepath"]
)
)
return scaler
def take_log(self, mels):
amp_to_db = AmplitudeToDB(stype="amplitude")
amp_to_db.amin = 1e-5 # amin= 1e-5 as in librosa
return amp_to_db(mels).clamp(min=-50, max=80) # clamp to reproduce old code
def detect(self, mel_feats, label, model):
return model(self.scaler(self.take_log(mel_feats)), label)
# 训练
def training_step(self, batch, batch_idx):
"""
Args:
batch: 从 train_dataloader 采样的一个batch的数据
batch_idx: 目前batch的索引
Returns: 返回要反向传播的loss
"""
if self.hparams["represent"]["domain_represent"]:
audio, class_labels, domain_labels = batch
else:
audio, class_labels = batch
features = self.feat_transform(audio)
if self.hparams["training"]["mixup"] and 0.5 > random.random():
features, class_labels = mixup(features, class_labels)
preds, embeddings = self.detect(features, class_labels, self.model)
loss = self.supervised_loss(preds, class_labels)
if self.center_loss is not None:
loss_cent = self.center_loss(embeddings, class_labels)
loss += loss_cent
self.log('train/step', self.global_step)
self.log('train/loss', loss)
self.log('train/lr', self.opt.param_groups[-1]["lr"])
return loss
def on_validation_epoch_start(self):
self.embedding_list = []
for batch in self.train_dataloader():
if self.hparams["represent"]["domain_represent"]:
audio, class_labels, domain_labels = batch
else:
audio, class_labels = batch
domain_labels = None
mels = self.feat_transform(audio.cuda())
self.model.eval()
with torch.no_grad():
_, embedding = self.detect(mels, class_labels.cuda(), self.model)
self.embedding_list.append({'embedding': embedding,
'domain_label': domain_labels,
'class_label': class_labels})
self.represent_embedding_dict = represent_extractor(embedding_list=self.embedding_list,
pooling_type=self.hparams["represent"]["pooling_type"],
domain_represent=self.hparams["represent"]["domain_represent"])
# 校验
def validation_step(self, batch, batch_indx):
"""
Args:
batch:
batch_indx:
Returns: 每1个epoch校验一次,即自动调用validation_step()函数
check_val_every_n_epoch=1
"""
if self.hparams["represent"]["domain_represent"]:
audio, class_labels, anomaly_labels, domain_labels = batch
else:
audio, class_labels, anomaly_labels = batch
domain_labels = None
mels = self.feat_transform(audio)
preds, embedding = self.detect(mels, class_labels, self.model)
detected_embedding_dict = {
'embedding': embedding,
'domain_label': domain_labels,
'class_label': class_labels
}
anomaly_score = compute_batch_anomaly_score(detected_embedding_dict,
self.represent_embedding_dict,
self.hparams["represent"]["score_type"],
self.hparams["represent"]["domain_represent"])
class_labels = class_labels.to(torch.int16)
self.accuracy_calculator.update(preds, class_labels)
self.auc_calculator.update(anomaly_score, anomaly_labels)
self.pauc_calculator.update(anomaly_score, anomaly_labels)
return
def validation_epoch_end(self, outputs):
accuracy = self.accuracy_calculator.compute()
auc = self.auc_calculator.compute()
pauc = self.pauc_calculator.compute()
self.log('valid/accuracy', accuracy)
self.log('valid/auc+pauc', auc + pauc)
self.log('valid/auc', auc, prog_bar=True)
self.log('valid/pauc', pauc)
return auc
def on_save_checkpoint(self, checkpoint):
checkpoint = self.model.state_dict()
return checkpoint
def on_test_epoch_start(self):
self.embedding_list = []
self.train_data.return_domain_label = True
for batch in self.train_dataloader():
audio, class_labels, domain_labels = batch
mels = self.feat_transform(audio.cuda())
self.model.eval()
with torch.no_grad():
_, embedding = self.detect(mels, class_labels.cuda(), self.model)
self.embedding_list.append({'embedding': embedding,
'domain_label': domain_labels,
'class_label': class_labels})
self.represent_embedding_dict = represent_extractor(embedding_list=self.embedding_list,
pooling_type=self.hparams["represent"]["pooling_type"],
domain_represent=self.hparams["represent"]["domain_represent"])
# 测试
def test_step(self, batch, batch_indx):
audio, class_labels, anomaly_labels, domain_labels, filenames = batch
mels = self.feat_transform(audio)
_, embedding = self.detect(mels, class_labels, self.model)
detected_embedding_dict = {
'embedding': embedding,
'domain_label': domain_labels,
'class_label': class_labels
}
anomaly_scores = compute_batch_anomaly_score(detected_embedding_dict,
self.represent_embedding_dict,
self.hparams["represent"]["score_type"],
self.hparams["represent"]["domain_represent"])
batch_predict_df = batched_preds(anomaly_scores,
class_labels,
anomaly_labels,
domain_labels,
filenames,
self.test_data_label_dict)
self.test_buffer = self.test_buffer.append(batch_predict_df)
def on_test_epoch_end(self):
self.test_buffer.columns = ["filename", "anomaly_score", "anomaly_label",
"machine_label", "section_label", "domain_label"]
result_dict = compute_test_auc(self.test_buffer, max_fpr=self.hparams["training"]["max_fpr"])
csv_line = []
auc_pauc = []
for machine, result_df in result_dict.items():
csv_line.append([machine])
csv_line.append(list(result_df.columns))
for _, row in result_df.iterrows():
csv_line.append(list(row.apply(lambda x: format(x, '.2%') if not isinstance(x, str) else x)))
auc_pauc.append('{:.2f}%/{:.2f}%'.format(result_df.loc[result_df['section'] == 'mean', 'auc'].values[0] * 100,
result_df.loc[result_df['section'] == 'mean', 'pauc'].values[0] * 100))
csv_line.append([])
csv_line.append(result_dict.keys())
csv_line.append(['AUC/pAUC'] * len(result_dict.keys()))
csv_line.append(auc_pauc)
csv_save_path = os.path.join(self.log_dir, 'result.csv')
with open(csv_save_path, "w", newline="") as f:
writer = csv.writer(f, lineterminator='\n')
writer.writerows(csv_line)
print(f"The results of the experiment are saved in {self.log_dir}")
def configure_optimizers(self):
if self.opt_center_loss is not None:
return [self.opt, self.opt_center_loss], [self.scheduler]
else:
return [self.opt], [self.scheduler]
def train_dataloader(self):
self.train_loader = torch.utils.data.DataLoader(
self.train_data,
batch_size=self.hparams["training"]["batch_size"],
shuffle=True,
num_workers=self.num_workers
)
return self.train_loader
def val_dataloader(self):
self.valid_loader = torch.utils.data.DataLoader(
self.valid_data,
batch_size=self.hparams["training"]["batch_size_val"],
shuffle=False,
num_workers=self.num_workers
)
return self.valid_loader
def test_dataloader(self):
self.test_loader = torch.utils.data.DataLoader(
self.test_data,
batch_size=self.hparams["training"]["batch_size_val"],
num_workers=self.num_workers,
shuffle=False,
drop_last=False
)
return self.test_loader