-
Notifications
You must be signed in to change notification settings - Fork 8
/
train.py
397 lines (370 loc) · 15.7 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
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
""" NormAE estimator, sklearn style """
import copy
from itertools import chain
from functools import partial
import torch
import torch.nn as nn
import torch.utils.data as data
import torch.optim as optim
import pandas as pd
from tqdm import tqdm
import matplotlib.pyplot as plt
from datasets import ConcatData
from networks import SimpleCoder, OrderLoss
import metrics as mm
from visual import VisObj, pca_for_dict, pca_plot
class BatchEffectTrainer:
def __init__(
self, in_features, batch_label_num, device, pre_transfer, opts
):
# architecture
self.in_features = in_features
self.batch_label_num = batch_label_num
self.device = device
self.encoder_hiddens = opts.ae_encoder_units
self.decoder_hiddens = opts.ae_decoder_units
self.disc_b_hiddens = opts.disc_b_units
self.disc_o_hiddens = opts.disc_o_units
self.bottle_num = opts.bottle_num
self.dropouts = opts.dropouts
# loss
self.use_batch_for_order = opts.use_batch_for_order
self.lambda_b, self.lambda_o = opts.lambda_b, opts.lambda_o
# optimizer
self.lr_rec = opts.lr_rec
self.lr_disc_b = opts.lr_disc_b
self.lr_disc_o = opts.lr_disc_o
# training
self.epoch = sum(opts.epoch)
self.rec_epoch, self.disc_epoch, self.iter_epoch = opts.epoch
self.bs, self.nw = opts.batch_size, opts.num_workers
self.train_with_qc = opts.train_data == "all"
# other
self.visdom_port = opts.visdom_port
self.visdom_env = opts.visdom_env
self.pre_transfer = pre_transfer
# build model
self._build_model()
# training record
self.history = {
'disc_b_loss': [], 'disc_o_loss': [], 'adv_b_loss': [],
'adv_o_loss': [], 'rec_loss': [], 'qc_rec_loss': [],
'qc_distance': []
}
# visdom
self.visobj = VisObj(self.visdom_port, env=self.visdom_env)
# early stop
self.early_stop_objs = {
'best_epoch': -1, 'best_qc_loss': 1000, 'best_qc_distance': 1000,
'best_models': None, 'index': 0, 'best_score': 2000
}
def fit(self, datas):
# get dataloaders
train_data = data.ConcatDataset([datas['subject'], datas['qc']]) \
if self.train_with_qc else datas['subject']
dataloaders = {
'train': data.DataLoader(train_data, batch_size=self.bs,
num_workers=self.nw, shuffle=True),
'qc': data.DataLoader(datas['qc'], batch_size=self.bs,
num_workers=self.nw)
}
# begin training
pbar = tqdm(total=self.epoch)
for e in range(self.epoch):
self.e = e
if e < self.rec_epoch:
self.phase = 'rec_pretrain'
elif e < self.rec_epoch + self.disc_epoch:
self.phase = 'disc_pretrain'
else:
self.phase = 'iter_train'
pbar.set_description(self.phase)
# --- train phase ---
for model in self.models.values():
model.train()
disc_b_loss_obj = mm.Loss()
disc_o_loss_obj = mm.Loss()
adv_b_loss_obj = mm.Loss()
adv_o_loss_obj = mm.Loss()
rec_loss_obj = mm.Loss()
for batch_x, batch_y in tqdm(dataloaders['train'], 'Batch: '):
batch_x = batch_x.to(self.device).float()
batch_y = batch_y.to(self.device).float()
bs0 = batch_x.size(0)
for optimizer in self.optimizers.values():
optimizer.zero_grad()
if self.phase in ['disc_pretrain', 'iter_train']:
disc_b_loss, disc_o_loss = \
self._forward_discriminate(batch_x, batch_y)
disc_b_loss_obj.add(disc_b_loss, bs0)
disc_o_loss_obj.add(disc_o_loss, bs0)
if self.phase in ['rec_pretrain', 'iter_train']:
rec_loss, adv_b_loss, adv_o_loss = \
self._forward_autoencode(batch_x, batch_y)
rec_loss_obj.add(rec_loss, bs0)
adv_b_loss_obj.add(adv_b_loss, bs0)
adv_o_loss_obj.add(adv_o_loss, bs0)
# record loss
self.history['disc_b_loss'].append(disc_b_loss_obj.value())
self.history['disc_o_loss'].append(disc_o_loss_obj.value())
self.history['adv_b_loss'].append(adv_b_loss_obj.value())
self.history['adv_o_loss'].append(adv_o_loss_obj.value())
self.history['rec_loss'].append(rec_loss_obj.value())
# visual epoch loss
self.visobj.add_epoch_loss(
winname='disc_losses',
disc_b_loss=self.history['disc_b_loss'][-1],
disc_o_loss=self.history['disc_o_loss'][-1],
adv_b_loss=self.history['adv_b_loss'][-1],
adv_o_loss=self.history['adv_o_loss'][-1],
)
self.visobj.add_epoch_loss(
winname='recon_losses',
recon_loss=self.history['rec_loss'][-1]
)
# --- valid phase ---
all_data = ConcatData(datas['subject'], datas['qc'])
all_reses_dict, qc_loss = self.generate(
all_data, verbose=False, compute_qc_loss=True)
# pca
subject_pca, qc_pca = pca_for_dict(all_reses_dict, 3)
# plot pca
pca_plot(subject_pca, qc_pca)
# display in visdom
self.visobj.vis.matplot(plt, win='PCA', opts={'title': 'PCA'})
plt.close()
# --- early stopping ---
qc_dist = mm.mean_distance(qc_pca['Rec_nobe'])
self.history['qc_rec_loss'].append(qc_loss)
self.history['qc_distance'].append(qc_dist)
self.visobj.add_epoch_loss(winname='qc_rec_loss', qc_loss=qc_loss)
self.visobj.add_epoch_loss(winname='qc_distance', qc_dist=qc_dist)
if e >= (self.epoch - 200):
self._check_qc(qc_dist, qc_loss)
# progressbar
pbar.update(1)
pbar.close()
# early stop information and save visdom env
if self.visdom_env != 'main':
self.visobj.vis.save([self.visdom_env])
print('')
print('The best epoch is %d' % self.early_stop_objs['best_epoch'])
print('The best qc loss is %.4f' %
self.early_stop_objs['best_qc_loss'])
print('The best qc distance is %.4f' %
self.early_stop_objs['best_qc_distance'])
for k, v in self.models.items():
v.load_state_dict(self.early_stop_objs['best_models'][k])
self.early_stop_objs.pop('best_models')
return self.models, self.history, self.early_stop_objs
def generate(self, data_loader, verbose=True, compute_qc_loss=False):
for m in self.models.values():
m.to(self.device).eval()
if isinstance(data_loader, data.Dataset):
data_loader = data.DataLoader(
data_loader, batch_size=self.bs, num_workers=self.nw)
x_ori, x_rec, x_rec_nobe, ys, codes = [], [], [], [], []
qc_loss = mm.Loss()
# encoding
if verbose:
print('----- encoding -----')
iterator = tqdm(data_loader, 'encode and decode: ')
else:
iterator = data_loader
with torch.no_grad():
for batch_x, batch_y in iterator:
# return x and y
x_ori.append(batch_x)
ys.append(batch_y)
# return latent representation
batch_x = batch_x.to(self.device, torch.float)
batch_y = batch_y.to(self.device, torch.float)
hidden = self.models['encoder'](batch_x)
codes.append(hidden)
# return rec with and without batch effects
batch_ys = [
torch.eye(self.batch_label_num)[batch_y[:, 1].long()].to(
hidden),
batch_y[:, [0]]
]
batch_ys = torch.cat(batch_ys, dim=1)
hidden_be = hidden + self.models['map'](batch_ys)
x_rec.append(self.models['decoder'](hidden_be))
x_rec_nobe.append(self.models['decoder'](hidden))
# return qc loss
if compute_qc_loss:
qc_index = batch_y[:, -1] == 0.
if qc_index.sum() > 0:
batch_qc_loss = self.criterions['rec'](
batch_x[qc_index], x_rec[-1][qc_index])
qc_loss.add(
batch_qc_loss,
qc_index.sum().detach().cpu().item()
)
else:
qc_loss.add(torch.tensor(0.), 0)
# return dataframe
res = {
'Ori': torch.cat(x_ori), 'Ys': torch.cat(ys),
'Codes': torch.cat(codes), 'Rec': torch.cat(x_rec),
'Rec_nobe': torch.cat(x_rec_nobe)
}
for k, v in res.items():
if v is not None:
if k == 'Ys':
res[k] = pd.DataFrame(
v.detach().cpu().numpy(),
index=data_loader.dataset.y_df.index,
columns=data_loader.dataset.y_df.columns
)
elif k != 'Codes':
res[k] = pd.DataFrame(
v.detach().cpu().numpy(),
index=data_loader.dataset.x_df.index,
columns=data_loader.dataset.x_df.columns
)
res[k] = self.pre_transfer.inverse_transform(
res[k], None)[0]
else:
res[k] = pd.DataFrame(
v.detach().cpu().numpy(),
index=data_loader.dataset.x_df.index,
)
if compute_qc_loss:
return res, qc_loss.value()
return res
def load_model(self, model_file):
saved_model = torch.load(model_file)
for k in self.early_stop_objs:
saved_model.pop(k, None)
self.models = saved_model
def _check_qc(self, qc_dist, qc_loss):
early_stop_score = qc_dist + qc_loss * 100
if early_stop_score < self.early_stop_objs['best_score']:
self.early_stop_objs['best_epoch'] = self.e
self.early_stop_objs['best_models'] = {
k: copy.deepcopy(v.state_dict())
for k, v in self.models.items()
}
self.early_stop_objs['best_qc_loss'] = qc_loss
self.early_stop_objs['best_qc_distance'] = qc_dist
self.early_stop_objs['best_score'] = early_stop_score
self.early_stop_objs['index'] = 0
else:
self.early_stop_objs['index'] += 1
def _build_model(self):
logit_dim = self.batch_label_num + 1
# build models
self.models = {
'encoder': SimpleCoder(
[self.in_features] + self.encoder_hiddens +
[self.bottle_num], dropout=self.dropouts[0]
).to(self.device),
'decoder': SimpleCoder(
[self.bottle_num] + self.decoder_hiddens +
[self.in_features], dropout=self.dropouts[1],
final_act=None
).to(self.device),
'map': SimpleCoder(
[logit_dim] + [500] + [self.bottle_num],
).to(self.device),
'disc_b': SimpleCoder(
[self.bottle_num] + self.disc_b_hiddens +
[self.batch_label_num], bn=True, dropout=self.dropouts[2]
).to(self.device),
"disc_o": SimpleCoder(
[self.bottle_num] + self.disc_o_hiddens + [1],
bn=False, dropout=self.dropouts[3]
).to(self.device)
}
# build loss
self.criterions = {
'disc_b': nn.CrossEntropyLoss(),
'disc_o': OrderLoss(),
"rec": nn.L1Loss()
}
# build optim
optimizer_obj = partial(optim.Adam, betas=(0.5, 0.9))
self.optimizers = {
'rec': optimizer_obj(
chain(
self.models['encoder'].parameters(),
self.models['decoder'].parameters(),
self.models['map'].parameters()
), lr=self.lr_rec
),
"disc_b": optimizer_obj(self.models['disc_b'].parameters(),
lr=self.lr_disc_b),
"disc_o": optimizer_obj(self.models['disc_o'].parameters(),
lr=self.lr_disc_o)
}
def _forward_autoencode(self, batch_x, batch_y):
''' autoencode进行训练的部分 '''
res = [None, None, None]
with torch.enable_grad():
all_loss = 0.
hidden = self.models['encoder'](batch_x)
# decoder
batch_ys = [
torch.eye(self.batch_label_num)[batch_y[:, 1].long()].to(
hidden),
batch_y[:, [0]]
]
batch_ys = torch.cat(batch_ys, dim=1)
hidden_be = hidden + self.models['map'](batch_ys)
batch_x_rec = self.models['decoder'](hidden_be)
# reconstruction losses
recon_loss = self.criterions['rec'](batch_x_rec, batch_x)
all_loss += recon_loss
res[0] = recon_loss
if self.phase == "iter_train":
# adversarial regularizations (disc_b)
logit_b = self.models['disc_b'](hidden)
loss_b = self.criterions['disc_b'](logit_b,
batch_y[:, 1].long())
all_loss -= self.lambda_b * loss_b
res[1] = loss_b
# adversarial regularizations (disc_o)
if self.use_batch_for_order:
group = batch_y[:, 1]
else:
group = None
logit_o = self.models['disc_o'](hidden)
loss_o = self.criterions['disc_o'](logit_o, batch_y[:, 0],
group)
all_loss -= self.lambda_o * loss_o
res[2] = loss_o
all_loss.backward()
nn.utils.clip_grad_norm_(
chain(self.models["encoder"].parameters(),
self.models["decoder"].parameters(),
self.models["map"].parameters()),
max_norm=1
)
self.optimizers['rec'].step()
return res
def _forward_discriminate(self, batch_x, batch_y):
with torch.no_grad():
hidden = self.models['encoder'](batch_x)
with torch.enable_grad():
# disc_b
logit_b = self.models['disc_b'](hidden)
adv_b_loss = self.criterions['disc_b'](logit_b,
batch_y[:, 1].long())
adv_b_loss.backward()
nn.utils.clip_grad_norm_(self.models["disc_b"].parameters(),
max_norm=1)
self.optimizers['disc_b'].step()
# disc_o
logit_o = self.models['disc_o'](hidden)
if self.use_batch_for_order:
group = batch_y[:, 1]
else:
group = None
adv_o_loss = self.criterions['disc_o'](
logit_o, batch_y[:, 0], group)
adv_o_loss.backward()
nn.utils.clip_grad_norm_(self.models["disc_b"].parameters(),
max_norm=1)
self.optimizers['disc_o'].step()
return [adv_b_loss, adv_o_loss]