-
Notifications
You must be signed in to change notification settings - Fork 0
/
solver.py
385 lines (333 loc) · 15.9 KB
/
solver.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
import torch
import wandb
import time
import datetime
from torch import nn
import torch.nn.functional as F
from evaluation import retrieval_evaluation, QR_evaluation, attribute_evaluation, L1_gen_evaluation
class Solver(object):
def __init__(self, opts, logfile, val_logfile, device, train_dataloader_list,
models):
self.opts = opts
self.logfile = logfile
self.val_logfile = val_logfile
self.device = device
self.num_dataset = len(train_dataloader_list)
self.num_steps_per_epoch = sum([len(tt) for tt in train_dataloader_list])
self.char_num = train_dataloader_list[0].dataset.char_num
self.fontemb_net = models['fontemb_net']
self.attrregressor_net = models['attrcls_net']
self.fontcls_net = models['fontcls_net']
self.charcls_net = models['charcls_net']
self.fontdec_net = models['fontdec_net']
# Loss criterion_sim
if opts.simclr:
from network.nt_xent import NT_Xent
# see appendix B.7.: temperature under different batch sizes
self.criterion_sim = NT_Xent(opts.batch_size, opts.temperature, 1).to(device)
elif opts.supcon:
from network.losses import SupConLoss
# see appendix B.7.: temperature under different batch sizes
self.criterion_sim = SupConLoss(temperature=opts.temperature).to(device)
else:
from network.losses import PUILoss
self.criterion_sim = PUILoss(device, lamda=2.0)
if opts.train_fontcls or opts.train_charcls and not opts.supcon:
self.criterion_ce = torch.nn.CrossEntropyLoss().to(device) ## used for font-classification / char-classification
if opts.train_ae or opts.train_cae:
self.criterion_pixel = torch.nn.L1Loss().to(device)
if opts.train_attr:
attr_loss_type = "BCE"
if attr_loss_type == "MSE":
self.criterion_attr = torch.nn.MSELoss(reduction = 'none').to(device)
def attr_logit(x):
return x
elif attr_loss_type == "BCE":
self.criterion_attr = torch.nn.BCEWithLogitsLoss(reduction = 'none').to(device)
sigmoid = nn.Sigmoid().to(device)
sigmoid.eval()
def attr_logit(x):
return sigmoid(x)
self.attr_logit = attr_logit
self.criterion_attr_eval = torch.nn.MSELoss(reduction = 'none').to(device)
self.criterion_attr_eval.eval()
# optimizers
params = [{"params": m.parameters()} for k,m in models.items() if m]
# params = [{"params": m.parameters()} for k,m in models.items() if not k in ['fontemb_net', 'fontcls_net'] and m ]
# params += [{"params": self.fontemb_net.parameters(), "lr":opts.lr_fontembcls}]
self.optimizer = torch.optim.Adam(params, lr=opts.lr, betas=(opts.b1, opts.b2))
self.best_ret_accuracy = 0
def train_batch(self, batch, dataset_i):
self.prev_time = time.time()
opts = self.opts
image, charclass, fontclass, attr_data, label = self.load_batch(batch) ## image, charclass could be tuple
feature = self.train_fontemb(image, charclass, fontclass, dataset_i)
if opts.train_attr:
self.train_attr(feature, attr_data, label)
# backward
self.optimizer.zero_grad()
self.loss.backward()
self.optimizer.step()
def load_batch(self, batch):
device = self.device
if self.opts.data_type == '2glyphs':
img_i = batch['img_i'].to(device)
charclass_i = batch['charclass_i'].to(device)
img_j = batch['img_j'].to(device)
charclass_j = batch['charclass_j'].to(device)
image = (img_i, img_j)
charclass = (charclass_i, charclass_j)
elif self.opts.data_type == '1glyph':
image = batch['img_i'].to(device)
charclass = batch['charclass_i'].to(device)
fontclass = batch['fontclass'].to(device)
attr_data = batch['attr'].to(device)
self.label = batch['label_A'].to(device)
return image, charclass, fontclass, attr_data, self.label
def train_fontemb(self, image, charclass, fontclass, dataset_i):
"""
forward pass of fontemb and compute loss
"""
if self.opts.data_type == '2glyphs':
img_i = image[0]
img_j = image[1]
charclass_i = charclass[0]
charclass_j = charclass[1]
loss = torch.zeros(1).to(self.device)
self.loss_dict = {}
## infer i
feat_i, output_i = self.fontemb_net(img_i)
if self.opts.simclr or self.opts.supcon:
"""
This trains model with contrastivel learning
"""
## model forward & backward
## infer j
feat_j, output_j = self.fontemb_net(img_j)
feature = (feat_i, feat_j)
z_i = output_i[dataset_i] # dataset_i used as index of head
z_j = output_j[dataset_i] # if opts.simclr: use simclr projection, else: use pui softmax head
if self.opts.simclr:
nt_xent_loss = self.criterion_sim(z_i, z_j)
else:
z_i = F.normalize(z_i, dim=1)
z_j = F.normalize(z_j, dim=1)
features = torch.cat([z_i.unsqueeze(1), z_j.unsqueeze(1)], dim=1)
if self.opts.train_fontcls:
print("fontcls for supcon but not effective.")
nt_xent_loss = self.criterion_sim(features, fontclass.view(-1))
else:
nt_xent_loss = self.criterion_sim(features)
self.loss_dict["loss_nt_xent"] = nt_xent_loss
loss += nt_xent_loss
if self.opts.train_cae:
feature = (feat_i, None)
if self.opts.train_fontcls:
font_out = self.fontcls_net(feat_i)
cls_loss = self.criterion_ce(font_out, fontclass.view(-1)) * self.opts.lambda_fontcls
self.loss_dict["loss_cls"] = cls_loss
loss += cls_loss
# feat_i, _ = self.fontemb_net(img_i)
if self.opts.grad_cut_fontdec:
fontdec_input = feat_i.detach()
else:
fontdec_input = feat_i
char_num = self.char_num
one_hot = torch.nn.functional.one_hot(charclass_j, num_classes=char_num).to(self.device) # torch.Size([64, 1, 52])
feat_char_cat = torch.cat([fontdec_input, one_hot.squeeze(1)], dim=1)
out = self.fontdec_net(feat_char_cat)
pixel_loss = self.criterion_pixel(out, img_j)
self.loss_dict["loss_pixel"] = pixel_loss
loss += pixel_loss
self.loss_dict["loss"] = loss
elif self.opts.data_type == '1glyph':
"""
This mode classifies or autoencode font. Technically, can be joint trained but we dont try
"""
if self.opts.train_fontcls:
feature, _ = self.fontemb_net(image)
font_out = self.fontcls_net(feature)
loss = self.criterion_ce(font_out, fontclass.view(-1)) * self.opts.lambda_fontcls
self.loss_dict = {"loss": loss}
if self.opts.train_ae:
"""
This mode classifies or autoencode font.
"""
# image_feature = self.fontemb_net.image_encode(image)
# out, feature = self.fontdec_net(image_feature)
feature, _ = self.fontemb_net(image)
out = self.fontdec_net(feature)
loss = self.criterion_pixel(out, image)
self.loss_dict = {"loss": loss}
if not self.opts.train_fontcls and not self.opts.train_ae and self.opts.train_attr:
'''
train attribute only
'''
feature, _ = self.fontemb_net(image)
loss = torch.zeros(1).to(self.device)
self.loss_dict = {"loss": loss}
self.loss = self.loss_dict['loss']
return feature
def train_attr(self, feature, attr_data, label):
# if self.opts.simclr or self.opts.supcon:
if self.opts.data_type == '2glyphs':
feat_i = feature[0]
feat_j = feature[1]
attr_i = self.attrregressor_net(feat_i)
attr_j = self.attrregressor_net(feat_j)
loss_i = self.criterion_attr(attr_i, attr_data)
loss_j = self.criterion_attr(attr_j, attr_data)
loss_attr = loss_i + loss_j
else:
attr_i = self.attrregressor_net(feature)
loss_attr = self.criterion_attr(attr_i, attr_data)
loss_attr= loss_attr * label
if label.sum() > 0:
# no supervised attr in batch
loss_attr = loss_attr.sum()/label.sum()/attr_data.size(1)
self.loss += loss_attr
self.loss_dict["attr_loss"] = loss_attr
def log_results(self, epoch, batch_idx, dataset_i):
opts = self.opts
batches_done = (epoch - opts.init_epoch) * self.num_steps_per_epoch + batch_idx
batches_left = (opts.n_epochs - opts.init_epoch) * self.num_steps_per_epoch - batches_done
time_left = datetime.timedelta(seconds=batches_left*(time.time() - self.prev_time))
self.prev_time = time.time()
message = (
f"Epoch: {epoch}/{opts.n_epochs}, Dataset: {dataset_i}/{self.num_dataset}, "
f"Batch: {batch_idx}/{self.num_steps_per_epoch}, ETA: {time_left}, "
f"loss_total: {self.loss.item():.6f}, "
)
self.curr_step = (epoch - 1)*self.num_steps_per_epoch + batch_idx + 1
wandb_dict = self.loss_dict
wandb_dict['step'] = self.curr_step
if opts.train_attr:
if self.label.sum() > 0: # no logging when no attr label
loss_attr = self.loss_dict["attr_loss"]
message += f"loss_attr: {loss_attr.item():.6f}"
if opts.use_wandb:
wandb.log(wandb_dict)
print(message)
self.logfile.write(message + '\n')
self.logfile.flush()
def infer_model(self, test_dataloader):
char_num = test_dataloader.dataset.char_num
feat_per_font = []
attr_per_font = []
attrgt_per_font = []
self.fontemb_net.eval()
device = next(self.fontemb_net.parameters()).device # NOTE: assume model in one gpu
do_attr_extraction = (self.attrregressor_net and test_dataloader.dataset.use_attr)
with torch.no_grad():
## load data for each font with batchsize=charnum
for ii, batch_test in enumerate(test_dataloader):
img_i = batch_test['img_i'].to(device)
feat_i, _ = self.fontemb_net(img_i)
feat_per_font.append(feat_i)
if do_attr_extraction:
attr_gt = batch_test['attr'].to(device)
attr_i = self.attrregressor_net(feat_i)
attr_i = self.attr_logit(attr_i)
attr_per_font.append(attr_i)
attrgt_per_font.append(attr_gt)
else:
attr_i = None
attr_gt = None
feat_per_font = torch.stack(feat_per_font) ## torch.Size([28, 52, 512])
feat_per_char = feat_per_font.permute(1,0,2) # torch.Size([52, 28, 512])
if do_attr_extraction:
attr_per_font = torch.stack(attr_per_font) ## torch.Size([28, 52, 512])
attr_per_char = attr_per_font.permute(1,0,2) # torch.Size([52, 28, 512])
attr_gt = torch.stack(attrgt_per_font)[:,0,:] ## torch.Size([28, 37]) ## donovan
else:
attr_per_char = []
print("Extract feat done: {}".format(len(feat_per_char)))
return feat_per_char, attr_per_char, attr_gt
def evaluate(self, epoch, test_dataloader):
opts = self.opts
self.fontemb_net.eval()
if self.fontdec_net:
self.fontdec_net.eval()
fontemb_net = self.fontemb_net
attrregressor_net = self.attrregressor_net
with torch.no_grad():
"""
TEST font retrieval
use init_char to test all, or capitals
"""
t = time.time()
feat_per_char, attr_per_char, attr_gt = self.infer_model(test_dataloader)
ret_accuracy, ret_per_char = retrieval_evaluation(feat_per_char, test_dataloader, init_char=0)
# ret_accuracy_upper, ret_per_char_upper = retrieval_evaluation(feat_per_char, test_dataloader, init_char=26)
if ret_accuracy > self.best_ret_accuracy:
self.best_ret_accuracy = ret_accuracy
if opts.use_wandb:
wandb.log({
"best retrieval accuracy" : self.best_ret_accuracy,
"best epoch" : epoch,
})
if opts.use_wandb:
wandb_dict = {}
wandb_dict["epoch"] = epoch
wandb_dict["step"] = self.curr_step
wandb_dict["retrieval accuracy_all"] = ret_accuracy
# wandb_dict["retrieval accuracy_upper"] = ret_accuracy_upper
wandb.log(wandb_dict)
ret_per_char["epoch"] = epoch
ret_per_char["step"] = self.curr_step
wandb.log(ret_per_char)
message = (
f"Evaluation on testset took {(time.time() - t)/60:.2f} minutes, "
f"Retrieval accuracy: {ret_accuracy:.6f}, "
)
print(message)
self.val_logfile.write(message + '\n')
self.val_logfile.flush()
"""
L1 generation measure
"""
if self.opts.train_cae and epoch % self.opts.check_L1_gen_freq == 0 :
t = time.time()
L1_error = L1_gen_evaluation(
self.fontemb_net,
self.fontdec_net,
test_dataloader, True)
message = (
f"Evaluation (L1 generation) on testset took {(time.time() - t)/60:.2f} minutes, "
f"L1 generation error: {L1_error:.6f}, "
)
print(message)
if opts.use_wandb:
wandb.log({
'epoch' : epoch,
'L1-image-error_{}'.format(test_dataloader.dataset.dataset_name) : L1_error,
})
"""
Attribute loss
"""
if opts.train_attr:
t = time.time()
attrregressor_net.eval()
error_attr, attr_error_dict, char_error_dict, message = \
attribute_evaluation(attr_per_char, attr_gt,
test_dataloader,
self.criterion_attr_eval)
message += (f"Evaluation on testset took {(time.time() - t)/60:.2f} minutes,"
f"total attr_error : {error_attr.item():.6f}\n")
if opts.use_wandb:
attr_error_dict["epoch"] = epoch
attr_error_dict["step"] = self.curr_step
attr_error_dict["attr_error"] = error_attr.item()
char_error_dict["epoch"] = epoch
char_error_dict["step"] = self.curr_step
wandb.log(attr_error_dict)
wandb.log(char_error_dict)
print(message)
self.val_logfile.write(message + '\n')
self.val_logfile.flush()
# train()
attrregressor_net.train()
# train()
fontemb_net.train()
if self.fontdec_net:
self.fontdec_net.train()