-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathtrainer.py
322 lines (266 loc) · 13.9 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
import numpy as np
import random
from utils import *
from logger import AverageMeter
import time
from calculate_error import *
import torch
from torch.autograd import Variable
from torchvision.utils import save_image
import csv
import os
import imageio
from tqdm import tqdm
from path import Path
import warnings
warnings.filterwarnings(action='ignore')
def validate(args, val_loader, model, decoder, logger, dataset = 'KITTI'):
##global device
batch_time = AverageMeter()
if dataset == 'KITTI':
error_names = ['abs_diff', 'abs_rel', 'sq_rel', 'a1', 'a2', 'a3','rmse','rmse_log','log10']
elif dataset == 'NYU':
error_names = ['abs_diff', 'abs_rel', 'sq_rel', 'log10', 'a1', 'a2', 'a3','rmse','rmse_log']
errors = AverageMeter(i=len(error_names))
# switch to evaluate mode
model.eval()
end = time.time()
logger.valid_bar.update(0)
for i, (rgb_data, gt_data) in enumerate(val_loader):
if gt_data.ndim != 4 and gt_data[0] == False:
continue
end = time.time()
rgb_data = rgb_data.cuda()
gt_data = gt_data.cuda()
# compute output
input_img = F.interpolate(rgb_data, scale_factor=0.5, mode='bilinear')
with torch.no_grad():
feat_list = model(input_img)
_,output_depth = decoder(feat_list)
batch_time.update(time.time() - end)
output_depth = F.interpolate(output_depth, scale_factor=2, mode='bilinear')
if dataset == 'KITTI':
err_result = compute_errors(gt_data, output_depth,crop=True, cap=args.cap)
elif dataset == 'NYU':
err_result = compute_errors_NYU(gt_data, output_depth,crop=True)
elif dataset == 'Make3D':
err_result = compute_errors_Make3D(depth, output_depth)
errors.update(err_result)
# measure elapsed time
logger.valid_bar.update(i+1)
if i % args.print_freq == 0:
logger.valid_writer.write('valid: Time {} Abs Error {:.4f} ({:.4f})'.format(batch_time, errors.val[0], errors.avg[0]))
logger.valid_bar.update(len(val_loader))
return errors.avg,error_names
def train_single(args,model,decoder, optimizer, dataset_loader,val_loader, batch_size, n_epochs,lr,logger):
num = 0
model_num = 0
encoder_save_dir = './' + args.dataset + '_KDDN_' + args.model + '_b'\
+ str(args.batch_size) + '_single_encoder'
decoder_save_dir = './' + args.dataset + '_KDDN_' + args.model + '_b'\
+ str(args.batch_size) + '_single_decoder'
if (args.rank == 0):
print("Training for %d epochs..." % n_epochs)
if not os.path.exists(encoder_save_dir):
os.makedirs(encoder_save_dir)
os.makedirs(decoder_save_dir)
test_loss_dir = Path(args.save_path)
test_loss_dir_rmse = str(test_loss_dir/'test_rmse_list.txt')
test_loss_dir = str(test_loss_dir/'test_loss_list.txt')
train_loss_dir = Path(args.save_path)
train_loss_dir_rmse = str(train_loss_dir/'train_rmse_list.txt')
a1_acc_dir = str(train_loss_dir/'a1_acc_list.txt')
train_loss_dir = str(train_loss_dir/'train_loss_list.txt')
loss_pdf = "train_loss.pdf"
rmse_pdf = "train_rmse.pdf"
a1_pdf = "train_a1.pdf"
loss_list = []
rmse_list = []
train_loss_list = []
train_rmse_list = []
a1_acc_list = []
num_cnt = 0
train_loss_cnt = 0
loss_sum = 0
n_iter = 0
iter_per_epoch = len(dataset_loader)
base_lr = args.lr
end_lr = args.end_lr
total_iter = n_epochs * iter_per_epoch
################ train mode ####################
model.train()
decoder.train()
################################################
for epoch in tqdm(range(n_epochs+5)):
#dataset_loader.sampler.set_epoch(epoch)
random.seed(epoch)
np.random.seed(epoch) # numpy-based func random setting
torch.manual_seed(epoch) # cpu operation random setting
torch.cuda.manual_seed(epoch) # gpu operation random setting
torch.cuda.manual_seed_all(epoch) # multi-gpu operation random setting
####################################### one epoch training #############################################
for i, (rgb_data, gt_data) in enumerate(dataset_loader):
# get the inputs
inputs = rgb_data
depths = gt_data
inputs = F.interpolate(inputs, scale_factor=0.5, mode='bilinear')
inputs = inputs.cuda()
depths = depths.cuda()
# wrap them in Variable
inputs, depths = Variable(inputs), Variable(depths)
'''Network loss'''
# Feed-forward pass
feat_list = model(inputs)
feat_list_, outputs = decoder(feat_list)
outputs = F.interpolate(outputs, scale_factor=2, mode='bilinear')
##################################### Valid mask definition ####################################
# masking valied area
valid_mask = make_mask(depths, args.dataset)
valid_out = outputs[valid_mask]
valid_gt_sparse = depths[valid_mask]
###################################### scale invariant loss #####################################
scale_inv_loss = scale_invariant_loss(valid_out, valid_gt_sparse)
#################################################################################################
loss = scale_inv_loss
# zero the parameter gradients and backward & optimize
optimizer.zero_grad()
loss.backward()
if n_iter == total_iter:
current_lr = end_lr
else:
current_lr = (base_lr - end_lr) * (1 - n_iter / total_iter) ** 0.5 + end_lr
n_iter += 1
optimizer.param_groups[0]['lr'] = current_lr
optimizer.param_groups[1]['lr'] = current_lr
optimizer.step()
if ((i+1) % 100 == 0):
if (args.rank == 0):
print("epoch: %d, %d/%d"%(epoch+1,i+1,args.epoch_size))
print("[%6d/%6d] total: %.5f, scale_inv: %.5f"%(n_iter, total_iter, loss.item(),scale_inv_loss.item()))
total_loss = loss.item()
rmse_loss = (torch.sqrt(torch.pow(valid_out-valid_gt_sparse,2))).mean()
rmse_loss = rmse_loss.item()
train_loss_cnt = train_loss_cnt + 1
all_plot(args.save_path,total_loss, rmse_loss, train_loss_list, train_rmse_list, train_loss_dir,train_loss_dir_rmse,loss_pdf, rmse_pdf, train_loss_cnt,True)
if (args.rank == 0):
print(" learning decay... current lr: %.6f"%(current_lr))
torch.save(model.state_dict(), encoder_save_dir+'/epoch_%02d_encoder_loss_%.4f.pkl' %(model_num+1,loss))
torch.save(decoder.state_dict(), decoder_save_dir+'/epoch_%02d_decoder_loss_%.4f.pkl' %(model_num+1,loss))
model_num = model_num + 1
return loss
def train_student(args,Teacher,decoder,Student, optimizer, dataset_loader,val_loader, batch_size, n_epochs,lr,logger):
num = 0
model_num = 0
save_dir = './' + args.dataset + '_KDDN_' + args.model + '_b'\
+ str(args.batch_size)
if (args.rank == 0):
print("Training for %d epochs..." % n_epochs)
if not os.path.exists(save_dir):
os.makedirs(save_dir)
test_loss_dir = Path(args.save_path)
test_loss_dir_rmse = str(test_loss_dir/'test_rmse_list.txt')
test_loss_dir = str(test_loss_dir/'test_loss_list.txt')
train_loss_dir = Path(args.save_path)
train_loss_dir_rmse = str(train_loss_dir/'train_rmse_list.txt')
a1_acc_dir = str(train_loss_dir/'a1_acc_list.txt')
train_loss_dir = str(train_loss_dir/'train_loss_list.txt')
loss_pdf = "train_loss.pdf"
rmse_pdf = "train_rmse.pdf"
a1_pdf = "train_a1.pdf"
loss_list = []
rmse_list = []
train_loss_list = []
train_rmse_list = []
a1_acc_list = []
num_cnt = 0
train_loss_cnt = 0
n_iter = 0
iter_per_epoch = len(dataset_loader)
base_lr = args.lr
end_lr = args.end_lr
total_iter = n_epochs * iter_per_epoch
total_iter_drop = (n_epochs-25) * iter_per_epoch
keep_prob = 1
################ train mode ####################
Student.train()
Teacher.eval()
decoder.eval()
################################################
for epoch in tqdm(range(n_epochs+5)):
dataset_loader.sampler.set_epoch(epoch)
random.seed(epoch)
np.random.seed(epoch) # numpy-based func random setting
torch.manual_seed(epoch) # cpu operation random setting
torch.cuda.manual_seed(epoch) # gpu operation random setting
torch.cuda.manual_seed_all(epoch) # multi-gpu operation random setting
####################################### one epoch training #############################################
for i, (rgb_data, gt_data) in enumerate(dataset_loader):
# get the inputs
inputs = rgb_data
depths = gt_data
inputs = inputs.cuda()
inputs = F.interpolate(inputs, scale_factor=0.5, mode='bilinear')
depths = depths.cuda()
# wrap them in Variable
inputs, depths = Variable(inputs), Variable(depths)
'''Network loss'''
# Feed-forward pass
with torch.no_grad():
featlist_T = Teacher(inputs)
[D_feat1_T, D_feat2_T, D_feat3_T, D_feat4_T, D_feat5_T], outputs_T = decoder(featlist_T)
D_feat_T_list = [D_feat1_T.detach(), D_feat2_T.detach(), D_feat3_T.detach(), D_feat4_T.detach(), D_feat5_T.detach()]
featlist = Student(inputs)
#D_feat_S_list, outputs = decoder(featlist)
D_feat_S_list, drop_stud_list, drop_tchr_list, outputs = decoder(featlist, keep_prob=keep_prob, T_dec_feat_list=D_feat_T_list, dense_feat_T = featlist_T[-1].detach())
outputs = F.interpolate(outputs, scale_factor=2, mode='bilinear')
##################################### Valid mask definition ####################################
# masking valied area
valid_mask = make_mask(depths, args.dataset)
valid_out = outputs[valid_mask]
valid_gt_sparse = depths[valid_mask]
###################################### scale invariant loss ########################################
scale_inv_loss = scale_invariant_loss(valid_out, valid_gt_sparse)
####################################################################################################
###################################### feature decomposition loss ##################################
feat_loss, lap1_loss, lap2_loss, lap3_loss = decomposition_loss(D_feat_S_list, D_feat_T_list)
feat_loss_ = feat_loss[0] + feat_loss[1] + feat_loss[2] + feat_loss[3] + feat_loss[4]
lap1_loss_ = lap1_loss[0] + lap1_loss[1] + lap1_loss[2] + lap1_loss[3] + lap1_loss[4]
lap2_loss_ = lap2_loss[0] + lap2_loss[1] + lap2_loss[2] + lap2_loss[3] + lap2_loss[4]
lap3_loss_ = lap3_loss[0] + lap3_loss[1] + lap3_loss[2] + lap3_loss[3] + lap3_loss[4]
####################################################################################################
###################################### Dropped feature loss ########################################
drop_loss = feature_loss_L1(drop_stud_list, drop_tchr_list)
drop_loss_ = drop_loss[0] + drop_loss[1] + drop_loss[2] + drop_loss[3] + drop_loss[4]
if n_iter > total_iter_drop:
keep_prob = 0.8
else:
keep_prob = (-0.2/total_iter_drop)*(n_iter) + 1.0
decoder.module.set_drop_prob(1-keep_prob)
####################################################################################################
loss = scale_inv_loss + 0.2*drop_loss_ + 0.3*feat_loss_ + 0.1*lap1_loss_ + 0.1*lap2_loss_ + 0.1*lap3_loss_
# zero the parameter gradients and backward & optimize
optimizer.zero_grad()
loss.backward()
if n_iter == total_iter:
current_lr = end_lr
else:
current_lr = (base_lr - end_lr) * (1 - n_iter / total_iter) ** 0.5 + end_lr
n_iter += 1
optimizer.param_groups[0]['lr'] = current_lr
optimizer.step()
if ((i+1) % 100 == 0):
if (args.rank == 0):
print("epoch: %d, %d/%d"%(epoch+1,i+1,args.epoch_size))
print("[%6d/%6d] total: %.5f, feat: %.5f, drop: %.5f, lap1_loss: %.5f, lap2_loss: %.5f, lap3_loss: %.5f, scale_inv: %.5f"%(n_iter, total_iter,\
loss.item(), 0.3*feat_loss_.item(), 0.2*drop_loss_.item(), 0.1*lap1_loss_.item(), 0.1*lap2_loss_.item(), 0.1*lap3_loss_.item(), scale_inv_loss.item()))
print("drop_prob: %.5f"%(1 - keep_prob))
total_loss = loss.item()
rmse_loss = (torch.sqrt(torch.pow(valid_out-valid_gt_sparse,2))).mean()
rmse_loss = rmse_loss.item()
train_loss_cnt = train_loss_cnt + 1
all_plot(args.save_path,total_loss, rmse_loss, train_loss_list, train_rmse_list, train_loss_dir,train_loss_dir_rmse,loss_pdf, rmse_pdf, train_loss_cnt,True)
if (args.rank == 0):
print(" learning decay... current lr: %.6f"%(current_lr))
torch.save(Student.state_dict(), save_dir+'/epoch_%02d_encoder_loss_%.4f.pkl' %(model_num+1,loss))
model_num = model_num + 1
return loss