-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathtraining.py
282 lines (255 loc) · 13.2 KB
/
training.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
'''Implements a generic training loop.
'''
import torch
import utils
from torch.utils.tensorboard import SummaryWriter
from tqdm.autonotebook import tqdm
import time
import numpy as np
import os
import shutil
from torch import nn
from torch.optim import AdamW
import time
import random
import pickle
import qtlib
import pdb
from torch_ema import ExponentialMovingAverage
torch.cuda.set_device(int(os.environ["CUDA_USING"]))
def trainMCvalidationLight(model, train_dataloader, epochs, lr, steps_til_summary, epochs_til_checkpoint, model_dir,
loss_fn, pruning_fn, summary_fn, double_precision=False, clip_grad=False,optimizer = 'Adam',
loss_schedules=None, resume_checkpoint={}, objs_to_save={}, epochs_til_pruning=4,subjpath=None,opt=None,coord_dataset = None):
# reset the batchnorm
def reset_batchnorm(m):
if type(m) == nn.BatchNorm3d: m.reset_parameters()
if type(m) == nn.BatchNorm1d: m.reset_parameters()
# model.apply(reset_batchnorm)'
if optimizer == 'Adam':
optim = torch.optim.Adam(lr=lr, params=model.parameters())
elif optimizer == 'AdamW':
print('using AdamW!!!')
optim = AdamW(lr = lr,params = model.parameters(),weight_decay=0.05,betas=(0.9,0.999))
start=time.time()
# load optimizer if supplied
if 'optimizer_state_dict' in resume_checkpoint:
optim.load_state_dict(resume_checkpoint['optimizer_state_dict'])
for g in optim.param_groups:
g['lr'] = lr
if opt.representation == 1 and opt.trainingstrate==1:
optim = torch.optim.Adam(lr = lr, params = model.parameters(), betas=(0.9, 0.99), eps=1e-15)
lr_scheduler = torch.optim.lr_scheduler.LambdaLR(optim, lambda iter: 0.1 ** min(iter / (len(train_dataloader) * epochs), 1))
ema = ExponentialMovingAverage(model.parameters(), decay=0.95)
os.makedirs(model_dir, exist_ok=True)
summaries_dir = os.path.join(model_dir, 'summaries')
utils.cond_mkdir(summaries_dir)
checkpoints_dir = os.path.join(model_dir, 'checkpoints')
utils.cond_mkdir(checkpoints_dir)
writer = SummaryWriter(summaries_dir)
total_steps = 0
if 'total_steps' in resume_checkpoint:
total_steps = resume_checkpoint['total_steps']
start_epoch = 0
if 'epoch' in resume_checkpoint:
start_epoch = resume_checkpoint['epoch']
minvalloss = 100000
minvalloss2 = 100000
epochloss = 100000
minepochloss = 100000
earlystopdone = 0
earlycount = 0
winsize = 10
minwinval = 10000000
minwinvalindex = 0
valid_loss = minvalloss
start=time.time()
if len(train_dataloader) * epochs>1000: # add for FS
tmpepochs_til_checkpoint = epochs_til_checkpoint # add for FS
epochs_til_checkpoint = 10 # add for FS
torch.save(model.state_dict(), os.path.join(checkpoints_dir, '1best_val_model_000000.pth'))
bestmodel = None
with tqdm(total=len(train_dataloader) * epochs) as pbar:
pbar.update(total_steps)
train_losses = []
valid_losses = []
es_valid_losses = []
es_valid_losses_markinepoch = []
es_valid_losses_markinepoch_win = []
for epoch in range(start_epoch, epochs):
epochstarttime = time.time()
if opt.MCchannel:
usechannel = coord_dataset.bvalnotzero
print(usechannel)
if valid_loss<minvalloss2:
minvalloss2 = valid_loss
torch.save(model.state_dict(),
os.path.join(checkpoints_dir, 'ck0_val_model.pth'))
if epoch>=100: # add for FS
epochs_til_checkpoint = tmpepochs_til_checkpoint # add for FS
if not epoch % epochs_til_checkpoint and epoch:
updateornot = int(valid_loss<minvalloss)
dpbest2now = os.path.join(checkpoints_dir, 'ck0_val_model.pth')
dpnow = os.path.join(checkpoints_dir, f'{updateornot:1d}best_val_model_{total_steps:06d}.pth')
os.system(f"cp {dpbest2now} {dpnow}")
np.savetxt(os.path.join(checkpoints_dir, 'train_losses_%06d.txt' % total_steps),
np.array(train_losses))
np.savetxt(os.path.join(checkpoints_dir, 'es_valid_losses_%06d.txt' % total_steps),
np.array(es_valid_losses))
if not (epoch-winsize//2) % epochs_til_checkpoint and epoch>winsize:
es_valid_losses_markinepoch_win.append( np.average(es_valid_losses_markinepoch[epoch-winsize:epoch]) )
if es_valid_losses_markinepoch_win[-1]<minwinval:
minwinval = es_valid_losses_markinepoch_win[-1]
minwinvalindex = epoch-winsize//2
if epoch >= 10*winsize and not earlystopdone:
if (es_valid_losses_markinepoch_win[-2]-es_valid_losses_markinepoch_win[-1])/es_valid_losses_markinepoch_win[-1]<0.0001:
earlycount = earlycount + 1
if earlycount > 2:
dpearlystop = os.path.join(checkpoints_dir,f"ES_{(epoch-winsize//2)}")
if not os.path.exists(dpearlystop):
os.mkdir(dpearlystop)
dpearlystop = os.path.join(checkpoints_dir,f"ES_min_{minwinvalindex}")
if not os.path.exists(dpearlystop):
os.mkdir(dpearlystop)
earlystopdone = 1
else:
earlycount = 0
if epoch in [10,50,100,300]:
dpbest2now = os.path.join(checkpoints_dir, 'ck0_val_model.pth')
dpnow = os.path.join(checkpoints_dir, f'epoch{epoch:06d}.pth')
os.system(f"cp {dpbest2now} {dpnow}")
if not (epoch + 1) % epochs_til_pruning:
retile = False
else:
retile = True
count = 0
length = len(train_dataloader)
if opt.cv:
splitnum = int(length*0.8)
else:
splitnum = int(length)
epochlosslist = []
validlosslist = []
esvalidloss_epochs = []
stependtime = time.time()
for step, (model_input, gt) in enumerate(train_dataloader):
start_time = time.time()
tmp = {}
for key, value in model_input.items():
if isinstance(value, torch.Tensor):
tmp.update({key: value.cuda(int(os.environ["CUDA_USING"]))})
# print(key,time.time()-start_time)
else:
tmp.update({key: value})
model_input = tmp
tmp = {}
for key, value in gt.items():
if isinstance(value, torch.Tensor):
tmp.update({key: value.cuda(int(os.environ["CUDA_USING"]))})
# print('gt',key,time.time()-start_time)
else:
tmp.update({key: value})
gt = tmp
# '''
if double_precision:
model_input = {key: value.double() for key, value in model_input.items()}
gt = {key: value.double() for key, value in gt.items()}
if step<=splitnum:
if opt.sz_block_mode in ['min32']:
if len(gt['NoneEdgeIndex'][0])==0:
continue
model_output = model(model_input)
losses = loss_fn(model_output, gt, total_steps, retile=retile)
train_loss = 0.
es_valid_loss = 0.
for loss_name, loss in losses.items():
single_loss = loss.mean()
if loss_schedules is not None and loss_name in loss_schedules:
writer.add_scalar(loss_name + "_weight", loss_schedules[loss_name](total_steps), total_steps)
single_loss *= loss_schedules[loss_name](total_steps)
writer.add_scalar(loss_name, single_loss, total_steps)
if loss_name == 'img_loss':
train_loss += single_loss
elif loss_name == 'valid_loss':
es_valid_loss += single_loss
train_losses.append(train_loss.item())
if opt.EarlyStopping or opt.MCchannel:
es_valid_losses.append(es_valid_loss.item())
esvalidloss_epochs.append(es_valid_loss.item())
else:
es_valid_losses.append(es_valid_loss)
esvalidloss_epochs.append(es_valid_loss)
writer.add_scalar("total_train_loss", train_loss, total_steps)
writer.add_scalar("total_es_valid_loss", es_valid_loss, total_steps)
optim.zero_grad()
train_loss.backward()
if clip_grad:
if isinstance(clip_grad, bool):
torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.)
else:
torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=clip_grad)
optim.step()
pbar.update(1)
epochlosslist.append(train_loss.detach().item())
total_steps += 1
if opt.representation == 1 and opt.trainingstrate==1:
lr_scheduler.step()
else:
with torch.no_grad():
model_output = model(model_input)
losses = loss_fn(model_output, gt, total_steps, retile=retile)
valid_loss = 0.
es_valid_loss = 0.
for loss_name, loss in losses.items():
single_loss = loss.mean()
if loss_schedules is not None and loss_name in loss_schedules:
writer.add_scalar(loss_name + "_weight", loss_schedules[loss_name](total_steps), total_steps)
single_loss *= loss_schedules[loss_name](total_steps)
writer.add_scalar(loss_name, single_loss, total_steps)
valid_loss += single_loss
if loss_name == 'img_loss':
train_loss += single_loss
elif loss_name == 'valid_loss':
es_valid_loss += single_loss
valid_losses.append(valid_loss.item())
validlosslist.append(valid_loss.item())
epochlosslist.append(valid_loss.item())
if opt.EarlyStopping or opt.MCchannel:
es_valid_losses.append(es_valid_loss.item())
esvalidloss_epochs.append(es_valid_loss.item())
else:
es_valid_losses.append(es_valid_loss)
esvalidloss_epochs.append(es_valid_loss)
writer.add_scalar("total_valid_loss", valid_loss, total_steps)
writer.add_scalar("total_es_valid_loss", es_valid_loss, total_steps)
if opt.representation == 1 and opt.trainingstrate==1:
ema.update()
train_loss = np.average(epochlosslist)
if opt.cv:
valid_loss = np.average(validlosslist)
else:
valid_loss = np.average(epochlosslist)
esvalidloss_epoch = np.average(esvalidloss_epochs)
es_valid_losses_markinepoch.append(esvalidloss_epoch)
# if opt.earlystop:
tqdm.write("Epoch %d, Total loss %0.6f, Valid loss %0.6f, es Valid loss %0.6f, iteration time %0.6f, epoch time %0.6f" % (epoch, train_loss, valid_loss, esvalidloss_epoch, time.time() - start_time, time.time()-epochstarttime))
if 0:
state = {
'epoch': epochs,
'global_step': total_steps,
}
state['optimizer'] = optim.state_dict()
state['lr_scheduler'] = lr_scheduler.state_dict()
state['ema'] = ema.state_dict()
state['model'] = model.state_dict()
torch.save(state, os.path.join(checkpoints_dir, 'model_final_%06d.pth' % total_steps))
else:
torch.save(model.state_dict(),
os.path.join(checkpoints_dir, 'model_final_%06d.pth' % total_steps))
np.savetxt(os.path.join(checkpoints_dir, 'train_losses_final_%06d.txt' % total_steps),
np.array(train_losses))
np.savetxt(os.path.join(checkpoints_dir, 'valid_losses_final_%06d.txt' % total_steps),
np.array(valid_losses))
np.savetxt(os.path.join(checkpoints_dir, 'es_valid_losses_final_%06d.txt' % total_steps),
np.array(es_valid_losses))
np.savetxt(os.path.join(checkpoints_dir, 'es_valid_losses_inepoch_final_%06d.txt' % total_steps),
np.array(es_valid_losses_markinepoch))