/
trainer.py
351 lines (323 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
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
import torch
import torch.nn as nn
import time
import numpy as np
import torch.optim as optim
import torch.nn.functional as F
from tqdm import tqdm
import os
import json
import random
from utils import Logger,ap_per_class
from utils import non_maximum_supression_soft as nms
from utils import cal_tp_per_item as cal_tp
tosave = ['mAP']
plot = [0.5,0.75]
thresholds = np.around(np.arange(0.5,0.76,0.05),2)
class Trainer:
def __init__(self,cfg,datasets,net,epoch,cmp_net=None):
self.cfg = cfg
if 'train' in datasets:
self.trainset = datasets['train']
if 'val' in datasets:
self.valset = datasets['val']
if 'trainval' in datasets:
self.trainval = datasets['trainval']
else:
self.trainval = False
if 'test' in datasets:
self.testset = datasets['test']
self.net = net
name = cfg.exp_name
self.name = name
self.checkpoints = os.path.join(cfg.checkpoint,name)
self.device = cfg.device
self.net_ = cmp_net
self.optimizer = optim.Adam(self.net.parameters(),lr=cfg.lr,weight_decay=cfg.weight_decay)
self.lr_sheudler = optim.lr_scheduler.ReduceLROnPlateau(self.optimizer,mode='min', factor=cfg.lr_factor, threshold=0.0001,patience=cfg.patience,min_lr=cfg.min_lr)
if not(os.path.exists(self.checkpoints)):
os.mkdir(self.checkpoints)
self.predictions = os.path.join(self.checkpoints,'pred')
if not(os.path.exists(self.predictions)):
os.mkdir(self.predictions)
start,total = epoch
self.start = start
self.total = total
log_dir = os.path.join(self.checkpoints,'logs')
if not(os.path.exists(log_dir)):
os.mkdir(log_dir)
self.logger = Logger(log_dir)
torch.cuda.empty_cache()
self.save_every_k_epoch = cfg.save_every_k_epoch #-1 for not save and validate
self.val_every_k_epoch = cfg.val_every_k_epoch
self.upadte_grad_every_k_batch = 1
self.best_mAP = 0
self.best_mAP_epoch = 0
self.movingLoss = 0
self.bestMovingLoss = 10000
self.bestMovingLossEpoch = 1e9
self.early_stop_epochs = 50
self.alpha = 0.95 #for update moving loss
self.lr_change= cfg.adjust_lr
self.base_epochs = cfg.base_epochs
self.nms_threshold = cfg.nms_threshold
self.conf_threshold = cfg.dc_threshold
self.save_pred = False
#load from epoch if required
if start:
if (start=='-1')or(start==-1):
self.load_last_epoch()
else:
self.load_epoch(start)
else:
self.start = 0
self.net = self.net.to(self.device)
if not (cmp_net is None):
self.net_ = self.net_.to(self.device)
def load_last_epoch(self):
files = os.listdir(self.checkpoints)
idx = 0
for name in files:
if name[-3:]=='.pt':
epoch = name[6:-3]
if epoch=='best' or epoch=='bestm':
continue
idx = max(idx,int(epoch))
if idx==0:
exit()
else:
self.load_epoch(str(idx))
def save_epoch(self,idx,epoch):
saveDict = {'net':self.net.state_dict(),
'optimizer': self.optimizer.state_dict(),
'lr_scheduler':self.lr_sheudler.state_dict(),
'epoch':epoch,
'mAP':self.best_mAP,
'mAP_epoch':self.best_mAP_epoch,
'movingLoss':self.movingLoss,
'bestmovingLoss':self.bestMovingLoss,
'bestmovingLossEpoch':self.bestMovingLossEpoch}
path = os.path.join(self.checkpoints,'epoch_'+idx+'.pt')
torch.save(saveDict,path)
def load_epoch(self,idx):
model_path = os.path.join(self.checkpoints,'epoch_'+idx+'.pt')
if os.path.exists(model_path):
print('load:'+model_path)
info = torch.load(model_path)
self.net.load_state_dict(info['net'])
if not(self.lr_change):
self.optimizer.load_state_dict(info['optimizer'])#might have bugs about device
for state in self.optimizer.state.values():
for k, v in state.items():
if isinstance(v, torch.Tensor):
state[k] = v.to(self.device)
self.lr_sheudler.load_state_dict(info['lr_scheduler'])
self.start = info['epoch']+1
self.best_mAP = info['mAP']
self.best_mAP_epoch = info['mAP_epoch']
self.movingLoss = info['movingLoss']
self.bestMovingLoss = info['bestmovingLoss']
self.bestMovingLossEpoch = info['bestmovingLossEpoch']
else:
print('no such model at:',model_path)
exit()
def _updateRunningLoss(self,loss,epoch):
if self.bestMovingLoss>loss:
self.bestMovingLoss = loss
self.bestMovingLossEpoch = epoch
self.save_epoch('bestm',epoch)
def logMemoryUsage(self, additionalString=""):
if torch.cuda.is_available():
print(additionalString + "Memory {:.0f}Mb max, {:.0f}Mb current".format(
torch.cuda.max_memory_allocated() / 1024 / 1024, torch.cuda.memory_allocated() / 1024 / 1024))
def set_lr(self,lr):
#adjust learning rate manually
for param_group in self.optimizer.param_groups:
param_group['lr']=lr
#tbi:might set different lr to different kind of parameters
def adjust_lr(self,lr_factor):
#adjust learning rate manually
for param_group in self.optimizer.param_groups:
param_group['lr']*=lr_factor
def warm_up(self,epoch):
if len(self.base_epochs)==0:
return False
if epoch <= self.base_epochs[-1]:
if epoch in self.base_epochs:
self.adjust_lr(0.1)
return True
else:
return False
def train_one_epoch(self):
self.optimizer.zero_grad()
running_loss ={'xy':0.0,'wh':0.0,'conf':0.0,'cls':0.0,'obj':0.0,'all':0.0,'iou':0.0,'giou':0.0}
self.net.train()
n = len(self.trainset)
for data in tqdm(self.trainset):
inputs,labels = data
labels = labels.to(self.device).float()
display,loss = self.net(inputs.to(self.device).float(),gts=labels)
#display,loss = self.net_(inputs.to(self.device).float(),gts=labels)
#exit()
del inputs,labels
for k in running_loss:
if k in display.keys():
if np.isnan(display[k]):
continue
running_loss[k] += display[k]/n
loss.backward()
#solve gradient explosion problem caused by large learning rate or small batch size
#nn.utils.clip_grad_value_(self.net.parameters(), clip_value=2.0)
#nn.utils.clip_grad_norm_(self.net.parameters(),max_norm=2.0)
self.optimizer.step()
self.optimizer.zero_grad()
del loss
self.logMemoryUsage()
print(f'#Gt not matched:{self.net.loss.not_match}')
self.net.loss.reset_notmatch()
return running_loss
def train(self):
print("strat train:",self.name)
print("start from epoch:",self.start)
print("=============================")
self.optimizer.zero_grad()
print(self.optimizer.param_groups[0]['lr'])
epoch = self.start
stop_epochs = 0
#torch.autograd.set_detect_anomaly(True)
while epoch < self.total and stop_epochs<self.early_stop_epochs:
running_loss = self.train_one_epoch()
lr = self.optimizer.param_groups[0]['lr']
self.logger.write_loss(epoch,running_loss,lr)
#step lr
self._updateRunningLoss(running_loss['all'],epoch)
if not self.warm_up(epoch):
self.lr_sheudler.step(running_loss['all'])
lr_ = self.optimizer.param_groups[0]['lr']
if lr_ == self.cfg.min_lr:
stop_epochs +=1
if (epoch+1)%self.save_every_k_epoch==0:
self.save_epoch(str(epoch),epoch)
if (epoch+1)%self.val_every_k_epoch==0:
metrics = self.validate(epoch,'val',self.save_pred)
self.logger.write_metrics(epoch,metrics,tosave)
mAP = metrics['mAP']
if mAP >= self.best_mAP:
self.best_mAP = mAP
self.best_mAP_epoch = epoch
print("best so far, saving......")
self.save_epoch('best',epoch)
if self.trainval:
metrics = self.validate(epoch,'train',self.save_pred)
self.logger.write_metrics(epoch,metrics,tosave,mode='Trainval')
mAP = metrics['mAP']
print(f"best so far with {self.best_mAP} at epoch:{self.best_mAP_epoch}")
epoch +=1
print("Best mAP: {:.4f} at epoch {}".format(self.best_mAP, self.best_mAP_epoch))
self.save_epoch(str(epoch-1),epoch-1)
def validate(self,epoch,mode,save=False):
self.net.eval()
res = {}
print('start Validation Epoch:',epoch)
if mode=='val':
valset = self.valset
else:
valset = self.trainval
with torch.no_grad():
mAP = 0
count = 0
batch_metrics={}
for th in thresholds:
batch_metrics[th] = []
gt_labels = []
pd_num = 0
for data in tqdm(valset):
inputs,labels,info = data
pds = self.net(inputs.to(self.device).float())
nB = pds.shape[0]
gt_labels += labels[:,1].tolist()
for b in range(nB):
pred = pds[b].view(-1,self.cfg.cls_num+5)
name = info['img_id'][b]
size = info['size'][b]
pad = info['pad'][b]
pred[:,:4] *= max(size)
pred[:,0] -= pad[1]
pred[:,1] -= pad[0]
if save:
pds_ = list(pred.cpu().numpy().astype(float))
pds_ = [list(pd) for pd in pds_]
result ={'bboxes':pds_,'pad':pad,'size':size}
res[name] = result
pred_nms = nms(pred,self.conf_threshold, self.nms_threshold)
gt = labels[labels[:,0]==b,1:].reshape(-1,5)
#pred_nms_ = np.round(pred_nms.cpu().numpy().astype(np.float32),1)
#print(pred_nms_)
#print(gt)
pd_num+=pred_nms.shape[0]
'''if save:
print(pred_nms)
print(gt)'''
count+=1
for th in batch_metrics:
batch_metrics[th].append(cal_tp(pred_nms,gt,th))
metrics = {}
for th in batch_metrics:
tps,scores,pd_labels = [np.concatenate(x, 0) for x in list(zip(*batch_metrics[th]))]
precision, recall, AP,_,_ = ap_per_class(tps, scores, pd_labels, gt_labels)
mAP += np.mean(AP)
if th in plot:
metrics['AP/'+str(th)] = np.mean(AP)
metrics['Precision/'+str(th)] = np.mean(precision)
metrics['Recall/'+str(th)] = np.mean(recall)
metrics['mAP'] = mAP/len(thresholds)
if save:
json.dump(res,open(os.path.join(self.predictions,'pred_epoch_'+str(epoch)+'.json'),'w'))
return metrics
def test(self):
self.net.eval()
res = {}
with torch.no_grad():
for data in tqdm(self.testset):
inputs,info = data
pds = self.net(inputs.to(self.device).float())
nB = pds.shape[0]
for b in range(nB):
pred = pds[b].view(-1,self.cfg.cls_num+5)
name = info['img_id'][b]
tsize = info['size'][b]
pad = info['pad'][b]
pred[:,:4]*= max(tsize)
pred[:,0] -= pad[1]
pred[:,1] -= pad[0]
cls_confs,cls_labels = torch.max(pred[:,5:],dim=1,keepdim=True)
pred_nms = torch.cat((pred[:,:5],cls_confs,cls_labels.float()),dim=1)
#pred_nms = nms(pred,self.conf_threshold, self.nms_threshold)
pds_ = list(pred_nms.cpu().numpy().astype(float))
pds_ = [list(pd) for pd in pds_]
res[name] = pds_
json.dump(res,open(os.path.join(self.predictions,'pred_test.json'),'w'))
def validate_random(self):
self.net.eval()
self.valset.shuffle = True
bs = self.valset.batch_size
imgs = list(range(bs))
preds = list(range(bs))
gts = list(range(bs))
sizes = list(range(bs))
with torch.no_grad():
inputs,labels,info = next(iter(self.valset))
pds = self.net(inputs.to(self.device).float())
for b in range(bs):
pred = pds[b].view(-1,self.cfg.cls_num+5)
pred_nms = nms(pred,self.conf_threshold, self.nms_threshold)
size = info['size'][b]
pad = info['pad'][b]
pred_nms[:,:4] *= max(size)
pred_nms[:,0] -= pad[1]
pred_nms[:,1] -= pad[0]
imgs[b] = inputs[b]
preds[b] = pred_nms
gts[b] = labels[labels[:,0]==b,1:].reshape(-1,5)
sizes[b] = size
return imgs,preds,gts,sizes