-
Notifications
You must be signed in to change notification settings - Fork 2
/
trainer.py
282 lines (210 loc) · 12.8 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
import numpy as np
import torch
from torch import optim
from torch.autograd import Variable
from models.CC import CrowdCounter
from config import cfg
from misc.utils import *
import datasets
from tqdm import tqdm
from torchvision.utils import make_grid
from misc.utils import adjust_learning_rate, adjust_double_learning_rate
class Trainer():
def __init__(self, cfg_data, pwd):
self.cfg_data = cfg_data
self.train_loader, self.val_loader = datasets.loading_data(cfg.DATASET)
self.data_mode = cfg.DATASET
self.exp_name = cfg.EXP_NAME
self.exp_path = cfg.EXP_PATH
self.pwd = pwd
self.net_name = cfg.NET
self.gpu_id = cfg.GPU_ID
self.writer, self.log_txt = logger(self.exp_path, self.exp_name, self.pwd, 'exp', resume=cfg.RESUME)
self.net = CrowdCounter(self.net_name, self.gpu_id)
self.categorys = self.cfg_data.CATEGORYS
self.num_classes = len(self.categorys)
if cfg.BACKBONE_FREEZE:
self.optimizer = optim.AdamW([
{'params': [param for name, param in self.net.named_parameters() if 'backbone' in name], 'lr': cfg.CONV_LR, 'weight_decay': cfg.WEIGHT_DECAY},
{'params': [param for name, param in self.net.named_parameters() if 'backbone' not in name], 'lr': cfg.BASE_LR, 'weight_decay': cfg.WEIGHT_DECAY},
])
else:
self.optimizer = optim.AdamW(self.net.parameters(), lr=cfg.BASE_LR, weight_decay=cfg.WEIGHT_DECAY)
self.train_record = {'best_cls_avg_mae': 1e20, 'best_cls_avg_mse':1e20, 'best_cls_weight_mse':1e20, 'best_model_name': ''}
self.timer = {'iter time' : Timer(),'train time' : Timer(),'val time' : Timer()}
self.epoch = 0
self.i_tb = 0
self.num_iters = cfg.MAX_EPOCH * np.int(len(self.train_loader))
self.train_loss_avg = 0
self.val_loss_avg = 0
if cfg.PRE_GCC:
self.net.load_state_dict(torch.load(cfg.PRE_GCC_MODEL))
if cfg.RESUME:
latest_state = torch.load(cfg.RESUME_PATH)
self.net.load_state_dict(latest_state['net'])
self.optimizer.load_state_dict(latest_state['optimizer'])
self.epoch = latest_state['epoch'] + 1
self.i_tb = latest_state['i_tb']
self.num_iters = latest_state['num_iters']
self.train_record = latest_state['train_record']
self.exp_path = latest_state['exp_path']
self.exp_name = latest_state['exp_name']
def forward(self):
# self.validate()
for epoch in range(self.epoch,cfg.MAX_EPOCH):
self.epoch = epoch
record_path = os.path.join(self.exp_path, self.exp_name, 'log.txt')
record = open(record_path, 'a+')
# training
self.timer['train time'].tic()
self.train()
self.timer['train time'].toc(average=False)
print( 'train time: {:.2f}s'.format(self.timer['train time'].diff) )
print( 'train time: {:.2f}s'.format(self.timer['train time'].diff), file=record )
if self.epoch % cfg.CP_FREQ == 0:
save_checkpoint(self)
print( '='*20 )
print( '='*20, file=record )
self.writer.add_scalar('train_loss_avg', self.train_loss_avg, self.epoch)
self.train_loss_avg = 0
# validation
if is_validation(cfg.VAL_STAGE,cfg.VAL_FREQ,self.epoch):
self.timer['val time'].tic()
self.validate() # only when pos_emb = False
self.timer['val time'].toc(average=False)
print( 'val time: {:.2f}s'.format(self.timer['val time'].diff) )
print( 'val time: {:.2f}s'.format(self.timer['val time'].diff), file=record )
self.writer.add_scalar('val_loss_avg', self.val_loss_avg, self.epoch)
self.val_loss_avg = 0
record.close()
def train(self): # training for all datasets
self.net.train()
train_losses = AverageMeter()
for it, data in enumerate(self.train_loader, 0):
self.i_tb += 1
self.timer['iter time'].tic()
for k, v in data.items(): # data to cuda
data[k] = v.cuda()
num_class = self.num_classes
rgb, nir, gt_map = data['rgb'], data['nir'], data['gt_map'][:, :num_class, :, :]
self.optimizer.zero_grad()
# input_grid = make_grid([rgb[0],nir[0]], nrow=2, normalize=True, scale_each=True)
# self.writer.add_image("inputs", input_grid, it)
outputs = self.net(data, num_class, it)
pred_map = outputs['pred_map']
gauss_map = outputs['gauss_map']
output_list = list()
# for c_idx in range(self.num_classes):
# output_list.extend([gt_map[0][c_idx].unsqueeze(dim=0), \
# gauss_map[0][c_idx].unsqueeze(dim=0), \
# pred_map[0][c_idx].unsqueeze(dim=0)])
# output_grid = make_grid(output_list, nrow=4, normalize=True, scale_each=True)#排列图像
# self.writer.add_image("outputs", output_grid, it)
loss = self.net.loss
loss.backward()
self.optimizer.step()
train_losses.update(loss.item())
if cfg.BACKBONE_FREEZE:
base_lr, conv_lr = adjust_double_learning_rate(self.optimizer, self.i_tb, self.num_iters, base_lr=cfg.BASE_LR, conv_lr = cfg.CONV_LR)
else:
base_lr = adjust_learning_rate(self.optimizer, self.i_tb, self.num_iters, lr=cfg.BASE_LR)
record_path = os.path.join(self.exp_path, self.exp_name, 'log.txt')
record = open(record_path, 'a+')
if (it + 1) % cfg.PRINT_FREQ == 0:
self.writer.add_scalar('train_base_lr', base_lr, self.i_tb)
self.writer.add_scalar('train_loss', loss.item(), self.i_tb)
self.timer['iter time'].toc(average=False)
if cfg.BACKBONE_FREEZE:
print( '[ep %d][it %d][loss %.4f][conv_lr %.8f][base_lr %.8f][%.2fs]' % \
(self.epoch + 1, it + 1, loss.item(), self.optimizer.param_groups[0]['lr'], self.optimizer.param_groups[1]['lr'], self.timer['iter time'].diff) )
print( '[ep %d][it %d][loss %.4f][conv_lr %.8f][base_lr %.8f][%.2fs]' % \
(self.epoch + 1, it + 1, loss.item(), self.optimizer.param_groups[0]['lr'], self.optimizer.param_groups[1]['lr'], self.timer['iter time'].diff), file=record )
else:
print( '[ep %d][it %d][loss %.4f][base_lr %.8f][%.2fs]' % \
(self.epoch + 1, it + 1, loss.item(), self.optimizer.param_groups[0]['lr'], self.timer['iter time'].diff) )
print( '[ep %d][it %d][loss %.4f][base_lr %.8f][%.2fs]' % \
(self.epoch + 1, it + 1, loss.item(), self.optimizer.param_groups[0]['lr'], self.timer['iter time'].diff), file=record )
for c_idx in range(self.num_classes):
print( ' ', self.categorys[c_idx],': [cnt: gt: %.1f pred: %.2f]' % (gt_map[0][c_idx].sum().data/self.cfg_data.LOG_PARA, pred_map[0][c_idx].sum().data/self.cfg_data.LOG_PARA) )
print( ' ', self.categorys[c_idx],': [cnt: gt: %.1f pred: %.2f]' % (gt_map[0][c_idx].sum().data/self.cfg_data.LOG_PARA, pred_map[0][c_idx].sum().data/self.cfg_data.LOG_PARA), file=record )
record.close()
self.train_loss_avg = train_losses.avg
def validate(self):
self.net.eval()
val_losses = AverageMeter()
maes = AverageCategoryMeter(self.num_classes)
mses = AverageCategoryMeter(self.num_classes)
cmses = AverageMeter()
for index, data in enumerate(self.val_loader, 0):
for k, v in data.items(): # data to cuda
data[k] = v.cuda()
num_class = self.num_classes
rgb, nir, gt_map = data['rgb'], data['nir'], data['gt_map'][:, :num_class, :, :]
with torch.set_grad_enabled(False):
if cfg.POS_EMBEDDING:
b, c, h, w = rgb.shape
rh, rw = self.cfg_data.TRAIN_SIZE
crop_RGBs, crop_Ts, crop_masks = [],[],[] # mask是用来记录overlapping的部分重叠过几次的
for i in range(0, h, rh): # i=0,256
# (0,256),(224,480),crop blcok overlapping
gis, gie = max(min(h-rh, i), 0), min(h, i+rh)
for j in range(0, w, rw):
gjs, gje = max(min(w-rw, j), 0), min(w, j+rw)
crop_RGBs.append(rgb[:, :, gis:gie, gjs:gje])
crop_Ts.append(nir[:, :, gis:gie, gjs:gje])
mask = torch.zeros(b, 1, h//self.cfg_data.LABEL_FACTOR, w//self.cfg_data.LABEL_FACTOR).cuda()
mask[:, :, gis//self.cfg_data.LABEL_FACTOR:gie//self.cfg_data.LABEL_FACTOR, gjs//self.cfg_data.LABEL_FACTOR:gje//self.cfg_data.LABEL_FACTOR].fill_(1.0)
crop_masks.append(mask)
crop_RGBs, crop_Ts, crop_masks = map(lambda x: torch.cat(x, dim=0), (crop_RGBs, crop_Ts, crop_masks))
crop_data = {'rgb': crop_RGBs, 'nir': crop_Ts,}
crop_outputs = self.net(crop_data, mode = 'val')
crop_preds = crop_outputs['pred_map']
h, w, rh, rw = h//self.cfg_data.LABEL_FACTOR, w//self.cfg_data.LABEL_FACTOR, rh//self.cfg_data.LABEL_FACTOR, rh//self.cfg_data.LABEL_FACTOR
idx = 0
pred_map = torch.zeros(b, 1, h, w).cuda()
for i in range(0, h, rh):
gis, gie = max(min(h-rh, i), 0), min(h, i+rh)
for j in range(0, w, rw):
gjs, gje = max(min(w-rw, j), 0), min(w, j+rw)
pred_map[:, :, gis:gie, gjs:gje] += crop_preds[idx]
idx += 1
mask = crop_masks.sum(dim=0)
pred_map = pred_map / mask
else:
outputs = self.net(data, num_class, mode = 'val')
pred_map = outputs['pred_map']
val_losses.update(self.net.loss.item())
# abs_errors, square_errors = eval(pred_map, gt_map, self.cfg_data.LOG_PARA)
abs_errors, square_errors, weights = eval_mc(pred_map, gt_map, self.cfg_data.LOG_PARA)
wmse = 0.0
for c_idx in range(self.num_classes):
maes.update(abs_errors[c_idx], c_idx)
mses.update(square_errors[c_idx], c_idx)
wmse += square_errors[c_idx] * weights[c_idx]
cmses.update(wmse)
N = len(self.val_loader)
self.val_loss_avg = val_losses.avg
overall_mae = maes.avg # list
# RMSE ??
overall_rmse = np.sqrt(mses.avg)
cls_weight_mse = cmses.avg
cls_avg_mae = sum(overall_mae) / self.num_classes
cls_avg_rmse = sum(overall_rmse) / self.num_classes
self.writer.add_scalar('val_loss', self.val_loss_avg, self.epoch + 1)
self.writer.add_scalar('cls_weight_mse', cls_weight_mse, self.epoch + 1)
self.writer.add_scalar('cls_avg_mae', cls_avg_mae, self.epoch + 1)
self.writer.add_scalar('cls_avg_rmse', cls_avg_rmse, self.epoch + 1)
self.train_record = update_model(self.net,self.optimizer,self.epoch,self.i_tb,self.exp_path,self.exp_name, \
[self.val_loss_avg, overall_mae, overall_rmse, cls_avg_mae, cls_avg_rmse, cls_weight_mse], self.train_record, self.log_txt, self.categorys)
record_path = os.path.join(self.exp_path, self.exp_name, 'log.txt')
record = open(record_path, 'a+')
log_str = 'Val{}, val_loss {val_loss:.4f} cls_avg_mae {cls_avg_mae:.4f} cls_avg_rmse {cls_avg_rmse:.4f} cls_weight_mse {cls_weight_mse:.4f}'.\
format(N, val_loss=self.val_loss_avg, cls_avg_mae=cls_avg_mae, cls_avg_rmse=cls_avg_rmse, cls_weight_mse=cls_weight_mse)
print(log_str)
print(log_str, file=record)
for c_idx in range(self.num_classes):
class_str = '{category}, mae: {mae:.4f} rmse: {rmse:.4f}'\
.format(category = self.categorys[c_idx], mae = overall_mae[c_idx], rmse = overall_rmse[c_idx])
print(class_str)
print(class_str, file=record)
record.close()