/
train.py
449 lines (353 loc) · 16.3 KB
/
train.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
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
#coding=utf-8
import torch
from torch.autograd import Variable
from torch.backends import cudnn
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
import numpy as np
import pprint
import os
import argparse
import pandas as pd
import warnings
warnings.filterwarnings("ignore")
# from data_load_gray_multiscale import MouseKeypointsDataset
from data_load_RGB_multiscale import MouseKeypointsDataset
from utilis.checkpointsave import savecheckpoint, saveepochcheckpont, saveepochcheckpont_PCK
from utilis.netparameter160 import config
from utilis.metrics import PCK
# from model.mainStackedHourGlass_full_finalcas2_C1GM2 import *
from model.mainStackedHourGlass_full_finalcas2_C3GM2 import *
from tensorboardX import SummaryWriter
cudnn.benchmark = True
device = torch.device("cuda:0,1" if torch.cuda.is_available() else "cpu")
#DeepLabCut dataset
# netname = 'mousefullnet_finalcas2_C1GM2_4fulltrain'
# validresultname = 'results/mouse/compare/results/validresult_fullnet_finalcas2_C1GM2_4fulltrain.csv'
# trainresultname = 'results/mouse/compare/results/trainresult_fullnet_finalcas2_C1GM2_4fulltrain.csv'
#PDMB
netname = 'pdmousefullnet_finalcas2_C3GM2'
validresultname = 'results/pdmouse/compare/results/validresult_fullnet_finalcas2_C3GM2.csv'
trainresultname = 'results/pdmouse/compare/results/trainresult_fullnet_finalcas2_C3GM2.csv'
#zebra dataset
# netname = 'zebrafullnet_finalcas2_C3GM2'
# validresultname = 'results/zebra/compare/results/validresult_fullnet_finalcas2_C3GM2.csv'
# trainresultname = 'results/zebra/compare/results/trainresult_fullnet_finalcas2_C3GM2.csv'
parser = argparse.ArgumentParser(description='PyTorch Mousepose Training')
parser.add_argument('--resume', '-r', action='store_true', help='resume from checkpoint')
args = parser.parse_args()
def get_peak_points(heatmaps):
"""
heatmap to keypoints
:param heatmaps: numpy array (N,4,240,320)
:return:numpy array (N,4,2)
"""
N,C,H,W = heatmaps.shape
all_peak_points = []
for i in range(N):
peak_points = []
for j in range(C):
yy,xx = np.where(heatmaps[i,j] == heatmaps[i,j].max())
y = yy[0]
x = xx[0]
peak_points.append([x,y])
all_peak_points.append(peak_points)
all_peak_points = np.array(all_peak_points)
return all_peak_points
def get_mse(pred_points,gts,indices_valid=None):
"""
:param pred_points: numpy (N,4,2)
:param gts: numpy (N,4,2)
:return:
"""
pred_points = pred_points[indices_valid[0],indices_valid[1],:]
gts = gts[indices_valid[0],indices_valid[1],:]
pred_points = torch.from_numpy(pred_points).float()
gts = torch.from_numpy(gts).float()
criterion = nn.MSELoss()
corloss = criterion(pred_points,gts)
rmse = np.sqrt(corloss)
#print("loss-----: %f" % corloss)
return rmse
def calculate_mask(heatmaps_targets):
"""
:param heatmaps_target: Variable (N,15,96,96)
:return: Variable (N,15,96,96)
"""
N,C,_,_ = heatmaps_targets.size()
N_idx = []
C_idx = []
NaN_N_idx = []
NaN_C_idx = []
for n in range(N):
for c in range(C):
max_v = heatmaps_targets[n,c,:,:].max()
if max_v != 0.0:
N_idx.append(n)
C_idx.append(c)
else:
NaN_N_idx.append(n)
NaN_C_idx.append(c)
mask = torch.zeros(heatmaps_targets.size())
mask[N_idx,C_idx,:,:] = 1.
# mask = mask.float().to(device)
return mask,[N_idx,C_idx],[NaN_N_idx,NaN_C_idx]
def makenotvalidpointsasnan(predicted, gts,indices_notvalid):
for i,j in zip(indices_notvalid[0],indices_notvalid[1]):
predicted[i,j,:] = [-1,-1]
gts[i, j, :] = [-1, -1]
return predicted, gts
def resume_checkpoint(net):
'''
:return:
'''
if args.resume:
# Load checkpoint.
print('==> Resuming from checkpoint..')
assert os.path.isdir('checkpoint'), 'Error: no checkpoint directory found!'
checkpoint = torch.load(os.path.join('./checkpoint', netname, 'epoch_6checkpoint.ckpt'))
net.load_state_dict(checkpoint['net'])
start_epoch = checkpoint['epoch']
config['start_epoch'] = start_epoch+1
def initnet():
pprint.pprint(config)
torch.manual_seed(0)
net = Gmscenet( nChannels=128, nStack=config['nstack'], nModules=1, numReductions=3, nJoints=4)
# using 2 Gpus
print("Let's use", torch.cuda.device_count(), "GPUs")
if torch.cuda.device_count() > 1:
print("Let's use", torch.cuda.device_count(), "GPUs")
net = torch.nn.DataParallel(net)
net.to(device)
print(device)
trainDataset = MouseKeypointsDataset(csv_file=config['trainsample'], root_dir='')
sample_num = len(trainDataset)
config['train_num'] = sample_num
validDataset = MouseKeypointsDataset(csv_file=config['validsample'], root_dir='')
sample_num = len(validDataset)
config['valid_num'] = sample_num
print('train:', config['train_num'], 'validation:', config['valid_num'])
# split train and valid
# train_size = int(0.8 * sample_num)
# valid_size = sample_num - train_size
# config['train_num'], config['valid_num'] = train_size, valid_size
# train_db, val_db = torch.utils.data.random_split(trainDataset, [train_size, valid_size])
# print('train:', len(train_db), 'validation:', len(val_db))
trainDataLoader = DataLoader(trainDataset, config['batch_size'], shuffle=True, num_workers=2)
validDataLoader = DataLoader(validDataset, config['batch_size_valid'], shuffle=True, num_workers=2)
# if (config['checkout'] != ''):
# net.load_state_dict(torch.load(config['checkout'],map_location='cpu'))
resume_checkpoint(net)
return net, trainDataLoader,validDataLoader
def train(net, trainDataLoader, epoch):
train_loss=0
train_rmse=0
train_PCK= 0
train_PCK2= [0, 0, 0, 0 ,0]
pckcount = [0, 0, 0, 0, 0]
net.train()
criterion = nn.MSELoss()
# optimizer = optim.SGD(net.parameters(), lr=config['lr'], momentum=config['momentum'] , weight_decay=config['weight_decay'])
optimizer = optim.Adam(net.parameters(), lr=config['lr'])
for i, data in enumerate(trainDataLoader):
inputs = data['image']
heatmaps_targets = data['heatmaps']
heatmaps_targets_2s = data['heatmaps_2s']
gts = data['keypoints']
mask, indices_valid, indices_notvalid = calculate_mask(heatmaps_targets)
heatmaps_targets = heatmaps_targets * mask
#2s
mask_2s, indices_valid_2s, indices_notvalid_2s = calculate_mask(heatmaps_targets_2s)
heatmaps_targets_2s = heatmaps_targets_2s * mask_2s
inputs = inputs.to(device)
heatmaps_targets = heatmaps_targets.to(device)
mask = mask.float().to(device)
#2s
heatmaps_targets_2s = heatmaps_targets_2s.to(device)
mask_2s = mask_2s.float().to(device)
optimizer.zero_grad()
outputs, outputs_2s = net(inputs)
if(i ==0):
print("outputs: ", len(outputs))
print("outputs_2s: ", len(outputs_2s))
#intermideate supervision
loss=0
loss_gm=0
loss_2s=0
#
for j in range(config['nstack']+2):
stackoutput = outputs[j] * mask
loss += criterion(stackoutput, heatmaps_targets)
# 2s
if (j < config['nstack']+1):
stackoutput_2s = outputs_2s[j] * mask_2s
loss_2s += criterion(stackoutput_2s, heatmaps_targets_2s)
loss = loss + loss_2s
# for j in range(config['nstack'] + 2):
# if(j!= config['nstack']-1):
# stackoutput = outputs[j] * mask
# loss += criterion(stackoutput, heatmaps_targets)
# else:
# stackoutput = outputs[j] * mask
# loss_gm = criterion(stackoutput, heatmaps_targets)
# # 2s
# if (j < config['nstack']+1):
# stackoutput_2s = outputs_2s[j] * mask_2s
# loss_2s += criterion(stackoutput_2s, heatmaps_targets_2s)
# loss = loss + 1.4 * loss_gm + loss_2s
loss.backward()
optimizer.step()
train_loss += loss.item()
#gt size
gts = gts.numpy()
gts = gts * [ 2 , 2]
# gts = gts.numpy() * [1 / 2, 1 / 2]
# evaluate
all_peak_points = get_peak_points(outputs_2s[config['nstack']].cpu().detach().numpy())
# all_peak_points = get_peak_points(outputs[config['nstack']].cpu().detach().numpy())
all_peak_points = all_peak_points * [2, 2]
rmse = get_mse(all_peak_points, gts, indices_valid_2s)
train_rmse += rmse
# PCK
acc,aveacc,_ = pck.eval(all_peak_points, gts, 0.2)
train_PCK += aveacc
# each part pck
for k in range(0, 5):
if acc[k] != -1:
pckcount[k] += 1
train_PCK2[k] += acc[k]
print('[ train---Epoch {:005d} -> {:005d} / {} ] loss : {:15} rmse : {:15} allPCK : {:5} {:5} {:5} {:5} {:5}'.format(
epoch, i * config['batch_size'], config['train_num'], loss, rmse, acc[0], acc[1], acc[2], acc[3], acc[4]))
trainallPCK = np.true_divide(train_PCK2, pckcount)
AvetrainPCK = (trainallPCK[1] + trainallPCK[2] + trainallPCK[3] + trainallPCK[4]) / 4
train_loss_everyepoch = train_loss / (i + 1)
train_rmse_everyepoch = train_rmse / (i + 1)
train_PCK_everyepoch = train_PCK / (i + 1)
# save result
dict = {'epoch': 0, 'lr': 0, 'loss': 0, 'rmse': 0, 'avePCKepoch': 0 , 'avePCK' : 0, 'part1PCK' : 0,
'part2PCK' :0, 'part3PCK' :0, 'part4PCK' :0}
dict['epoch'] = epoch
dict['lr'] = config['lr']
dict['loss'] = train_loss_everyepoch
dict['rmse'] = train_rmse_everyepoch.numpy()
dict['avePCKepoch'] = train_PCK_everyepoch
dict['avePCK'] = AvetrainPCK
dict['part1PCK'] = trainallPCK[1]
dict['part2PCK'] = trainallPCK[2]
dict['part3PCK'] = trainallPCK[3]
dict['part4PCK'] = trainallPCK[4]
df = pd.DataFrame([dict])
if epoch == 0:
df.to_csv(trainresultname)
else:
df.to_csv(trainresultname, mode='a', header=False)
print('[ train---Epoch {:005d} ] loss : {:15} rmse : {:15} allPCK : {:5} {:5} {:5} {:5} {:5} {:5} '.format(epoch, train_loss_everyepoch, train_rmse_everyepoch, train_PCK_everyepoch,
AvetrainPCK, trainallPCK[1], trainallPCK[2],trainallPCK[3], trainallPCK[4]))
return net, (train_loss_everyepoch,train_rmse_everyepoch, train_PCK_everyepoch)
def valid(net, validDataLoader,epoch):
lowrmse = 100000
valid_loss = 0
valid_rmse = 0
valid_PCK = 0
valid_PCK2 = [0, 0, 0, 0, 0]
pckcount = [0, 0, 0, 0, 0]
net.eval()
criterion = nn.MSELoss()
with torch.no_grad():
for i, data in enumerate(validDataLoader):
inputs = data['image']
heatmaps_targets = data['heatmaps']
heatmaps_targets_2s = data['heatmaps_2s']
gts = data['keypoints']
mask, indices_valid, indices_notvalid = calculate_mask(heatmaps_targets)
heatmaps_targets = heatmaps_targets * mask
#2s
mask_2s, indices_valid_2s, indices_notvalid_2s = calculate_mask(heatmaps_targets_2s)
heatmaps_targets_2s = heatmaps_targets_2s * mask_2s
inputs = inputs.to(device)
heatmaps_targets = heatmaps_targets.to(device)
mask = mask.float().to(device)
#2s
heatmaps_targets_2s = heatmaps_targets_2s.to(device)
mask_2s = mask_2s.float().to(device)
outputs, outputs_2s = net(inputs)
# intermideate supervision
loss = 0
loss_2s = 0
for j in range(config['nstack']+2):
stackoutput = outputs[j] * mask # outputs为8次stack每次输出预测值,8哥tensor的列表
loss += criterion(stackoutput, heatmaps_targets)
# 2s
if (j < config['nstack']+1 ):
stackoutput_2s = outputs_2s[j] * mask_2s # outputs为8次stack每次输出预测值,8哥tensor的列表
loss_2s += criterion(stackoutput_2s, heatmaps_targets_2s)
loss = loss + loss_2s
valid_loss += loss.item()
# gt size
gts = gts.numpy()
gts = gts* [ 2 , 2]
# gts = gts.numpy() * [1 / 2, 1 / 2]
all_peak_points = get_peak_points(outputs_2s[config['nstack']].cpu().detach().numpy())
# all_peak_points = get_peak_points(outputs[config['nstack']].cpu().detach().numpy())
all_peak_points = all_peak_points * [2, 2]
rmse = get_mse(all_peak_points, gts, indices_valid_2s) # 关键点坐标的loss
valid_rmse += rmse
# PCK
acc, aveacc, _ = pck.eval(all_peak_points, gts, 0.2)
valid_PCK += aveacc
# 计算每个部位的pck
for k in range(0, 5):
if acc[k] != -1:
pckcount[k] += 1
valid_PCK2[k] += acc[k]
print('[ valid---Epoch {:005d} -> {:005d} / {} ] loss : {:15} rmse : {:15} allPCK : {:5} {:5} {:5} {:5} {:5} {:5}'.format(
epoch, i, config['valid_num'], loss, rmse, aveacc, acc[0], acc[1], acc[2], acc[3], acc[4]))
validallPCK = np.true_divide(valid_PCK2, pckcount)
AvevalidPCK = (validallPCK[1] + validallPCK[2] + validallPCK[3] + validallPCK[4]) / 4
valid_loss_everyepoch = valid_loss / (i + 1)
valid_rmse_everyepoch = valid_rmse / (i + 1)
valid_PCK_everyepoch = valid_PCK / (i + 1)
# save result
validdict = {'epoch': 0, 'lr': 0,'loss': 0, 'rmse': 0, 'avePCKepoch': 0, 'avePCK': 0, 'part1PCK': 0,
'part2PCK': 0, 'part3PCK': 0, 'part4PCK': 0, }
validdict['epoch'] = epoch
validdict['r'] = config['lr']
validdict['loss'] = valid_loss_everyepoch
validdict['rmse'] = valid_rmse_everyepoch.numpy()
validdict['avePCKepoch'] = valid_PCK_everyepoch
validdict['avePCK'] = AvevalidPCK
validdict['part1PCK'] = validallPCK[1]
validdict['part2PCK'] = validallPCK[2]
validdict['part3PCK'] = validallPCK[3]
validdict['part4PCK'] = validallPCK[4]
df = pd.DataFrame([validdict])
if epoch == 0:
df.to_csv(validresultname)
else:
df.to_csv(validresultname, mode='a', header=False)
print('[ valid---Epoch {:005d} ] loss : {:15} rmse : {:15} allPCK : {:5}, {:5} {:5} {:5} {:5} {:5}'.format( epoch, valid_loss_everyepoch, valid_rmse_everyepoch, valid_PCK_everyepoch,
AvevalidPCK, validallPCK[1],
validallPCK[2], validallPCK[3],
validallPCK[4]
))
# Save checkpoint.
saveepochcheckpont_PCK(valid_PCK_everyepoch, config['highPCK'], epoch, net, netname)
return (valid_loss_everyepoch,valid_rmse_everyepoch, valid_PCK_everyepoch)
if __name__ == '__main__':
pck = PCK(njoints=4)
writer = SummaryWriter(comment='mousepose_train')
writer_total = SummaryWriter(comment='mousepose_valid')
net,trainDataLoader, validDataLoader = initnet()
updatelr = []
for epoch in range(config['start_epoch'],config['epoch_num']+config['start_epoch']):
running_loss = 0.0
nettrain, (train_loss_everyepoch,train_rmse_everyepoch, train_PCK_everyepoch) = train(net, trainDataLoader, epoch)
(valid_loss_everyepoch,valid_rmse_everyepoch, valid_PCK_everyepoch) = valid(nettrain, validDataLoader, epoch)
updatelr.append(valid_rmse_everyepoch)
writer.add_scalar('loss', train_loss_everyepoch, global_step=epoch)
writer.add_scalar('RMSE', train_rmse_everyepoch, global_step=epoch)
writer.add_scalar('PCK', train_PCK_everyepoch, global_step=epoch)
writer_total.add_scalar('loss', valid_loss_everyepoch, global_step=epoch)
writer_total.add_scalar('RMSE', valid_rmse_everyepoch, global_step=epoch)
writer_total.add_scalar('PCK', valid_PCK_everyepoch, global_step=epoch)