/
train_8_2.py
449 lines (379 loc) · 19.7 KB
/
train_8_2.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
449
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on 2020年3月6日18:52:46
2020年3月15日 20:21:37
@author: bdus
(5,3,16,112,112)
loss是修改过的 predict + construction 修改之前的用train_5_4.py
2020年3月18日 01:16:55
修改自train61
改动了光流loss 把loss3的系数减小了
"""
from __future__ import division
import argparse, time, logging, os, sys, math
os.environ['MXNET_CUDNN_AUTOTUNE_DEFAULT']='0'
os.environ['CUDA_VISIBLE_DEVICES']='0' #0,1
import numpy as np
import mxnet as mx
import gluoncv as gcv
from mxnet import gluon, nd, init, context
from mxnet import autograd as ag
from mxnet.gluon import nn
from mxnet.gluon.data.vision import transforms
from mxboard import SummaryWriter
from gluoncv.data.transforms import video
from gluoncv.data import UCF101, Kinetics400, SomethingSomethingV2, HMDB51
#from gluoncv.model_zoo import get_model
from gluoncv.utils import makedirs, LRSequential, LRScheduler, split_and_load, TrainingHistory
from model_zoo import get_model as myget
class AttrDisplay:
def gatherAttrs(self):
return ",".join("{}={}"
.format(k, getattr(self, k))
for k in self.__dict__.keys())
# attrs = []
# for k in self.__dict__.keys():
# item = "{}={}".format(k, getattr(self, k))
# attrs.append(item)
# return attrs
# for k in self.__dict__.keys():
# attrs.append(str(k) + "=" + str(self.__dict__[k]))
# return ",".join(attrs) if len(attrs) else 'no attr'
def __str__(self):
return "[{}:{}]".format(self.__class__.__name__, self.gatherAttrs())
class config(AttrDisplay):
def __init__(self):
self.predict_T = 3
self.use_take = True
self.new_length = 2*(16 + self.predict_T-1)
self.new_step = 1
self.model = 'r2plus1d_resnet18_kinetics400_custom'
self.use_kinetics_pretrain = False
self.TranConv_model = 'r2plus1d_resnet18_aetfc'#'r2plus1d_resnet34_tranconv_lateral_tanhbn'
self.use_lateral=True
self.freeze_lateral=False #True
self.save_dir = 'logs/param_rgb_r2plus1d_resnet18_kinetics400_custom_aet_train82'
self.num_classes = 51#101
#self.new_length_diff = self.new_length +1
self.dataset = 'hmdb51'#'ucf101'
self.train_dir = os.path.expanduser('~/.mxnet/datasets/ucf101/rawframes')#'/media/hp/mypan/BGSDecom/cv_MOG2/fgs')#
self.train_setting = '/home/hp/.mxnet/datasets/ucf101/ucfTrainTestlist/ucf101_train_split_1_rawframes.txt'
self.val_setting = '/home/hp/.mxnet/datasets/ucf101/ucfTrainTestlist/ucf101_val_split_1_rawframes.txt'
self.train_dir_hmdb51 = os.path.expanduser('~/.mxnet/datasets/hmdb51/rawframes')
self.train_setting_hmdb51 = '/home/hp/.mxnet/datasets/hmdb51/testTrainMulti_7030_splits/hmdb51_train_split_1_rawframes.txt'
self.val_setting_hmdb51 = '/home/hp/.mxnet/datasets/hmdb51/testTrainMulti_7030_splits/hmdb51_val_split_1_rawframes.txt'
self.logging_file = 'train.log'
self.name_pattern='img_%05d.jpg'
self.input_size=112#204#112
self.new_height=128#256#128
self.new_width=171#340#171
self.input_channel=3
self.num_segments= 1
self.num_workers =1
self.num_gpus = 1
self.per_device_batch_size = 8
self.lr = 0.1
self.lr_decay = 0.1
self.warmup_lr = 0.001
self.warmup_epochs = 25
self.momentum = 0.9
self.wd = 0.0005
self.prefetch_ratio = 1.0
self.use_amp = False
self.epochs = 300
self.lr_decay_epoch = [85,120]
self.lr_decay_period = 0
self.scale_ratios = [1.0, 0.8]#[1.0, 0.875, 0.75, 0.66]
self.dtype = 'float32'
self.pretrained_lateral_path = 'logs/param_rgb_r2plus1d_resnet18_kinetics400_custom_hmdb51_nlength16_lateral_scratch_tanh_pre_train61'
self.pretrained_lateral_file = '0.2824-hmdb51-r2plus1d_resnet34_tranconv_lateral_tanhbn-075-best.params'
self.use_pretrained = False
self.partial_bn = False
self.train_patterns = 'r2plus1d0_dense'#'r2plus1d1_dense'
self.use_train_patterns = False#True
self.freeze_patterns = '' #'net1'
self.freeze_lr_mult = 10 #set freezed base layer lr = self.lr * self.freeze_lr_mult
self.use_mult = False
self.clip_grad = 40
self.log_interval = 10
self.lr_mode = 'cosine'
self.resume_epoch = 0 #32
self.resume_path = ''#'logs/param_rgb_r2plus1d_resnet18_kinetics400_custom_hmdb51_nlength16_lateral_scratch_tanh_T_eq1'
self.resume_params =''# os.path.join(self.resume_path,'0.3758-hmdb51-r2plus1d_resnet18_kinetics400_custom-077-best.params')
self.resume_states =''# os.path.join(self.resume_path,'0.3758-hmdb51-r2plus1d_resnet18_kinetics400_custom-077-best.states')
self.reshape_type = 'tsn' #mxc3d c3d tsn tsn_newlength
opt = config()
makedirs(opt.save_dir)
filehandler = logging.FileHandler(os.path.join(opt.save_dir, opt.logging_file))
streamhandler = logging.StreamHandler()
logger = logging.getLogger('')
logger.setLevel(logging.INFO)
logger.addHandler(filehandler)
logger.addHandler(streamhandler)
logger.info(opt)
# number of GPUs to use
num_gpus = opt.num_gpus
ctx = [mx.gpu(i) for i in range(num_gpus)]
#ctx = [mx.gpu(1)]
# Get the model
net = myget(name=opt.model, nclass=opt.num_classes, num_segments=opt.num_segments,input_channel=opt.input_channel,batch_normal=opt.partial_bn,use_lateral=opt.use_lateral,use_kinetics_pretrain=opt.use_kinetics_pretrain)
net.cast(opt.dtype)
net.collect_params().reset_ctx(ctx)
net1 = myget(name=opt.TranConv_model)
net1.cast(opt.dtype)
net1.collect_params().reset_ctx(ctx)
#logger.info(net)
if opt.resume_params is not '':
net.load_parameters(opt.resume_params, ctx=ctx)
#if opt.use_pretrained:
#net.features_3d.load_parameters(opt.pretrained_ECOfeature3d,ctx=ctx,allow_missing=True)
#net.output.load_parameters(opt.pretrained_ECOoutput,ctx=ctx,allow_missing=True)
#logger.info('use pretrained model : %s , %s',opt.pretrained_ECOfeature3d,opt.pretrained_ECOoutput)
if opt.use_pretrained:
modelpath = os.path.join(opt.pretrained_lateral_path,opt.pretrained_lateral_file)
modelfile = os.path.expanduser(modelpath)
net1.load_parameters(modelfile,ctx=ctx,allow_missing=True)
logger.info('use pretrained model : %s',modelfile)
if opt.use_mult:
#net.collect_params(opt.freeze_patterns).setattr('lr_mult',opt.freeze_lr_mult)
net1.collect_params().setattr('lr_mult',opt.freeze_lr_mult)
logger.info(net)
net.collect_params().reset_ctx(ctx)
transform_train = video.VideoGroupTrainTransform(size=(opt.input_size, opt.input_size), scale_ratios=opt.scale_ratios, mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
transform_test = video.VideoGroupValTransform(size=opt.input_size, mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
# Batch Size for Each GPU
per_device_batch_size = opt.per_device_batch_size
# Number of data loader workers
num_workers = opt.num_workers
# Calculate effective total batch size
batch_size = per_device_batch_size * num_gpus
# Set train=True for training data. Here we only use a subset of UCF101 for demonstration purpose.
# The subset has 101 training samples, one sample per class.
if opt.dataset == 'ucf101':
train_dataset = UCF101(setting=opt.train_setting, root=opt.train_dir, train=True,
new_width=opt.new_width, new_height=opt.new_height, new_length=opt.new_length,new_step=opt.new_step,
target_width=opt.input_size, target_height=opt.input_size,
num_segments=opt.num_segments, transform=transform_train)
val_dataset = UCF101(setting=opt.val_setting, root=opt.train_dir, train=False,
new_width=opt.new_width, new_height=opt.new_height, new_length=opt.new_length,new_step=opt.new_step,
target_width=opt.input_size, target_height=opt.input_size,
num_segments=opt.num_segments, transform=transform_test)
elif opt.dataset == 'hmdb51':
train_dataset = HMDB51(setting=opt.train_setting_hmdb51, root=opt.train_dir_hmdb51, train=True,
new_width=opt.new_width, new_height=opt.new_height, new_length=opt.new_length, new_step=opt.new_step,
target_width=opt.input_size, target_height=opt.input_size,
num_segments=opt.num_segments, transform=transform_train)
val_dataset = HMDB51(setting=opt.val_setting_hmdb51, root=opt.train_dir_hmdb51, train=False,
new_width=opt.new_width, new_height=opt.new_height, new_length=opt.new_length, new_step=opt.new_step,
target_width=opt.input_size, target_height=opt.input_size,
num_segments=opt.num_segments, transform=transform_test)
else:
logger.info('Dataset %s is not supported yet.' % (opt.dataset))
train_data = gluon.data.DataLoader(train_dataset, batch_size=batch_size, shuffle=True, num_workers=num_workers,
prefetch=int(opt.prefetch_ratio * num_workers), last_batch='rollover')
val_data = gluon.data.DataLoader(val_dataset, batch_size=batch_size, shuffle=False, num_workers=num_workers,
prefetch=int(opt.prefetch_ratio * num_workers), last_batch='discard')
train_data = gluon.data.DataLoader(train_dataset, batch_size=batch_size,
shuffle=True, num_workers=num_workers)
logger.info('Load %d training samples.' % len(train_dataset))
val_data = gluon.data.DataLoader(val_dataset, batch_size=batch_size,
shuffle=False, num_workers=num_workers)
def batch_fn(batch, ctx):
data = split_and_load(batch[0], ctx_list=ctx, batch_axis=0, even_split=False)
label = split_and_load(batch[1], ctx_list=ctx, batch_axis=0, even_split=False)
return data, label
# Learning rate decay factor
lr_decay = opt.lr_decay
# Epochs where learning rate decays
#lr_decay_epoch = opt.lr_decay_epoch
lr_decay_period = opt.lr_decay_period
if opt.lr_decay_period > 0:
lr_decay_epoch = list(range(lr_decay_period, opt.num_epochs, lr_decay_period))
else:
lr_decay_epoch = opt.lr_decay_epoch
lr_decay_epoch = [e - opt.warmup_epochs for e in lr_decay_epoch]
# Stochastic gradient descent
optimizer = 'sgd'
# Set parameters
optimizer_params = {'learning_rate': opt.lr, 'wd': opt.wd, 'momentum': opt.momentum}
num_batches = len(train_data)
lr_scheduler = LRSequential([
LRScheduler('linear', base_lr=opt.warmup_lr, target_lr=opt.lr,
nepochs=opt.warmup_epochs, iters_per_epoch=num_batches),
LRScheduler(opt.lr_mode, base_lr=opt.lr, target_lr=0,
nepochs=opt.epochs - opt.warmup_epochs,
iters_per_epoch=num_batches,
step_epoch=lr_decay_epoch,
step_factor=lr_decay, power=2)
])
optimizer_params['lr_scheduler'] = lr_scheduler
if opt.partial_bn:
train_patterns = None
if 'inceptionv3' in opt.model:
train_patterns = '.*weight|.*bias|inception30_batchnorm0_gamma|inception30_batchnorm0_beta|inception30_batchnorm0_running_mean|inception30_batchnorm0_running_var'
else:
logger.info('Current model does not support partial batch normalization.')
trainer = gluon.Trainer(net.collect_params(train_patterns), optimizer, optimizer_params, update_on_kvstore=False)
elif opt.partial_bn == False and opt.use_train_patterns == True:
logger.info('========\n %s' % net.collect_params() )
trainer = gluon.Trainer(net.collect_params(opt.train_patterns), optimizer, optimizer_params, update_on_kvstore=False)
logger.info('trainner.patterns: %s.' % opt.train_patterns )
logger.info('========\n %s' % net.collect_params(opt.train_patterns) )
elif opt.use_lateral and not opt.freeze_lateral:
print("============== use_lateral")
lst = list(net.collect_params().values()) + list(net1.collect_params().values())
trainer = gluon.Trainer(lst, optimizer, optimizer_params, update_on_kvstore=False)
else:
print("============== training net0. net1 is not included in training")
trainer = gluon.Trainer(net.collect_params(), optimizer, optimizer_params, update_on_kvstore=False)
if opt.resume_states is not '':
trainer.load_states(opt.resume_states)
# Define our trainer for net
#trainer = gluon.Trainer(net.collect_params(), optimizer, optimizer_params)
loss_fn = gluon.loss.SoftmaxCrossEntropyLoss()
loss_l2 = gluon.loss.L2Loss(weight=1.0)
loss_l2.initialize()
def takeT(X,T=0):
#idx = nd.array(nd.arange(T,opt.new_length,2),ctx=ctx[0])
idx = nd.array([2*n+T for n in range(16)],ctx=ctx[0])
return nd.take(X,idx,axis=3)
if opt.use_take:
print('==============================================',opt.new_length)
print([2*n+0 for n in range(16)])#nd.arange(0,opt.new_length,2))
for T in range(opt.predict_T+1):
print([2*n+T for n in range(16)])#nd.arange(opt.predict_T,opt.new_length,2))
train_metric = mx.metric.Accuracy()
train_metric_aet = mx.metric.Accuracy()
train_history = TrainingHistory(['training-acc','val-top1-acc','val-top5-acc','training-loss','loss_val','loss_aets','valaet'])
# train_history.update([acc,acc_top1_val,acc_top5_val,train_loss/(i+1),loss_val,loss_aet,valaet])
lr_decay_count = 0
best_val_score = 0
acc_top1 = mx.metric.Accuracy()
acc_top5 = mx.metric.TopKAccuracy(5)
acc_valAET = mx.metric.Accuracy()
def test(ctx,val_data):
acc_top1.reset()
acc_top5.reset()
acc_valAET.reset()
L = gluon.loss.SoftmaxCrossEntropyLoss()
num_test_iter = len(val_data)
val_loss_epoch = 0
val_loss_epoch1 = 0
# val_loss_epoch2 = 0
# val_loss_epoch3 = 0
for i, batch in enumerate(val_data):
data, label = batch_fn(batch, ctx)
val_outputs = []
if opt.use_lateral:
val_l2out = []
val_l2label = []
for X, y in zip(data,label):
X1 = takeT(X)
X1 = X1.reshape((-1,) + X1.shape[2:]) # for reconstraction
pred, latel= net(X1.astype(opt.dtype, copy=False))
val_outputs.append(pred)
for T in range(opt.predict_T+1):
X2 = takeT(X,T=T)
X2 = X2.reshape((-1,) + X2.shape[2:]) # for prodiction
pred1, latel1= net(X2.astype(opt.dtype, copy=False))
val_t_hat = net1(latel[3].astype(opt.dtype, copy=False),
latel1[3].astype(opt.dtype, copy=False))
val_l2label.append(T* nd.ones(shape=val_t_hat.shape[0],ctx=ctx[0]))
val_l2out.append(val_t_hat)
loss0 = loss_fn(pred, y.astype(opt.dtype, copy=False)) + loss_fn(pred1, y.astype(opt.dtype, copy=False))
loss1 = loss_fn(val_t_hat, T* nd.ones(shape=val_t_hat.shape[0],ctx=ctx[0]) )
loss = loss0+loss1
# Update metrics
val_loss_epoch += loss0.mean().asscalar() / len(label)
val_loss_epoch1 += loss1.mean().asscalar() / len(label)
acc_top1.update(label, val_outputs)
acc_top5.update(label, val_outputs)
acc_valAET.update(val_l2label, val_l2out)
_, top1 = acc_top1.get()
_, top5 = acc_top5.get()
_, valaet = acc_valAET.get()
val_loss = val_loss_epoch / num_test_iter
loss_aet = val_loss_epoch1 / num_test_iter
return (top1, top5, val_loss, loss_aet,valaet)
# acc_top1_val, acc_top5_val, loss_val, loss_mse, valaet = test(ctx, val_data)
# training
for epoch in range(opt.resume_epoch, opt.epochs):
tic = time.time()
train_metric.reset()
train_metric_aet.reset()
train_loss = 0
mse_loss = 0
pre_loss = 0
flow_loss = 0
btic = time.time()
# Loop through each batch of training data
for i, batch in enumerate(train_data):
# Extract data and label
data, label = batch_fn(batch, ctx)
# AutoGrad
output = []
l2out = []
l2label = []
# l2latel1 = []
for X, y in zip(data,label):
#print('==================================',X.shape,(opt.new_length - opt.predict_T))
X1 = takeT(X)
X1 = X1.reshape((-1,) + X1.shape[2:]) # for reconstraction
with ag.record():
pred, latel= net(X1.astype(opt.dtype, copy=False))
output.append(pred)
for T in range(opt.predict_T+1):
X2 = takeT(X,T=T)
X2 = X2.reshape((-1,) + X2.shape[2:]) # for prodiction
with ag.record():
pred1, latel1= net(X2.astype(opt.dtype, copy=False))
t_hat = net1(latel[3].astype(opt.dtype, copy=False),
latel1[3].astype(opt.dtype, copy=False))
#l2latel.append(latel)
#l2latel1.append(latel1)
loss0 = loss_fn(pred, y.astype(opt.dtype, copy=False)) + loss_fn(pred1, y.astype(opt.dtype, copy=False))
loss1 = loss_fn(t_hat, T* nd.ones(shape=t_hat.shape[0],ctx=ctx[0]) )
loss = loss0+loss1
loss.backward()
l2label.append(T* nd.ones(shape=t_hat.shape[0],ctx=ctx[0]))
l2out.append(t_hat)
# Optimize
trainer.step(batch_size,ignore_stale_grad=True)
# Update metrics
train_loss += loss0.mean().asscalar() / len(label) #sum([l.mean().asscalar() for l in loss]) / len(loss)
mse_loss += loss1.mean().asscalar() / len(label) #sum([l.mean().asscalar() for l in loss1]) / len(loss1)
# pre_loss += 0
# flow_loss += 0
train_metric.update(label, output)
train_metric_aet.update(l2label, l2out)
if i % opt.log_interval == 0:
name, acc = train_metric.get()
_, accaet = train_metric_aet.get()
logger.info('[Epoch %d] [%d | %d] train=%f tAET=%f loss=%f aetloss=%f time: %f' %
(epoch,i,len(train_data), acc,accaet, train_loss / (i+1),mse_loss/(i+1), time.time()-btic) )
btic = time.time()
name, acc = train_metric.get()
_, accaet = train_metric_aet.get()
# test
#acc_top1_val, acc_top5_val, loss_val = test(ctx, val_data)
acc_top1_val, acc_top5_val, loss_val, loss_aet, valaet = test(ctx, val_data)
# Update history and print metrics
train_history.update([acc,acc_top1_val,acc_top5_val,train_loss/(i+1),loss_val,loss_aet,valaet])
train_history.plot(save_path=os.path.join(opt.save_dir,'trainlog_wth.jpg'))
#train_history.plot(save_path=os.path.join(opt.save_dir,'trainlog_acc.jpg'),labels = ['training-acc','val-top1-acc','val-top5-acc'])
#train_history.plot(save_path=os.path.join(opt.save_dir,'trainlog_loss.jpg'),labels = ['training-loss','cross-loss','mse-loss','pre-loss'])
logger.info('[Epoch %d] train=%f loss=%f time: %f' %
(epoch, acc, train_loss / (i+1), time.time()-tic))
#logger.info('[Epoch %d] val top1 =%f top5=%f val loss=%f,lr=%f' %
# (epoch, acc_top1_val, acc_top5_val, loss_val ,trainer.learning_rate ))
logger.info('[Epoch %d] val top1 =%f top5=%f val loss=%f,loss_aet=%f,valaet = %f, lr=%f' %
(epoch, acc_top1_val, acc_top5_val, loss_val ,loss_aet,valaet,trainer.learning_rate ))
if acc_top1_val > best_val_score and epoch > 5:
best_val_score = acc_top1_val
net.save_parameters('%s/%.4f-%s-%s-%03d-best.params'%(opt.save_dir, best_val_score, opt.dataset, opt.model, epoch))
trainer.save_states('%s/%.4f-%s-%s-%03d-best.states'%(opt.save_dir, best_val_score, opt.dataset, opt.model, epoch))
if opt.use_lateral:
net1.save_parameters('%s/%.4f-%s-%s-%03d-best.params'%(opt.save_dir, best_val_score, opt.dataset, opt.TranConv_model, epoch))
# We can plot the metric scores with:
train_history.plot(save_path=os.path.join(opt.save_dir,'trainlog_final.jpg'))