-
Notifications
You must be signed in to change notification settings - Fork 4
/
train_gazenet_mxJiang.py
246 lines (219 loc) · 12.7 KB
/
train_gazenet_mxJiang.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
# convert PyTorch gazenet scripts to MXNet
from mxboard import SummaryWriter
import sys, os, argparse, time
import numpy as np
import cv2
import matplotlib.pyplot as plt
import mxnet as mx
from mxnet import gluon
from mxnet.gluon import nn
from mxnet.autograd import record
from mxnet.gluon.data import DataLoader
from mxnet.gluon.data.vision import transforms
import mxnet.gluon.data.vision
import datasets_mxJiang, gazenet_mxJiang_yaw_zoo, gazenet_mxJiang_pitch_zoo,gazenet_mxJiang_roll_zoo
import mxnet.gluon.model_zoo as model_zoo
import pickle as pk
from mxnet import ndarray as nd
from mxnet.gluon import HybridBlock
import datetime
import copy
from mxnet.gluon.loss import Loss
import time
def parse_args():
"""Parse input arguments."""
parser = argparse.ArgumentParser(description='Head pose estimation using the gazenet network.')
parser.add_argument('--gpu', dest='gpu_id', help='GPU device id to use [0]',
default=0, type=int)
parser.add_argument('--num_epochs', dest='num_epochs', help='Maximum number of training epochs.',
default=5, type=int)
parser.add_argument('--batch_size', dest='batch_size', help='Batch size.',
default=16, type=int)
parser.add_argument('--lr', dest='lr', help='Base learning rate.',
default=0.001, type=float)
parser.add_argument('--dataset', dest='dataset', help='Dataset type.', default='Pose_300W_LP', type=str)
parser.add_argument('--data_dir', dest='data_dir', help='Directory path for data.',
default='', type=str)
parser.add_argument('--filename_list', dest='filename_list', help='Path to text file containing relative paths for every example.',
default='', type=str)
parser.add_argument('--output_string', dest='output_string', help='String appended to output snapshots.', default = '', type=str)
parser.add_argument('--alpha', dest='alpha', help='Regression loss coefficient.',
default=0.001, type=float)
parser.add_argument('--snapshot', dest='snapshot', help='Path of model snapshot.',
default='', type=str)
args = parser.parse_args()
return args
def softmax(x, axis=None):
x = x - x.max(axis=axis, keepdims=True)
y = np.exp(x)
return y / y.sum(axis=axis, keepdims=True)
def reg_criterion(A, B):
mse = ((A - B)**2)
return mse
class compute_cost(Loss):
def __init__(self, weight=1., batch_axis=0, **kwargs):
super(compute_cost, self).__init__(weight, batch_axis, **kwargs)
def hybrid_forward(self, F, yaw, label_yaw, label_yaw_cont, idx_tensor):
loss_yaw = criterion(yaw, label_yaw)
yaw_predicted = softmax(yaw.asnumpy(),axis=1)
#print('loss_entropy',loss_yaw)
yaw_predicted = np.sum(yaw_predicted * idx_tensor, 1) * 3 - 99
loss_reg_yaw = nd.array(reg_criterion(yaw_predicted, label_yaw_cont.asnumpy())).as_in_context(ctx)
#print('loss_reg',loss_reg_yaw)
total_loss_yaw = loss_yaw + alpha * loss_reg_yaw
#print('total_loss', total_loss_yaw)
return yaw_predicted, total_loss_yaw
if __name__ == '__main__':
mx.random.seed(int(time.time()))
args = parse_args()
num_epochs = args.num_epochs
batch_size = args.batch_size
gpu = args.gpu_id
lr = args.lr
lr_period = 15
lr_decay = 0.1
ctx = mx.gpu(gpu)
model_yaw = gazenet_mxJiang_yaw_zoo.Gazenet_mxJiang_yaw_zoo(66,ctx=ctx)
model_pitch = gazenet_mxJiang_pitch_zoo.Gazenet_mxJiang_pitch_zoo(66,ctx=ctx)
model_roll = gazenet_mxJiang_roll_zoo.Gazenet_mxJiang_roll_zoo(66,ctx=ctx)
print('Loading data.')
transformations = transforms.Compose([transforms.Resize(240),
#transforms.RandomResizedCrop(224),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])
if args.dataset == 'Pose_300W_LP':
pose_dataset = datasets_mxJiang.Pose_300W_LP(args.data_dir, args.filename_list, transformations)
elif args.dataset == 'Pose_300W_LP_random_ds':
pose_dataset = datasets_mxJiang.Pose_300W_LP_random_ds(args.data_dir, args.filename_list, transformations)
elif args.dataset == 'Synhead':
pose_dataset = datasets_mxJiang.Synhead(args.data_dir, args.filename_list, transformations)
elif args.dataset == 'AFLW2000':
pose_dataset = datasets_mxJiang.AFLW2000(args.data_dir, args.filename_list, transformations)
elif args.dataset == 'BIWI':
pose_dataset = datasets_mxJiang.BIWI(args.data_dir, args.filename_list, transformations)
elif args.dataset == 'AFLW':
pose_dataset = datasets_mxJiang.AFLW(args.data_dir, args.filename_list, transformations)
elif args.dataset == 'AFLW_aug':
pose_dataset = datasets_mxJiang.AFLW_aug(args.data_dir, args.filename_list, transformations)
elif args.dataset == 'AFW':
pose_dataset = datasets_mxJiang.AFW(args.data_dir, args.filename_list, transformations)
else:
print('Error: not a valid dataset name')
sys.exit()
train_loader = gluon.data.DataLoader(dataset=pose_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=1)
criterion = gluon.loss.SoftmaxCrossEntropyLoss()
criterion_reg = gluon.loss.L2Loss()
comp_cost = compute_cost()
alpha = args.alpha
idx_tensor = [idx for idx in range(66)]
idx_tensor = np.float32(idx_tensor)
#get_ignored_params(model) lr=0
model_yaw.collect_params('resnetv10_conv0_weight.*').setattr('lr_mult', 0)
model_yaw.collect_params('resnetv10_batchnorm0_gamma').setattr('lr_mult', 0)
model_yaw.collect_params('resnetv10_batchnorm0_betta').setattr('lr_mult', 0)
model_pitch.collect_params('resnetv10_conv0_weight.*').setattr('lr_mult', 0)
model_pitch.collect_params('resnetv10_batchnorm0_gamma').setattr('lr_mult', 0)
model_pitch.collect_params('resnetv10_batchnorm0_betta').setattr('lr_mult', 0)
model_roll.collect_params('resnetv10_conv0_weight.*').setattr('lr_mult', 0)
model_roll.collect_params('resnetv10_batchnorm0_gamma').setattr('lr_mult', 0)
model_roll.collect_params('resnetv10_batchnorm0_betta').setattr('lr_mult', 0)
#get_ignored_params(model) model.eval()
#model_yaw.collect_params('resnetv10_batchnorm0_running_mean').setattr('momentum',1)
#model_yaw.collect_params('resnetv10_batchnorm0_running_var').setattr('momentum',1)
#model_pitch.collect_params('resnetv10_batchnorm0_running_mean').setattr('momentum',0.01)
#model_pitch.collect_params('resnetv10_batchnorm0_running_var').setattr('momentum',0.01)
#model_roll.collect_params('resnetv10_batchnorm0_running_mean').setattr('momentum',0.01)
#model_roll.collect_params('resnetv10_batchnorm0_running_var').setattr('momentum',0.01)
#get_fc_params(model)
model_yaw.collect_params('dense.*').setattr('lr_mult', 5)
model_pitch.collect_params('dense.*').setattr('lr_mult', 5)
model_roll.collect_params('dense.*').setattr('lr_mult', 5)
model_yaw.model_yaw.bias.set_data(mx.nd.random.uniform(-0.2,0.2,shape=(66,),ctx=ctx))
optimizer_yaw = gluon.Trainer(model_yaw.collect_params(),'adam', {'learning_rate': lr})
optimizer_pitch = gluon.Trainer(model_pitch.collect_params(),'adam', {'learning_rate': lr})
optimizer_roll = gluon.Trainer(model_roll.collect_params(),'adam', {'learning_rate': lr})
#print('model_yaw collect prams==> ',model_yaw.collect_params())
# ResNet50 structure
model_yaw.hybridize()
model_pitch.hybridize()
model_roll.hybridize()
unique_id = str(datetime.datetime.now()).replace(' ', '-').replace(':', '-').replace('.','-')
with SummaryWriter(logdir='/data/deep-head-pose/logs_gluon/logs-'+unique_id, flush_secs=5) as sw:
try:
global_step = 0
for epoch in range(num_epochs):
thefile = open('/data/deep-head-pose/test'+'-'+str(epoch)+str(datetime.datetime.now())+'.txt', 'w')
#if epoch > 0 and epoch % lr_period == 0:
# optimizer.set_learning_rate(optimizer.learning_rate * lr_decay)
#if epoch == 20:
# model_yaw.net.collect_params('resnetv10_stage.*_.*').setattr('lr_mult', 0.001)
# model_pitch.net.collect_params('resnetv10_stage.*_.*').setattr('lr_mult', 0.001)
# model_roll.net.collect_params('resnetv10_stage.*_.*').setattr('lr_mult', 0.001)
for batch_idx, (images, labels, cont_labels) in enumerate(train_loader):
#print('images=>',images)
global_step += 1
images = images.as_in_context(ctx)
labels.as_in_context(ctx)
cont_labels = cont_labels.as_in_context(ctx)
# Binned labels
label_yaw = labels[:,0].as_in_context(ctx)
label_pitch = labels[:,1].as_in_context(ctx)
label_roll = labels[:,2].as_in_context(ctx)
# Continuous labels
label_yaw_cont = cont_labels[:,0].as_in_context(ctx)
label_pitch_cont = cont_labels[:,1].as_in_context(ctx)
label_roll_cont = cont_labels[:,2].as_in_context(ctx)
yaw = nd.ones((1,66))
yaw.attach_grad()
with mx.autograd.record():
yaw = model_yaw(images)
#print('yaw',yaw.mean())
yaw_predicted, total_loss_yaw = comp_cost(yaw, label_yaw, label_yaw_cont, idx_tensor)
#print('yaw_predicted', yaw_predicted)
mx.autograd.backward(total_loss_yaw.nansum().as_in_context(ctx))
if global_step%10 == 0:
sw.add_scalar(tag='loss_yaw',value=total_loss_yaw.nansum().asnumpy(), global_step=global_step)
#print('loss_yaw', total_loss_yaw)
#print('fc_yaw grad==>', [w.grad() for u, w in model_yaw.model_yaw.collect_params().items()])
optimizer_yaw.step(batch_size)
#print('yaw_grad==>', yaw.grad)
with mx.autograd.record():
pitch = model_pitch(images)
#print('pitch',pitch)
pitch_predicted, total_loss_pitch = comp_cost(pitch, label_pitch, label_pitch_cont, idx_tensor)
#print('pitch_predicted',pitch_predicted)
mx.autograd.backward(total_loss_pitch.nansum().as_in_context(ctx))
if global_step%10 == 0:
sw.add_scalar(tag='loss_pitch',value=total_loss_pitch.mean().asnumpy(), global_step=global_step)
optimizer_pitch.step(batch_size)
with mx.autograd.record():
roll = model_roll(images)
#print('roll',roll)
roll_predicted, total_loss_roll = comp_cost(roll, label_roll, label_roll_cont, idx_tensor)
#print('roll_predicted',roll_predicted)
mx.autograd.backward(total_loss_roll.nansum().as_in_context(ctx))
if global_step%10 == 0:
sw.add_scalar(tag='loss_roll',value=total_loss_roll.mean().asnumpy(), global_step=global_step)
sw.add_scalar(tag='epoch', value=epoch, global_step=global_step)
optimizer_roll.step(batch_size)
loss_yaw_sum = total_loss_yaw.mean().asscalar()
loss_pitch_sum = total_loss_pitch.mean().asscalar()
loss_roll_sum = total_loss_roll.mean().asscalar()
now_lr = optimizer_yaw.learning_rate
thefile.write("Epoch: %d; Batch %d; Loss Yaw %f; Loss Pitch %f; Loss Roll %f \n" % (epoch, batch_idx, loss_yaw_sum, loss_pitch_sum, loss_roll_sum))
if batch_idx%100 == 0:
print ("Epoch: %d; Batch %d; Losses of Yaw: %f; Pitch: %f; Roll: %f; Now_Lr: %f" % (epoch, batch_idx, loss_yaw_sum, loss_pitch_sum, loss_roll_sum, now_lr))
print('predicted==>', yaw_predicted, pitch_predicted, roll_predicted)
thefile.close()
model_yaw.export('/data/deep-head-pose/yawparams-'+unique_id+str(epoch))
model_pitch.export('/data/deep-head-pose/pitchparams-'+unique_id+str(epoch))
model_roll.export('/data/deep-head-pose/rollparams-'+unique_id+str(epoch))
sw.close()
except KeyboardInterrupt:
print("KeyboardInterrupted")
sw.close()
exit(0)
pass