-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
358 lines (338 loc) · 15.9 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
import os
import cv2
import tensorflow as tf
# import tensorflow.contrib as tc
slim = tf.contrib.slim
import numpy as np
import pdb
from blaze_loss import _Loss
from utils import bbox_utils,train_utils
from angment import _crop,_fliph,tcs,_expand,_distort
batch_norm_params = {
'scale': True,
'decay': 0.99,
'epsilon': 0.0001
}
hyper_params = train_utils.get_hyper_params()
def uint82float(a, m, stdv):
a = a.astype('float32')
return (a - m) * stdv
def bbox_iou(pb, gt):
tl = np.maximum(pb[0, 0:2], gt[0, 0:2])
br = np.minimum(pb[0, 2:4], gt[0, 2:4])
wl = np.minimum(pb[0, 0:2], gt[0, 0:2])
wr = np.maximum(pb[0, 2:4], gt[0, 2:4])
insect = br - tl
insect = np.clip(insect, 0, 99999)
pw = (pb[0, 2] - pb[0, 0])
ph = (pb[0, 3] - pb[0, 1])
gw = (gt[0, 2] - gt[0, 0])
gh = (gt[0, 3] - gt[0, 1])
area_i = insect[0] * insect[1]
area_a = pw * ph
area_b = gw * gh
outRect = wr - wl
outRect = outRect[0] * outRect[1]
iou = area_i / (area_a + area_b - area_i)
# outRect = iou - (outRect-(area_a + area_b - area_i))/outRect
return iou
def nms(conf, boxes):#conf [n] boxes[n*4]
ind = np.argsort(conf)
ind = ind[::-1]
conf = conf[ind]
boxes = boxes[ind, :]
dp = []
for i in range(conf.shape[0]):
if i in dp:
continue
for j in range(i + 1, conf.shape[0]):
if j in dp:
continue
iou = bbox_iou(boxes[i:i+1, :], boxes[j:j+1, :])
if iou > 0.15:
dp.append(j)
res = []
for i in range(ind.shape[0]):
if i not in dp:
res.append(ind[i])
return np.asarray(res)
def _parse_function(serialized_example):
features = tf.parse_single_example(
serialized_example,
features={
'boxes': tf.FixedLenFeature([9*4],tf.float32),
'img_raw':tf.FixedLenFeature([],tf.string)
}
)
image = tf.decode_raw(features['img_raw'], tf.uint8) # compare with input of tfrecord
image = tf.cast(image, tf.float32)
boxes = tf.cast(features['boxes'], tf.float32)
boxes = tf.reshape(boxes,(9,4))
image = tf.reshape(image,(320,320,3))
imbatch = image / 255.0
return imbatch,boxes
class BlazeFace(object):
def __init__(self, is_training=True, input_w=128,input_h=128):
self.input_w = input_w
self.input_h = input_h
self.is_training = is_training
self.normalizer = tf.keras.layers.BatchNormalization
self._create_placeholders()
xs = np.zeros((1,1344,4))
xs[:,:,0:2] = -3
self.mybias = tf.constant(xs, dtype=tf.float32)
self._build_model()
def _create_placeholders(self):
self.input = tf.placeholder(dtype=tf.float32, shape=[None, self.input_h, self.input_w, 3], name='input')
def _build_model(self):
self.output = self._Conv2d(self.input,64,5,2)
self.output = self.BlazeBlock(self.output, 64, 5, 1)
self.output = self.BlazeBlock(self.output, 64, 5, 1)
self.output = self.BlazeBlock(self.output, 96, 5, 2)
self.output = self.BlazeBlock(self.output, 96, 5, 1)
self.output = self.BlazeBlock(self.output, 96, 5, 1)
self.output = self.Double_BlazeBlock(self.output, 64, 192, 5, 2)
self.output = self.Double_BlazeBlock(self.output, 64, 192, 5, 1)
self.output16 = self.Double_BlazeBlock(self.output, 64, 192, 5, 1)
self.loc16 = self._Loc16(self.output16,5,3,1)*0.16666
def BlazeBlock(self, input, out, kernel_size,stride):
with slim.arg_scope([slim.conv2d], padding='SAME',
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params,
activation_fn=tf.nn.relu6,
weights_initializer=tf.contrib.layers.xavier_initializer(),
weights_regularizer=slim.l2_regularizer(0.),
biases_initializer=None):
with slim.arg_scope([slim.batch_norm], center=True, scale=True, is_training=self.is_training):
output = slim.separable_conv2d(input, num_outputs=None, kernel_size=kernel_size, stride=stride,
padding='SAME',
normalizer_fn=None,
normalizer_params=None,
activation_fn=None,
weights_initializer=tf.contrib.layers.xavier_initializer(),
weights_regularizer=slim.l2_regularizer(0.),
biases_initializer=None)
output = slim.conv2d(output, num_outputs=out, kernel_size=1)
# shortcut = slim.max_pool2d(input,kernel_size = stride,stride = stride,padding = 'SAME')
# shortcut = slim.conv2d(shortcut, num_outputs=out, kernel_size=1)
return output#+shortcut
def Double_BlazeBlock(self, input, mid,out, kernel_size,stride):
with slim.arg_scope([slim.conv2d], padding='SAME',
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params,
activation_fn=tf.nn.relu6,
weights_initializer=tf.contrib.layers.xavier_initializer(),
weights_regularizer=slim.l2_regularizer(0.),
biases_initializer=None):
with slim.arg_scope([slim.batch_norm], center=True, scale=True, is_training=self.is_training):
output = slim.separable_conv2d(input, num_outputs=None, kernel_size=kernel_size, stride=stride,
padding='SAME',
normalizer_fn=None,
normalizer_params=None,
activation_fn=None,
weights_initializer=tf.contrib.layers.xavier_initializer(),
weights_regularizer=slim.l2_regularizer(0.1),
biases_initializer=None)
output = slim.conv2d(output, num_outputs=mid, kernel_size=1)
output = slim.separable_conv2d(output, num_outputs=None, kernel_size=kernel_size, stride=1,
padding='SAME',
normalizer_fn=None,
normalizer_params=None,
activation_fn=None,
weights_initializer=tf.contrib.layers.xavier_initializer(),
weights_regularizer=slim.l2_regularizer(0.),
biases_initializer=None)
output = slim.conv2d(output, num_outputs=out, kernel_size=1)
return output
def _Conv2d(self, input, out, kernel_size,stride):
with slim.arg_scope([slim.conv2d], padding='SAME',
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params,
activation_fn=tf.nn.relu6,
weights_initializer=tf.contrib.layers.xavier_initializer(),
weights_regularizer=slim.l2_regularizer(0.),
biases_initializer=None):
with slim.arg_scope([slim.batch_norm], center=True, scale=True, is_training=self.is_training):
output = slim.conv2d(input, num_outputs=out, kernel_size=kernel_size,stride = stride)
return output
def _Loc16(self, input, out, kernel_size,stride):
with slim.arg_scope([slim.conv2d], padding='SAME',
normalizer_fn=None,
normalizer_params=None,
activation_fn=tf.nn.relu6,
weights_initializer=tf.contrib.layers.xavier_initializer(),
weights_regularizer=slim.l2_regularizer(0.),
biases_initializer=None):
output = slim.separable_conv2d(input, num_outputs=None, kernel_size=kernel_size, stride=stride,
padding='SAME',
normalizer_fn=None,
normalizer_params=None,
activation_fn=None,
weights_initializer=tf.contrib.layers.xavier_initializer(),
weights_regularizer=slim.l2_regularizer(0.),
biases_initializer=None)
output = slim.conv2d(output, num_outputs=out, kernel_size=1)
return output
def bbox_iou(pb, gt):
tl = np.maximum(pb[0,:], gt[0,:])
br = np.minimum(pb[1,:], gt[1,:])
wl = np.minimum(pb[0,:], gt[0,:])
wr = np.maximum(pb[1,:], gt[1,:])
insect = br-tl
insect = np.clip(insect,0,99999)
pw = (pb[1,0]-pb[0,0])
ph = (pb[1,1]-pb[0,1])
gw = (gt[1,0]-gt[0,0])
gh =(gt[1,1]-gt[0,1])
area_i = insect[0]*insect[1]
area_a = pw*ph
area_b = gw*gh
outRect = wr-wl
outRect = outRect[0]*outRect[1]
iou = area_i / (area_a + area_b - area_i)
# outRect = iou - (outRect-(area_a + area_b - area_i))/outRect
return iou
# boxes = np.asarray([[[0.3125-0.1484375,0.3125-0.1484375,0.3125+0.1484375,0.3125+0.1484375]]])
# boxes = np.zeros([10,1,4])
# landmark = np.asarray([[[[0.3125,0.3125],[0.3225,0.3025],[0.3135,0.3115],[0.3125,0.3125],[0.3125,0.3125],[0.3125,0.3125]]]])
# gt_landmarks = tf.placeholder(dtype=tf.float32,shape=[1,1,6,2])
# sess2 = tf.Session()
# gloc,glabel,m = sess2.run([gt_loc,gt_labels,mask],feed_dict={gt_boxes:boxes,gt_landmarks:landmark})
# p = sess2.run(prior_boxes)
def warmup(lr,step,warm_steps,epoch_size,warm_ratio,decay_ratio):
if step<warm_steps:
lr = float(step)/warm_steps*warm_ratio+0.00001
return lr
if step%epoch_size ==0:
lr = lr*decay_ratio
return lr
def box2center(box):
box[:,:,2] = box[:,:,2]-box[:,:,0]
box[:,:,3] = box[:,:,3]-box[:,:,1]
box[:,:,1] += 0.5*box[:,:,3]
box[:, :, 0] += 0.5 * box[:,:, 2]
return box
def Gau(d,throld = 0.03):
std = 2
m = 2*std**2
t = np.exp(-(d)/m)/(np.pi*m)
return t
def gener_heatmap(cwh,width,height,input = 128.0):
stride = input/width
heatmap = np.zeros((cwh.shape[0],height,width,5),dtype = 'float32')
for i in range(cwh.shape[0]):
for b in range(cwh.shape[1]):
if np.sum(cwh[i,b])==0:
break
m = int(cwh[i,b,0]//stride)
n = int(cwh[i,b,1]//stride)
bm = int(cwh[i,b,0]%stride)
bn = int(cwh[i,b,1] % stride)
heatmap[i,n,m,1] = cwh[i,b,2]/input
heatmap[i, n,m, 2] = cwh[i,b, 3]/input
heatmap[i,n,m,3] = bm/stride
heatmap[i, n,m, 4] = bn/stride
for j in range(height):
for k in range(width):
heatmap[i,k,j,0] = max(Gau((j-m)**2+(k-n)**2),heatmap[i,k,j,0])#多个heatmap时的argmax
if np.max(heatmap[i,:,:, 0])>0:
heatmap[i,:,:, 0] /= np.max(heatmap[i,:,:, 0])
tmp = heatmap[i,:,:, 0]
tmp[np.where(tmp<0.5)] = 0
heatmap[i, :, :, 0] = tmp
# for i in range(heatmap.shape[0]):
# imh = heatmap[i, :, :, 0] * 255
# imh = imh.astype('uint8')
# cv2.imwrite('./test/'+str(i)+'.png',imh)
# cv2.imshow('imh', imh)
# cv2.waitKey(0)
heat_mask = heatmap.copy()
heat_mask[np.where(heat_mask!=0)] = 1
return heatmap,heat_mask
if __name__ == '__main__':
dira = "/root/src/People_Detect/tfdata_big/"
fls = os.listdir(dira)
d = []
for i in fls:
d.append(dira + i)
bgrpath = "/root/src/Face_Detect/dataset/bg/"
bglist = os.listdir(bgrpath)
bglist = [bgrpath + i for i in bglist]
batch_size = 32
epoch_size = int(len(fls)/batch_size)
dataset = tf.data.TFRecordDataset(d)
dataset = dataset.map(_parse_function)
dataset = dataset.shuffle(batch_size*10).repeat().prefetch(batch_size*10).batch(batch_size)
iterator = dataset.make_one_shot_iterator()
imbatch,bboxes = iterator.get_next()
train = True
heatmap_ = tf.placeholder(dtype=tf.float32, shape=[None, 16,16, 5])
heatmask_ = tf.placeholder(dtype=tf.float32, shape=[None, 16,16, 5])
model = BlazeFace(True)
keep_prob = tf.placeholder(tf.float32, name='keep_prob')
HeatL = _Loss()
g = tf.get_default_graph()
if train:
loss = HeatL.loc(heatmap_,model.loc16,heatmask_)
#tf.contrib.quantize.create_training_graph(input_graph=g, quant_delay=2000)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_step = tf.train.AdamOptimizer(keep_prob).minimize(loss)
else:
tf.contrib.quantize.create_eval_graph(input_graph=g)
config = tf.ConfigProto(allow_soft_placement=True)
config.gpu_options.per_process_gpu_memory_fraction = 0.1
sess = tf.Session(config=config)
init_op = tf.global_variables_initializer()
sess.run(init_op)
lr = 0.001
dec = 0.95
once = 1
vlist = []
for l in tf.trainable_variables():
vlist.append(l)
saver = tf.compat.v1.train.Saver(max_to_keep=10,var_list = vlist)
saver1 = tf.compat.v1.train.Saver()
#saver.restore(sess, "./checkpoint/model/Fnetquan_30000_mioutest4.3910613_mioutrain0.029082093.ckpt")
####################数据集测试
for step in range(50001):
lr = warmup(lr,step,3000,epoch_size,0.01,dec)
image, boxes = sess.run([imbatch, bboxes])
# image = (image*255)
# image = image.astype('uint8')
##数据处理
# print(boxes)
for j in range(image.shape[0]):##数据增强前就已经是图像已经除以255 范围[0,1],box是在256尺度下
# #print(np.max(boxes))
image[j],boxes[j] = _crop(image[j], boxes[j], bglist[np.random.randint(0, len(bglist))])
# #print(np.max(boxes))
# if np.random.randint(2)==0:
# image[j],boxes[j] =_fliph(image[j], boxes[j])
# #
#
# if np.random.randint(2)==0:
# image[j], boxes[j] = _expand(image[j], boxes[j], bglist[np.random.randint(0, len(bglist))])
image[j],boxes[j] = tcs(image[j],boxes[j],bglist[np.random.randint(0, len(bglist))])
#
# if np.random.randint(2)==0:
# image[j] = _distort(image[j])
img_input = np.zeros((image.shape[0],128,128,3),dtype='float32')
for j in range(image.shape[0]):
img_input[j,:,:,:] = cv2.resize(image[j,:,:,:],(128,128))
boxes[j] = boxes[j]*(128.0/image.shape[1])
cwh = box2center(boxes)
heatmap_gt ,heat_mask = gener_heatmap(cwh,16,16)
sess.run(train_step,feed_dict={model.input:img_input,heatmap_:heatmap_gt,heatmask_:heat_mask,keep_prob:lr})
if step%500==0:
pre_heat,l = sess.run([model.loc16,loss],feed_dict={model.input:img_input,heatmap_:heatmap_gt,heatmask_:heat_mask})#gloc[n,1344,4] label [n,1344,1]
for p in range(pre_heat.shape[0]):
m,n = np.where(heatmap_gt[p,:,:,0]==np.max(heatmap_gt[p,:,:,0]))
mp, np = np.where(pre_heat[p, :, :, 0] == np.max(pre_heat[p, :, :, 0]))
print("gt max is %d %d ,predict is %d %d",m,n,mp,np)
print("w,h is %f %f pre is %f %f",heatmap_gt[p,m,n,1],heatmap_gt[p,m,n,2],pre_heat[p,m,n,1],pre_heat[p,m,n,2])
print("bias is %f %f pre is %f %f", heatmap_gt[p, m, n, 3], heatmap_gt[p, m, n, 4], pre_heat[p, m, n,3],
pre_heat[p, m, n, 4])
#print(label.shape,label2.shape)
print("%d step lr %f loc_loss: %f "%(step,lr,l))
if step%1000==0:
saver1.save(sess, "./model_netbig2/Fnet_" + str(step).zfill(5) + "_heatmaploss" + str(l) +".ckpt")