-
Notifications
You must be signed in to change notification settings - Fork 42
/
Copy pathmain_whole.py
344 lines (295 loc) · 18.2 KB
/
main_whole.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
#tensorflow 1.2.0 is needed
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os,time,cv2,scipy.io
import tensorflow as tf
import tensorflow.contrib.slim as slim
import numpy as np
import utils as utils
import myflowlib as flowlib
import flow_warp as flow_warp_op
import scipy.misc as sic
import subprocess
import network as net
import loss as loss
import argparse
from sklearn.neighbors import NearestNeighbors
parser = argparse.ArgumentParser()
parser.add_argument("--model", default='Result_whole', type=str, help="Model Name")
parser.add_argument("--div_num", default=4, type=int, help="diverse num")
parser.add_argument("--save_freq", default=1, type=int, help="save frequency")
parser.add_argument("--test_dir", default='./data/test/JPEGImages/480p/cows', type=str, help="Test dir path")
parser.add_argument("--train_root", default="./data/train/JPEGImages/480p/", type=str, help="Test dir path")
parser.add_argument("--test_root", default="./data/test/JPEGImages/480p/", type=str, help="Test dir path")
parser.add_argument("--imgs_dir", default='../data/Imagenet', type=str, help="Test dir path")
parser.add_argument("--is_training", default=1, type=int, help="Training or test")
parser.add_argument("--continue_training", default=1, type=int, help="Restore checkpoint")
ARGS = parser.parse_args()
print(ARGS)
model=ARGS.model
div_num=ARGS.div_num
save_freq = ARGS.save_freq
test_dir = ARGS.test_dir
train_root = [ARGS.train_root]
test_root = [ARGS.test_root]
is_training=ARGS.is_training
continue_training=ARGS.continue_training
imgs_dir = ARGS.imgs_dir
num_frame = 2 # number of read in frames
os.environ["CUDA_VISIBLE_DEVICES"]=str(np.argmax( [int(x.split()[2]) for x in subprocess.Popen("nvidia-smi -q -d Memory | grep -A4 GPU | grep Free", shell=True, stdout=subprocess.PIPE).stdout.readlines()]))
def occlusion_mask(im0, im1, flow10):
warp_im0 = flow_warp_op.flow_warp(im0, flow10)
diff = tf.abs(im1 - warp_im0)
mask = tf.reduce_max(diff, axis=3, keep_dims=True)
mask = tf.less(mask, 0.05)
mask = tf.cast(mask, tf.float32)
mask = tf.tile(mask, [1,1,1,3])
return mask, warp_im0
def Bilateral_NN(color_image, neigh):
h,w = color_image.shape[:2]
color_image_X = np.tile(np.arange(w),[h,1])/500.
color_image_Y = np.tile(np.arange(h),[w,1]).T/500.
color_image_all = np.concatenate([color_image,color_image_X[:,:,np.newaxis], color_image_Y[:,:,np.newaxis]],axis=2)
neigh.fit(np.reshape(color_image_all, [-1, 5]))
idxs = neigh.kneighbors(np.reshape(color_image_all, [-1, 5]), 5, return_distance=False)
# idxs = np.zeros([h*w, 5],dtype="int32")
return idxs
def prepare_input_w_flow(path, num_frames,gray=False):
file_id=os.path.splitext(os.path.basename(path))[0]
input_image_src, input_image_target = utils.read_image_sequence(path, num_frames=num_frame)
if input_image_target is None:
return None, None, None, None
if not gray:
input_flow_forward,input_flow_backward = utils.read_flow_sequence(path.replace("JPEGImages","FLOWImages"), num_frames=num_frame)
else:
input_flow_forward,input_flow_backward = utils.read_flow_sequence(path.replace("JPEGImages","FLOWImages_GRAY"), num_frames=num_frame)
h=input_image_src.shape[0]//32*32
w=input_image_src.shape[1]//32*32
return np.float32(np.expand_dims(input_image_src[:h:2,:w:2,:],axis=0)),\
np.float32(np.expand_dims(input_image_target[:h:2,:w:2,:],axis=0)),\
np.expand_dims(input_flow_forward[:h:2,:w:2,:],axis=0)/2.0,\
np.expand_dims(input_flow_backward[:h:2,:w:2,:],axis=0)/2.0
config=tf.ConfigProto()
config.gpu_options.allow_growth=True
sess=tf.Session(config=config)
train_low=utils.read_image_path(train_root)
test_low=utils.read_image_path(test_root)
input_idx=tf.placeholder(tf.int32,shape=[None,5*num_frame])
input_i=tf.placeholder(tf.float32,shape=[None,None,None,1*num_frame])
input_target=tf.placeholder(tf.float32,shape=[None,None,None,3*num_frame])
input_flow_forward=tf.placeholder(tf.float32,shape=[None,None,None,2*(num_frame-1)])
input_flow_backward=tf.placeholder(tf.float32,shape=[None,None,None,2*(num_frame-1)])
gray_flow_forward=tf.placeholder(tf.float32,shape=[None,None,None,2*(num_frame-1)])
gray_flow_backward=tf.placeholder(tf.float32,shape=[None,None,None,2*(num_frame-1)])
c0=tf.placeholder(tf.float32,shape=[None,None,None,3])
c1=tf.placeholder(tf.float32,shape=[None,None,None,3])
lossDict = {}
objDict={}
# X0, X1: Gray frames
# Y0, Y1: Ground truth color frames
# C0, C1: Colorized frames
with tf.variable_scope(tf.get_variable_scope()):
X0, X1 = input_i[:,:,:,0:1], input_i[:,:,:,1:2]
Y0, Y1 = input_target[:,:,:,0:3], input_target[:,:,:,3:6]
with tf.variable_scope('individual'):
C0=net.VCN(utils.build(tf.tile(X0, [1,1,1,3])),reuse=False)
C1=net.VCN(utils.build(tf.tile(X1, [1,1,1,3])),reuse=True)
objDict["mask"],_=occlusion_mask(Y0,Y1,input_flow_backward[:,:,:,0:2])
objDict["warped"]=flow_warp_op.flow_warp(C0,input_flow_backward[:,:,:,0:2])
lossDict["RankDiv_im1"]=loss.RankDiverse_loss(C0, tf.tile(input_target[:,:,:,0:3], [1,1,1,div_num]),div_num)
lossDict["RankDiv_im2"]=loss.RankDiverse_loss(C1, tf.tile(input_target[:,:,:,3:6], [1,1,1,div_num]),div_num)
lossDict["RankDiv"]=lossDict["RankDiv_im1"]+lossDict["RankDiv_im2"]
lossDict['Bilateral_im1']= sum([loss.KNN_loss(C0[:,:,:,3*i:3*i+3], input_idx[:,0:5]) for i in range(4)])
lossDict['Bilateral_im2']= sum([loss.KNN_loss(C1[:,:,:,3*i:3*i+3], input_idx[:,5:10])for i in range(4)])
lossDict['Bilateral']= lossDict['Bilateral_im2'] + lossDict['Bilateral_im1']
lossDict["temporal"]=tf.reduce_mean(tf.multiply(tf.abs(objDict["warped"]-C1),tf.tile(objDict["mask"],[1,1,1,4])))*5
lossDict["total"]=lossDict["RankDiv"]+lossDict["temporal"]#+lossDict['Bilateral']
objDict["prediction_0"]=tf.concat([C0[:,:,:,0:3],C0[:,:,:,9:12],C0[:,:,:,3:6],C0[:,:,:,6:9]],axis=2)
objDict["prediction_1"]=tf.concat([C1[:,:,:,0:3],C1[:,:,:,9:12],C1[:,:,:,3:6],C1[:,:,:,6:9]],axis=2)
#-------------RefineNet---------------#
cmap_C, warp_C0 = occlusion_mask(c0, c1, gray_flow_backward[:,:,:,0:2])
cmap_X, warp_X0 = occlusion_mask(tf.tile(input_i[:,:,:,0:1], [1,1,1,3]), tf.tile(input_i[:,:,:,1:2],[1,1,1,3]),gray_flow_backward[:,:,:,0:2])
low_conf_mask = tf.cast(tf.greater(cmap_X - cmap_C, 0), tf.float32)
coarse_C1 = c1*(-low_conf_mask+1) + tf.tile(input_i[:,:,:,1:2],[1,1,1,3])*low_conf_mask
ref_input = tf.concat([coarse_C1, warp_C0, c1, low_conf_mask, cmap_C, cmap_X], axis=3)
# ref_input = tf.concat([warp_C0, c1, cmap_C, cmap_X, low_conf_mask], axis=3)
final_r1 = net.VCRN(ref_input)
# temporal_g_loss = tf.reduce_mean(tf.abs(input_target[:,:,:,3:6]-final_r1)) + tf.reduce_mean(tf.multiply(tf.abs(c1-final_r1),-low_conf_mask+1))
temporal_g_loss = tf.reduce_mean(tf.abs(input_target[:,:,:,3:6]-final_r1)) + \
tf.reduce_mean(tf.multiply(tf.abs(warp_C0-final_r1),low_conf_mask)) + \
tf.reduce_mean(tf.multiply(tf.abs(c1-final_r1),-low_conf_mask+1))
opt=tf.train.AdamOptimizer(learning_rate=0.0001).minimize(lossDict["total"],var_list=[var for var in tf.trainable_variables()])
opt2=tf.train.AdamOptimizer(learning_rate=0.0001).minimize(0.2*lossDict["RankDiv_im1"],var_list=[var for var in tf.trainable_variables()])
opt_refine=tf.train.AdamOptimizer(learning_rate=0.0001).minimize(temporal_g_loss,var_list=[var for var in tf.trainable_variables() if var.name.startswith('VCRN')])
print([var for var in tf.trainable_variables() if var.name.startswith('VCRN')])
saver=tf.train.Saver(max_to_keep=1000)
sess.run([tf.global_variables_initializer()])
var_restore = [v for v in tf.trainable_variables()]
saver_restore=tf.train.Saver(var_restore)
ckpt=tf.train.get_checkpoint_state(model)
print("contain checkpoint: ", ckpt)
if ckpt and continue_training:
print('loaded '+ ckpt.model_checkpoint_path)
saver_restore.restore(sess,ckpt.model_checkpoint_path)
neigh=NearestNeighbors(n_neighbors=5)
maxepoch=1001
num_train=len(train_low)
print("Number of training images: ", num_train)
print("is_training: ", is_training )
if is_training:
for epoch in range(1,maxepoch):
print("Processing epoch %d"%epoch)
input_list_src=[None]*num_train
input_list_target=[None]*num_train
input_list_flow_forward=[None]*num_train
input_list_flow_backward=[None]*num_train
gray_list_flow_forward=[None]*num_train
gray_list_flow_backward=[None]*num_train
if os.path.isdir("%s/%04d"%(model,epoch)):
continue
cnt=0
all_RD,all_Bi = 0., 0.
#Images
for id in np.random.permutation(1431167):#1431167
st=time.time()
color_image=np.float32(scipy.misc.imread("%s/%06d.jpg"%(imgs_dir, id+1)))/255.0
if len(color_image.shape)==2:
continue
h=color_image.shape[0]//32*32
w=color_image.shape[1]//32*32
color_image=color_image[:h:2,:w:2,:]
gray_image=cv2.cvtColor(color_image,cv2.COLOR_RGB2GRAY)
if gray_image is None:
print(id)
continue
idxs = Bilateral_NN(color_image, neigh)
gray_image=gray_image[:,:,np.newaxis]
_, crt_RDLoss, crt_BiLoss = sess.run([opt2,lossDict["RankDiv_im1"], lossDict["Bilateral_im1"]],feed_dict={input_i:gray_image[np.newaxis,:,:,[0,0]],\
input_target:color_image[np.newaxis,:,:,[0,1,2,0,1,2]], input_idx:np.tile(idxs,[1,2])})
cnt+=1
all_RD += crt_RDLoss
all_Bi += crt_BiLoss
print("Image iter: %d %d || RankDiv: %.4f %.4f|| Bi: %.4f %.4f || Time: %.4f"%(epoch,cnt,crt_RDLoss,all_RD/cnt,crt_BiLoss,all_Bi/cnt,time.time()-st))
if cnt>=5000:
break
# Video VCN
cnt=0
all_D1, all_D2, all_B1, all_B2, all_T, all_loss = 0,0,0,0,0,0
for id in np.random.permutation(num_train):
st=time.time()
if input_list_src[id] is None:
input_list_src[id], input_list_target[id], input_list_flow_forward[id], input_list_flow_backward[id] = prepare_input_w_flow(train_low[id], num_frames=num_frame)
if input_list_src[id] is None or input_list_target[id] is None or input_list_flow_forward[id] is None:
continue
input_frames_processed = input_list_src[id]
idxs1 = Bilateral_NN(input_list_target[id][0,:,:,:3], neigh)
idxs2 = Bilateral_NN(input_list_target[id][0,:,:,3:6], neigh)
_, out_loss, C0_im, C1_im = sess.run([opt, lossDict,C0,C1],\
feed_dict={input_i:input_list_src[id],input_target:input_list_target[id],\
input_flow_backward:input_list_flow_backward[id],\
input_idx: np.concatenate([idxs1,idxs2],axis=1)})
all_D1 += out_loss["RankDiv_im1"]
all_D2 += out_loss["RankDiv_im2"]
all_B1 += out_loss["Bilateral_im1"]
all_B2 += out_loss["Bilateral_im2"]
all_T += out_loss["temporal"]
all_loss += out_loss["total"]
cnt+=1
print("iter: %d %d %.2fs loss: %.4f %.4f|| (D1) %.4f %.4f (D2) %.4f %.4f || (B1) %.4f %.4f (B2) %.4f %.4f (T) %.4f %.4f"\
%(epoch,cnt,out_loss["total"],all_loss/cnt, time.time()-st,\
out_loss["RankDiv_im1"], all_D1/cnt, out_loss["RankDiv_im2"], all_D2/cnt,\
out_loss["Bilateral_im1"], all_B1/cnt, out_loss["Bilateral_im2"], all_B2/cnt,\
out_loss["temporal"], all_T/cnt))
# Video Refine
if epoch > 0:
_, _, gray_list_flow_forward[id], gray_list_flow_backward[id] = prepare_input_w_flow(train_low[id], num_frames=num_frame)
_, out_loss, final_C1 = sess.run([opt_refine, temporal_g_loss, final_r1],\
feed_dict={c0:C0_im[:,:,:,0:3], c1: C1_im[:,:,:,0:3], input_i:input_list_src[id],input_target:input_list_target[id],\
input_flow_backward:input_list_flow_backward[id],\
gray_flow_backward:gray_list_flow_backward[id],\
input_idx: np.concatenate([idxs1,idxs2],axis=1)})
print("iter: %d %d || Refine || loss: %.4f %.4f"%(epoch,cnt,out_loss,time.time()-st))
if cnt>=1000:
break
# Validation
if not os.path.isdir("%s/%04d"%(model,epoch)):
os.makedirs("%s/%04d"%(model,epoch))
if epoch % save_freq == 0:
numtest=len(test_low)
all_loss_test=np.zeros(numtest, dtype=float)
for ind in range(numtest):
if ind>30 and epoch%25>0:
break
input_image_src, input_image_target, input_flow_forward_src, input_flow_backward_src = prepare_input_w_flow(test_low[int(ind*60/pow(60,int(epoch%25==0)))],num_frames=num_frame)
if input_image_src is None or input_image_target is None or input_flow_forward_src is None:
print("Not able to read the images/flows.")
flag=True
continue
st=time.time()
C0_imall,C1_imall,C0_im, C1_im, warped,mask=sess.run([objDict["prediction_0"],objDict["prediction_1"],C0, C1,objDict["warped"],objDict['mask']],feed_dict={input_i:input_image_src,
input_target:input_image_target,
input_flow_backward:input_flow_backward_src
})
print("test time for %s --> %.3f"%(ind, time.time()-st))
input_image_src, input_image_target, gray_flow_forward_src, gray_flow_backward_src = prepare_input_w_flow(test_low[int(ind*60/pow(60,int(epoch%25==0)))],num_frames=num_frame)
h,w = C0_im.shape[1:3]
outputs= []
for ref_i in range(4):
output, out_cmap_C, out_cmap_X, out_low_conf_mask = sess.run([final_r1, cmap_C,cmap_X,low_conf_mask],feed_dict={c0:C0_im[:,:,:,ref_i*3:ref_i*3+3], c1:C1_im[:,:,:,ref_i*3:ref_i*3+3], \
input_i:input_image_src, input_target:input_image_target, \
gray_flow_backward:gray_flow_backward_src, input_flow_backward:input_flow_backward_src})
outputs.append(output[0,:,:,:])
# Debug
if not os.path.isdir("%s/%04d/predictions" % (model, epoch)):
os.makedirs("%s/%04d/predictions" % (model, epoch))
sic.imsave("%s/%04d/predictions/mask_%06d.jpg"%(model, epoch, ind),np.uint8(np.maximum(np.minimum(np.concatenate([out_cmap_C[0],out_cmap_X[0],out_low_conf_mask[0]],axis=1)* 255.0,255.0),0.0)))
sic.imsave("%s/%04d/predictions/final_%06d.jpg"%(model, epoch, ind),np.uint8(np.maximum(np.minimum(output[0] * 255.0,255.0),0.0)))
sic.imsave("%s/%04d/predictions/predictions_%06d.jpg"%(model, epoch, ind),np.uint8(np.maximum(np.minimum(C0_imall[0] * 255.0,255.0),0.0)))
saver.save(sess,"%s/model.ckpt"%model)
if epoch%10==0:
saver.save(sess,"%s/%04d/model.ckpt"%(model,epoch))
# Inference
else:
test_low=utils.get_names(test_dir)
numtest=len(test_low)
print(test_low[0])
out_folder = test_dir.split('/')[-1]
outputs= [None]*4
for ind in range(numtest):
input_image_src, input_image_target, input_flow_forward_src, input_flow_backward_src = prepare_input_w_flow(test_low[ind],num_frames=num_frame,gray=True)
if input_image_src is None or input_image_target is None or input_flow_forward_src is None:
print("Not able to read the images/flows.")
continue
st=time.time()
C0_imall,C1_imall,C0_im, C1_im=sess.run([objDict["prediction_0"],objDict["prediction_1"],C0, C1],feed_dict={input_i:input_image_src,
input_target:input_image_target,
input_flow_backward:input_flow_backward_src
})
print("test time for %s --> %.3f"%(ind, time.time()-st))
h,w = C0_im.shape[1:3]
print(C0_im.shape)
if not os.path.isdir("%s/%s" % (model, out_folder)):
os.makedirs("%s/%s/predictions" % (model, out_folder))
os.makedirs("%s/%s/predictions0" % (model, out_folder))
os.makedirs("%s/%s/predictions1" % (model, out_folder))
os.makedirs("%s/%s/predictions2" % (model, out_folder))
os.makedirs("%s/%s/predictions3" % (model, out_folder))
if ind == 0:
for ref_i in range(4):
output,_ = sess.run([final_r1, temporal_g_loss],feed_dict={c0:C0_im[:,:,:,ref_i*3:ref_i*3+3], c1:C1_im[:,:,:,ref_i*3:ref_i*3+3], \
input_i:input_image_src, input_target:input_image_target, \
gray_flow_backward:input_flow_backward_src, input_flow_backward:input_flow_backward_src})
outputs[ref_i] = output
sic.imsave("%s/%s/predictions%d/final_%06d.jpg"%(model, out_folder, ref_i, ind),np.uint8(np.maximum(np.minimum(C0_im[0,:,:,ref_i*3:ref_i*3+3] * 255.0,255.0),0.0)))
sic.imsave("%s/%s/predictions%d/final_%06d.jpg"%(model, out_folder, ref_i, ind+1),np.uint8(np.maximum(np.minimum(output[0,:,:,:] * 255.0,255.0),0.0)))
sic.imsave("%s/%s/predictions/predictions_%06d.jpg"%(model, out_folder, ind+1),np.uint8(np.maximum(np.minimum(C1_imall[0,:,:,:] * 255.0,255.0),0.0)))
sic.imsave("%s/%s/predictions/final_%06d.jpg"%(model, out_folder, ind+1),np.uint8(np.maximum(np.minimum(np.concatenate(outputs,axis=2)[0,:,:,:] * 255.0,255.0),0.0)))
else:
for ref_i in range(4):
output,_ = sess.run([final_r1, temporal_g_loss],feed_dict={c0:outputs[ref_i], c1:C1_im[:,:,:,:3], \
input_i:input_image_src, input_target:input_image_target, \
gray_flow_backward:input_flow_backward_src, input_flow_backward:input_flow_backward_src})
outputs.append(output[0,:,:,:])
sic.imsave("%s/%s/predictions%d/final_%06d.jpg"%(model, out_folder, ref_i, ind+1),np.uint8(np.maximum(np.minimum(output[0,:,:,:] * 255.0,255.0),0.0)))
sic.imsave("%s/%s/predictions/predictions_%06d.jpg"%(model, out_folder, ind+1),np.uint8(np.maximum(np.minimum(C1_imall[0,:,:,:] * 255.0,255.0),0.0)))