/
RenderNet_Texture_Face_Normal.py
335 lines (293 loc) · 21.5 KB
/
RenderNet_Texture_Face_Normal.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
import numpy as np
import os
import tensorflow as tf
import sys
import time
import scipy.ndimage
import json
import random
import shutil
from tools.model_util import tf_transform_voxel_to_match_image, tf_random_crop_voxel_texture_image_normal
from tools.data_util import data_loader_image_texture_normal_face
from tools.layer_util import conv3d_transpose, conv3d, prelu, fully_connected, conv2d, conv2d_transpose, res_block_3d, res_block_2d, keep_prob, projection_unit
from tools.resampling_voxel_grid import tf_rotation_resampling
#=======================================================================================================================
with open(sys.argv[1], 'r') as fh:
cfg = json.load(fh)
IMAGE_PATH = cfg['image_path']
IMAGE_PATH_VALID = cfg['image_path_valid']
NORMAL_PATH = cfg['normal_path']
TEXTURE_PATH = cfg['texture_path']
MODEL_PATH = cfg['model_path']
SAMPLE_SAVE = cfg['sample_save']
MODEL_SAVE = os.path.join(SAMPLE_SAVE, cfg['trained_model_name'])
LOGDIR = SAMPLE_SAVE + "/log"
os.environ["CUDA_VISIBLE_DEVICES"] = "{0}".format(cfg['gpu'])
#=======================================================================================================================
def decoder_texture(z_in):
with tf.variable_scope("texture_encoder"):
batch_size = tf.shape(z_in)[0]
with tf.variable_scope('e_tex_fc1'):
zP = prelu((fully_connected(z_in, 32 * 32 * 32 * 4)))
z_resized = tf.reshape(zP, [batch_size, 32, 32, 32, 4])
with tf.variable_scope('e_tex_conv0'):
conv0 = prelu(conv3d_transpose(z_resized, 4, kernel_size=[4, 4, 4], stride=[1, 1, 1]))
with tf.variable_scope('e_tex_conv1'):
conv1 = prelu(conv3d_transpose(conv0, 8, kernel_size=[4, 4, 4], stride=[2, 2, 2]))
with tf.variable_scope('e_tex_conv2'):
conv2 = prelu(conv3d(conv1, 4, kernel_size=[4, 4, 4], stride=[1, 1, 1]))
return conv2
def RenderNet(models_in, prob = 0.75, reuse=False):
batch_size = tf.shape(models_in)[0]
with tf.variable_scope("encoder"):
with tf.variable_scope('e_conv1'):
enc1 = prelu(conv3d(models_in, 8, kernel_size=[5, 5, 5], stride=[2, 2, 2],
reuse=reuse, pad="SAME", scope='e_conv1',
weight_initializer_type=tf.contrib.layers.xavier_initializer()))
enc1 = tf.nn.dropout(enc1, keep_prob(prob, is_training))
with tf.variable_scope('e_conv2'):
enc2 = prelu(conv3d(enc1, 16, kernel_size=[3, 3, 3], stride=[1, 1, 2],
reuse=reuse, pad="SAME", scope='e_conv2',
weight_initializer_type=tf.contrib.layers.xavier_initializer()))
enc2 = tf.nn.dropout(enc2, keep_prob(prob, is_training))
with tf.variable_scope('e_conv3'):
enc3 = prelu(conv3d(enc2, 16, kernel_size=[3, 3, 3], stride=[1, 1, 1],
reuse=reuse, pad="SAME", scope='e_conv3',
weight_initializer_type=tf.contrib.layers.xavier_initializer()))
enc3 = tf.nn.dropout(enc3, keep_prob(prob, is_training))
shortcut = enc3
res1_1 = res_block_3d(enc3, 16, scope='res1_1')
res1_2 = res_block_3d(res1_1, 16, scope='res1_2')
res1_3 = res_block_3d(res1_2, 16, scope='res1_3')
res1_4 = res_block_3d(res1_3, 16, scope='res1_4')
res1_5 = res_block_3d(res1_4, 16, scope='res1_5')
res1_6 = res_block_3d(res1_5, 16, scope='res1_6')
res1_7 = res_block_3d(res1_6, 16, scope='res1_7')
res1_8 = res_block_3d(res1_7, 16, scope='res1_8')
res1_9 = res_block_3d(res1_8, 16, scope='res1_9')
res1_10 = res_block_3d(res1_9, 16, scope='res1_10')
with tf.variable_scope('res1_skip'):
enc3_skip = conv3d(res1_10, 16, kernel_size=[3, 3, 3], stride=[1, 1, 1], pad="SAME", scope="con1_3X3", weight_initializer_type=tf.contrib.layers.xavier_initializer())
enc3_skip = tf.add(tf.cast(enc3_skip, tf.float32), tf.cast(shortcut, tf.float32))
enc4 = projection_unit(enc3_skip)
shortcut = enc4
res2_1 = res_block_2d(enc4, 32 * 16, scope='res2_1')
res2_2 = res_block_2d(res2_1, 32 * 16, scope='res2_2')
res2_3 = res_block_2d(res2_2, 32 * 16, scope='res2_3')
res2_4 = res_block_2d(res2_3, 32 * 16, scope='res2_4')
res2_5 = res_block_2d(res2_4, 32 * 16, scope='res2_5')
res2_6 = res_block_2d(res2_5, 32 * 16, scope='res2_6')
res2_7 = res_block_2d(res2_6, 32 * 16, scope='res2_7')
res2_8 = res_block_2d(res2_7, 32 * 16, scope='res2_8')
res2_9 = res_block_2d(res2_8, 32 * 16, scope='res2_9')
res2_10 = res_block_2d(res2_9, 32 * 16, scope='res2_10')
with tf.variable_scope('res2_skip'):
enc4_skip = conv2d(res2_10, 32 * 16, kernel_size=[3, 3], stride=[1, 1], scope="con1_3X3", weight_initializer_type=tf.contrib.layers.xavier_initializer())
enc4_skip = tf.add(tf.cast(enc4_skip, tf.float32), tf.cast(shortcut, tf.float32))
with tf.variable_scope('e_conv5'):
enc5 = prelu(conv2d(enc4_skip, 32 * 8, kernel_size=[4,4], stride = [1, 1], scope='e_conv5', weight_initializer_type=tf.contrib.layers.xavier_initializer()))
enc5 = tf.nn.dropout(enc5, keep_prob(prob, is_training))
shortcut = enc5
res3_1 = res_block_2d(enc5, 32 * 8, scope='res3_1')
res3_2 = res_block_2d(res3_1, 32 * 8, scope='res3_2')
res3_3 = res_block_2d(res3_2, 32 * 8, scope='res3_3')
res3_4 = res_block_2d(res3_3, 32 * 8, scope='res3_4')
res3_5 = res_block_2d(res3_4, 32 * 8, scope='res3_5')
with tf.variable_scope('res3_skip'):
enc5_skip = conv2d(res3_5, 32 * 8, kernel_size=[3, 3], stride=[1, 1], scope="con1_3X3", weight_initializer_type=tf.contrib.layers.xavier_initializer())
enc5_skip = tf.add(tf.cast(enc5_skip, tf.float32), tf.cast(shortcut, tf.float32))
with tf.variable_scope("Image"):
with tf.variable_scope('e_conv6_1'):
enc6_1 = prelu(conv2d(enc5_skip, 32 * 4, kernel_size=[4, 4], stride=[1, 1], scope='e_conv6_1', weight_initializer_type=tf.contrib.layers.xavier_initializer()))
enc6_1 = tf.nn.dropout(enc6_1, keep_prob(prob, is_training))
with tf.variable_scope('e_conv7_1'):
enc7_1 = prelu(conv2d_transpose(enc6_1, 32 * 2, [4, 4], stride=[2, 2], scope='e_conv7_2', weight_initializer_type=tf.contrib.layers.xavier_initializer()))
enc7_1 = tf.nn.dropout(enc7_1, keep_prob(prob, is_training))
with tf.variable_scope('e_conv8_1'):
enc8_1 = prelu(conv2d_transpose(enc7_1, 32, [4, 4], stride=[2, 2], weight_initializer_type=tf.contrib.layers.xavier_initializer()))
enc8_1 = tf.nn.dropout(enc8_1, keep_prob(prob, is_training))
with tf.variable_scope('e_conv9_1'):
enc9_1 = prelu(conv2d_transpose(enc8_1, 16, [4, 4], stride=[2, 2], weight_initializer_type=tf.contrib.layers.xavier_initializer()))
enc9_1 = tf.nn.dropout(enc9_1, keep_prob(prob, is_training))
with tf.variable_scope('e_conv10_1'):
enc10_1 = conv2d_transpose(enc9_1, 3, [4, 4], stride=[1, 1], weight_initializer_type=tf.contrib.layers.xavier_initializer())
enc10_1 = tf.nn.sigmoid(enc10_1, name="encoder_output")
with tf.variable_scope("Normal"):
with tf.variable_scope('e_conv6_2'):
enc6_2 = prelu(conv2d(enc5_skip, 32 * 4, kernel_size=[4, 4], stride=[1,1], scope='e_conv6_2', weight_initializer_type=tf.contrib.layers.xavier_initializer()))
enc6_2 = tf.nn.dropout(enc6_2, keep_prob(prob, is_training))
with tf.variable_scope('e_conv7_2'):
enc7_2 = prelu(conv2d_transpose(enc6_2, 32 * 2, [4, 4], stride=[2, 2], scope='e_conv7_2', weight_initializer_type=tf.contrib.layers.xavier_initializer()))
enc7_2 = tf.nn.dropout(enc7_2, keep_prob(prob, is_training))
with tf.variable_scope('e_conv8_2'):
enc8_2 = prelu(conv2d_transpose(enc7_2, 32, [4, 4], stride=[2, 2], scope='e_conv8_2', weight_initializer_type=tf.contrib.layers.xavier_initializer()))
enc8_2 = tf.nn.dropout(enc8_2, keep_prob(prob, is_training))
with tf.variable_scope('e_conv9_2'):
enc9_2 = prelu(conv2d_transpose(enc8_2, 16, [4, 4], stride=[2, 2], scope='e_conv9_2', weight_initializer_type=tf.contrib.layers.xavier_initializer()))
enc9_2 = tf.nn.dropout(enc9_2, keep_prob(prob, is_training))
with tf.variable_scope('e_conv10_2'):
enc10_2 = conv2d_transpose(enc9_2, 3, [4, 4], stride=[1, 1], scope='e_conv10_2', weight_initializer_type=tf.contrib.layers.xavier_initializer())
enc10_2 = tf.nn.sigmoid(enc10_2, name="encoder_output")
return enc10_1, enc10_2
#============================================================================================================================
#============================================================================================================================
graph=tf.Graph()
new_res = 128
with graph.as_default():
model_in = tf.placeholder(shape=[None, 64, 64, 64, 1], dtype=tf.float32, name = "real_model_in") #Real 3D voxel objects
normal_in = tf.placeholder(shape=[None, 512, 512, 3], dtype=tf.float32, name = "real_normal_in") #2D render of voxel objects
image_in = tf.placeholder(shape=[None, 512, 512, 3], dtype=tf.float32, name = "real_image_in") #2D render of voxel objects
texture_in = tf.placeholder(shape=[None, 199], dtype=tf.float32, name = "real_texture_in") #2D render of voxel objects
param_in = tf.placeholder(shape=[None, 3], dtype=tf.float32, name = "view_name")
is_training = tf.placeholder(tf.bool, [], name="is_training")
patch_size_in = tf.placeholder(tf.int32, [], name="patch_size")
#Rotate and resample 3D input
rotated_models = tf_rotation_resampling(model_in, param_in, new_size = new_res)
rotated_models = tf_transform_voxel_to_match_image(rotated_models) #Transform voxel array to match image array (ijk -> xyz)
#Create 3D representation of the texture of the vector input
texture_decoded = decoder_texture (z_in=texture_in)
#Rotate and resample 3D input
texture_rotated = tf_rotation_resampling(texture_decoded, param_in, new_size = new_res)
texture_rotated = tf_transform_voxel_to_match_image(texture_rotated)
#Crop voxel gris and images for patch training
crop_vox, crop_tex, crop_img, crop_normal = tf_random_crop_voxel_texture_image_normal(rotated_models, texture_rotated, image_in, normal_in, patch_size_in)
#Concateante voxel geometry and texture for rendering
model_texture_concat = tf.concat([crop_vox, crop_tex], 4)
images_pred, normal_pred = RenderNet(model_texture_concat, prob=cfg['keep_prob'])
#Comput MSE loss on the rendered image and rendered normal
recon_loss = tf.reduce_mean(tf.losses.mean_squared_error(crop_img, images_pred)) + \
tf.reduce_mean(tf.losses.mean_squared_error(crop_normal, normal_pred))
global_step = tf.Variable(0, name='global_step', trainable=False)
learning_rate = tf.train.exponential_decay(cfg['e_eta'], global_step, cfg['decay_steps'], 0.96, staircase=True)
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate,beta1=0.5).minimize(recon_loss, global_step=global_step)
tf.summary.scalar('Recon loss train', recon_loss)
merged_summary_op = tf.summary.merge_all()
init = tf.global_variables_initializer()
sess_saver = tf.train.Saver()
summary_writer = tf.summary.FileWriter(LOGDIR, graph=graph)
#======================================================================================================================
#======================================================================================================================
sv = tf.train.Supervisor(logdir=LOGDIR,
graph=graph,
is_chief=True,
saver=sess_saver,
summary_op=None,
summary_writer=summary_writer,
save_summaries_secs = 300,
save_model_secs=cfg['checkpoint_secs'],
global_step=global_step)
sess_config = tf.ConfigProto(allow_soft_placement=True)
L1_ALL = list()
train = True
with sv.managed_session(config=sess_config) as sess:
with tf.device("/gpu:0"):
while not sv.should_stop():
if not os.path.exists(SAMPLE_SAVE):
os.makedirs(SAMPLE_SAVE)
shutil.copyfile(sys.argv[1], os.path.join(SAMPLE_SAVE, 'config.json'))
train_writer = tf.summary.FileWriter(os.path.join(SAMPLE_SAVE, 'train'), graph=graph)
if train:
for epoch in range(cfg['max_epochs']):
print ("Epoch starting")
if epoch < 4:
batch_patch_size = new_res // 4
else:
batch_patch_size = new_res // 2
start_time=time.time()
train_loader = data_loader_image_texture_normal_face(cfg, img_path=IMAGE_PATH,
model_path=MODEL_PATH, normal_path = NORMAL_PATH,
texture_path=TEXTURE_PATH,
validation_mode=False,
img_res=512)
counter=0
print ("Loading data")
for real_images, real_normals, real_models, real_texture, real_params, real_names in train_loader:# Loop across chunks, READER object will grab the next chunk after every time a chunk is completed
print ("Data chunk loaded")
chunk_start = time.time()
bceLoss_chunk = 0.0
# Check input
real_images /= 255.0
real_normals /= 255.0
num_batches = len(real_images) // cfg['batch_size']
print("Start stepping" + str(time.time() - chunk_start))
for idx in range(num_batches):
step_start = time.time()
batch_models = real_models[idx * cfg['batch_size']:(idx + 1) * cfg['batch_size']]
batch_images = real_images[idx * cfg['batch_size']:(idx + 1) * cfg['batch_size']]
batch_normals = real_normals[idx * cfg['batch_size']:(idx + 1) * cfg['batch_size']]
batch_texture = real_texture[idx * cfg['batch_size']:(idx + 1) * cfg['batch_size']]
batch_names = real_names[idx * cfg['batch_size']:(idx + 1) * cfg['batch_size']]
batch_params = real_params[idx * cfg['batch_size']:(idx + 1) * cfg['batch_size']]
feed_train = {model_in: batch_models,
image_in: batch_images,
normal_in: batch_normals,
texture_in: batch_texture,
param_in:batch_params,
patch_size_in: batch_patch_size,
is_training: True}
train, summary, step = sess.run([optimizer, merged_summary_op, global_step], feed_dict = feed_train)
train_writer.add_summary(summary, global_step=step)
print("Time/step" + str(time.time() - step_start))
if step % 600 == 0:
print( "Epoch {2} {0} BCE Loss: {1}".format(step, bceLoss_chunk, epoch))
rendered_samples, rendered_normals, out_cropped_im, out_cropped_normal = sess.run([images_pred, normal_pred, crop_img, crop_normal],
feed_dict = feed_train)
out_cropped_im = np.clip(255 * out_cropped_im, 0, 255).astype(np.uint8)
out_cropped_normal = np.clip(255 * out_cropped_normal, 0, 255).astype(np.uint8)
rendered_samples = np.clip(255 * rendered_samples, 0, 255).astype(np.uint8)
rendered_normals = np.clip(255 * rendered_normals, 0, 255).astype(np.uint8)
for i in range(1):
idx=random.randint(0, cfg['batch_size'] - 1)
scipy.misc.imsave(os.path.join(SAMPLE_SAVE, "{0}_train_target_{1}_patch.png".format(batch_names[idx], step)), out_cropped_im[idx])
scipy.misc.imsave(os.path.join(SAMPLE_SAVE, "{0}_train_target_normal_{1}_patch.png".format(batch_names[idx], step)), out_cropped_normal[idx])
scipy.misc.imsave(os.path.join(SAMPLE_SAVE, "{0}_train_{1}_patch.png".format(batch_names[idx], step)), rendered_samples[idx])
scipy.misc.imsave(os.path.join(SAMPLE_SAVE, "{0}_train_{1}_patch_normal.png".format(batch_names[idx],step)), rendered_normals[idx])
print("Time/chunk " + str(time.time() - chunk_start))
#===================================================================================================
#Validation
#===================================================================================================
save_path = sess_saver.save(sess, MODEL_SAVE)
valid_loader = data_loader_image_texture_normal_face(cfg, img_path=IMAGE_PATH_VALID, normal_path=NORMAL_PATH,
model_path=MODEL_PATH, texture_path=TEXTURE_PATH,
validation_mode=True,
img_res=512)
counter = 0
L1_valid = 0.
for real_images, real_normals, real_models, real_texture, real_params, real_names in valid_loader:
real_images /= 255.0
real_normals /= 255.0
num_batches = len(real_images) // cfg['batch_size']
for idx in range(num_batches):
valid_batch_models = real_models[idx * cfg['batch_size']:(idx + 1) * cfg['batch_size']]
valid_batch_images = real_images[idx * cfg['batch_size']:(idx + 1) * cfg['batch_size']]
valid_batch_normals = real_normals[idx * cfg['batch_size']:(idx + 1) * cfg['batch_size']]
valid_batch_texture = real_texture[idx * cfg['batch_size']:(idx + 1) * cfg['batch_size']]
valid_batch_names = real_names[idx * cfg['batch_size']:(idx + 1) * cfg['batch_size']]
valid_batch_params = real_params[idx * cfg['batch_size']:(idx + 1) * cfg['batch_size']]
feed_valid = {model_in: valid_batch_models,
image_in: valid_batch_images,
normal_in: valid_batch_normals,
texture_in: valid_batch_texture,
param_in: valid_batch_params,
patch_size_in: new_res,
is_training:False}
rendered_samples, rendered_normals = sess.run([images_pred, normal_pred],
feed_dict = feed_valid)
rendered_samples = np.clip(255 * rendered_samples, 0, 255).astype(np.uint8)
rendered_normals = np.clip(255 * rendered_normals, 0, 255).astype(np.uint8)
if counter % 600 == 0:
for i in range(1):
index = random.randint(0, cfg['batch_size'] - 1)
scipy.misc.imsave(os.path.join(SAMPLE_SAVE, "VALID_{0}_target_{1}.png".format(valid_batch_names[index], epoch)), np.squeeze(valid_batch_images[index]))
scipy.misc.imsave(os.path.join(SAMPLE_SAVE, "VALID_{0}_target_normal_{1}.png".format(valid_batch_names[index], epoch)), np.squeeze(valid_batch_normals[index]))
scipy.misc.imsave(os.path.join(SAMPLE_SAVE, "VALID_{0}_pred_{1}.png".format(valid_batch_names[index], epoch)), np.squeeze(rendered_samples[index]))
scipy.misc.imsave(os.path.join(SAMPLE_SAVE, "VALID_{0}_pred_normal_{1}.png".format(valid_batch_names[index], epoch)), np.squeeze(rendered_normals[index]))
validation_accuracy = np.mean(np.absolute(valid_batch_images - rendered_samples))
L1_valid += validation_accuracy
counter +=1
# Across test set (multi mini-batches) (after 1 epoch)
L1_valid = L1_valid / counter
L1_ALL.append(L1_valid)
np.savez(os.path.join(SAMPLE_SAVE, "L1 All.txt"), L1_ALL)
print("Time elapsed: " + str(time.time() - start_time))
sv.request_stop()