-
Notifications
You must be signed in to change notification settings - Fork 1
/
k-styleGAN.py
163 lines (144 loc) · 7.57 KB
/
k-styleGAN.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
import time
from collections import namedtuple
from utils import linear_sample, read_images, decoder
import os
import pickle
import numpy as np
import PIL.Image
import dnnlib
import tensorflow as tf
import dnnlib.tflib as tflib
import config
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
model_params = namedtuple("parameters",
'data_size,'
'mini_batch_size,'
'learning_rate,'
'total_epoch_num,'
'outputdir,'
'height,'
'width,'
'channel,'
'cutTop,'
'cutBottom,'
'cutLeft,'
'cutRight,'
'dataset_name,'
'cPointRow,'
'cPointCol,'
'compensate_disp,')
print_str = 'train_tps_Step：{:4} | Reonstruction loss is {:4} | Total loss is {:4} | z_var_mean is{:4}'
def get_uninitialized_variables(sess):
global_vars = tf.global_variables()
is_not_initialized = sess.run([tf.is_variable_initialized(var) for var in global_vars])
not_initialized_vars = [v for (v, f) in zip(global_vars, is_not_initialized) if not f]
print([str(i.name) for i in not_initialized_vars])
return not_initialized_vars
def train_z(params, left_ims, right_ims, Gs, batch_idx, result_path):
# solve dlatents
latents = np.zeros([params.data_size, Gs.input_shape[1]])
latents_in = tf.constant(latents)
dlatents = Gs.components.mapping.get_output_for(latents_in, None)
# build tf graph
with tf.variable_scope(tf.get_variable_scope()):
left = tf.constant(left_ims, dtype=tf.float32)
right = tf.constant(right_ims, dtype=tf.float32)
z_input = tf.Variable(dlatents[:, 0, :], dtype=tf.float32, name='contr_val')
# repeat 14 times
z_input_tile0 = tf.reshape(z_input, [params.data_size, 1, Gs.input_shape[1]])
z_input_tile = tf.tile(z_input_tile0, [1, 14, 1])
out_expr = Gs.components.synthesis.get_output_for(z_input_tile, truncation_psi=0.7,
randomize_noise=False)
compensate_disp = tf.Variable(params.compensate_disp, dtype=tf.float32, name='contr_val')
linear_interpolator = linear_sample.LinearInterpolator(params) # initialize linear interpolator
disp = decoder.decoder_forward1(z_input, linear_interpolator.sz_params, out_expr, compensate_disp * 10)
right_est = linear_interpolator.interpolate(left, disp)
# compute loss of reconstruction
if params.dataset_name == 'invivo':
compensateI = tf.Variable(4.3, dtype=tf.float32, name='contr_val')
loss_rec, compa_sum, loss_rec_sum = linear_sample.compute_rec_loss_per(right_est, right, compensateI, left,
linear_interpolator.sz_params, 160.0)
else:
# other dataset is 'phantom' as paper tested
compensateI = tf.Variable(-2.0, dtype=tf.float32, name='contr_val')
loss_rec, compa_sum, loss_rec_sum = linear_sample.scompute_rec_loss(right_est, right, compensateI, left,
linear_interpolator.sz_params, 160.0)
disp_i1 = tf.slice(disp, [0, params.cutTop, params.cutLeft, 0], [-1, 256, 255, -1],
name='r_clip')
disp_i2 = tf.slice(disp, [0, params.cutTop, params.cutLeft + 1, 0], [-1, 256, 255, -1],
name='r_clip')
disp_i3 = tf.slice(disp, [0, params.cutTop, params.cutLeft, 0], [-1, 255, 256, -1],
name='r_clip')
disp_i4 = tf.slice(disp, [0, params.cutTop + 1, params.cutLeft, 0], [-1, 255, 256, -1],
name='r_clip')
loss_wt_norm = tf.multiply(1e-4, tf.reduce_sum(tf.square(disp_i2 - disp_i1)), name='punishment')
loss_wt_norm2 = tf.multiply(1e-4, tf.reduce_sum(tf.square(disp_i4 - disp_i3)), name='punishment2')
loss = tf.add(loss_rec, loss_wt_norm, name='Total_loss')
loss = tf.add(loss, loss_wt_norm2, name='Total_loss')
train_op = optimize_op.minimize(loss, var_list=[z_input, compensate_disp]) # z_input_tile
# run session
tf.get_default_session().run(tf.variables_initializer(get_uninitialized_variables(tf.get_default_session())))
start_time = time.time()
loss_rec_temp = 0.
z_before = dlatents[:, 0, :].eval()
z_val = 0.
disp_val = np.zeros([params.data_size, params.height, params.width, 1])
est_right_val = np.zeros([params.data_size, params.height, params.width, params.channel])
step = 0
res_loss = [[], []]
for step in range(0, max_step):
_, z_val, loss_rec_val, compa_sum_val, loss_rec_sum_val, loss_val, compensate_disp_val, disp_val, est_right_val = tf.get_default_session().run(
[train_op, z_input, loss_rec, compa_sum, loss_rec_sum, loss, compensate_disp, disp, right_est]
)
res_loss[0].append(loss_rec_sum_val * params.data_size)
res_loss[1].append(compa_sum_val)
if 0 == step % 10 or step + 1 == max_step:
z_before = z_val
loss_var_interp = np.abs(loss_val - loss_rec_temp)
loss_rec_temp = loss_val
z_var_mean = np.mean(z_val - z_before)
print(print_str.format(step, loss_rec_val, loss_val, z_var_mean))
if loss_var_interp < 1e-3 or step >= 200:
break
z = z_val
print('time spent {:8} '.format(time.time() - start_time))
np.save(os.path.join(result_path, 'W_disp_batch' + str(batch_idx) + '.npy'), disp_val)
np.save(os.path.join(result_path, 'G_est_right_real_batch' + str(batch_idx) + '.npy'), est_right_val)
if __name__ == '__main__':
# application setting
params = model_params(data_size=10,
mini_batch_size=10,
learning_rate=5e-2,
total_epoch_num=np.int32(20000),
outputdir=r'output',
height=288,
width=360,
channel=3,
cutTop=32,
cutBottom=0,
cutLeft=14,
cutRight=90,
dataset_name='invivo',
cPointRow=5,
cPointCol=6,
compensate_disp=8.52
)
# modify the path of your own datasets and pre-trained styleGAN model
source_img_path = 'datasets/invivo1_rect/'
model_path = 'results/00001-sgan-MNdatasets-1gpu/network-snapshot-001520.pkl'
range_idx = [0] # in range of(range_idx)*data_size
# load pre-trained G
tflib.init_tf()
url = os.path.abspath(model_path)
# input zeros to solve latent vector w
with open(url, 'rb') as f:
_, _, Gs = pickle.load(f)
max_step = np.int32(params.total_epoch_num)
# learning policy
learning_rate_init = np.float32(params.learning_rate)
optimize_op = tf.train.AdamOptimizer(learning_rate_init) # ,0.9,0.999,1e-08
optimize_op_compensate_disp = tf.train.AdamOptimizer(5e-1) # ,0.9,0.999,1e-08
for idx in range_idx: # equal to from frame 601 to 610.
ids = range(idx * params.data_size, (idx + 1) * params.data_size)
left_ims, right_ims = read_images.read_stereo_images(source_img_path, ids)
train_z(params, left_ims, right_ims, Gs, idx, result_path = 'gt_z/')