-
Notifications
You must be signed in to change notification settings - Fork 6
/
AerialDepthEstimator.py
235 lines (182 loc) · 10.8 KB
/
AerialDepthEstimator.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
from __future__ import division
import os
import time
import math
import numpy as np
import tensorflow as tf
from data_loader import DataLoader
from nets import *
from utils import *
from layer import *
class AerialDepthEstimator(object):
def __init__(self):
pass
def train(self, opt):
self.opt = opt
with tf.Graph().as_default() as g:
self.build_train_graph()
self.build_summaries()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.allow_soft_placement = True
with tf.name_scope("parameter_count"):
parameter_count = tf.reduce_sum([tf.reduce_prod(tf.shape(v)) for v in tf.trainable_variables()])
self.saver = tf.train.Saver(max_to_keep=100)
sv = tf.train.Supervisor(logdir=opt.checkpoint_dir, save_summaries_secs=0, saver=None)
with sv.managed_session(config=config) as sess:
print("parameter_count =", sess.run(parameter_count))
if opt.continue_train:
if opt.init_checkpoint_file is None:
checkpoint = tf.train.latest_checkpoint(opt.checkpoint_dir)
else:
checkpoint = opt.init_checkpoint_file
print("Resume training from previous checkpoint: %s" % checkpoint)
self.saver.restore(sess, checkpoint)
start_time = time.time()
for step in range(1, opt.max_epochs*self.steps_per_epoch):
fetches = {
"train": self.train_op,
"global_step": self.global_step,
"incr_global_step": self.incr_global_step
}
if step % opt.summary_freq == 0:
fetches["loss"] = self.total_loss
fetches["mean_inverse_depth"] = self.mean_inverse_depth
fetches["smoth_loss"] = self.smooth_loss
fetches["summary"] = sv.summary_op
results = sess.run(fetches)
gs = results["global_step"]
if step % opt.summary_freq == 0:
sv.summary_writer.add_summary(results["summary"], gs)
train_epoch = math.ceil(gs / self.steps_per_epoch)
train_step = gs - (train_epoch - 1) * self.steps_per_epoch
print("Epoch: [%2d] [%5d/%5d] time: %4.4f/it loss: %.3f smooth_loss: %.10f mean inverse depth: %.3f" % (train_epoch, train_step, self.steps_per_epoch,
(time.time() - start_time)/opt.summary_freq, results["loss"],
results["smoth_loss"],results["mean_inverse_depth"]))
start_time = time.time()
if step % opt.save_latest_freq == 0:
print("save1")
self.save(sess, opt.checkpoint_dir, 'latest')
print("sucess")
if step % (self.steps_per_epoch*opt.save_epoch) == 0:
print("save2")
self.save(sess, opt.checkpoint_dir, gs)
print("sucess")
self.save(sess, opt.checkpoint_dir, gs)
def build_train_graph(self):
opt = self.opt
loader = DataLoader(opt.dataset_dir,
opt.batch_size,
opt.img_height,
opt.img_width,
opt.num_source,
opt.num_scales)
with tf.device('/cpu:0'):
with tf.name_scope("data_loading"):
images, intrinsics = loader.load_train_batch()
self.steps_per_epoch = loader.steps_per_epoch
for i in range(len(images)):
images[i]=self.preprocess_image( images[i])
tgt_image = images[int(opt.num_source/2)]
del images[int(opt.num_source/2)]
src_image_stack = tf.concat(images, axis=3)
self.tgt_image = tgt_image
self.src_image_stack = src_image_stack
with tf.device('/gpu:'+opt.gpu_id):
tgt_image_transposed=tgt_image
src_image_stack_transposed=src_image_stack
pred_disp = disp_net(tgt_image_transposed)
pred_depth = [convert_disp_to_depth(i ,opt.img_height, opt.img_width) for i in pred_disp]
pred_poses_01=pose_net_new(src_image_stack_transposed[:, :, :, 0:3],tgt_image_transposed)
pred_poses_12=pose_net_new(src_image_stack_transposed[:, :, :, 3:6],tgt_image_transposed)
pred_poses=tf.concat([pred_poses_01,pred_poses_12],axis=1)
pred_poses = pred_poses*0.01
with tf.name_scope("compute_loss"):
pixel_loss, smooth_loss, mean_inverse_depth = 0, 0, 0
proj_image_stack=[]
proj_error_stack=[]
for s in range(opt.num_scales):
mean_inverse_depth += tf.reduce_mean(1/pred_depth[s])
#current_downscaled_tgt_image = tf.compat.v1.image.resize_area(tgt_image, [int(opt.img_height/(2**s)), int(opt.img_width/(2**s))])
current_downscaled_tgt_image = tf.image.resize_area(tgt_image, [int(opt.img_height/(2**s)), int(opt.img_width/(2**s))])
smooth_loss += opt.smooth_weight / (2**s) * compute_edge_aware_smooth_loss(pred_disp[s], current_downscaled_tgt_image)
pixel_loss_stack = []
for i in range(opt.num_source):
# Inverse warp the source image to the target image frame
curr_proj_image, rot_mat_pred = projective_inverse_warp(
src_image_stack_transposed[:, :, :, 3*i:3*(i+1)],
tf.squeeze(pred_depth[s], axis=3),
pred_poses[:, i, :],
intrinsics[:, 0, :, :],
None)
proj_image_stack.append(curr_proj_image)
curr_proj_error_l1 = tf.abs(curr_proj_image - tgt_image)
curr_proj_error_l2 = tf.square(curr_proj_image - tgt_image)
curr_proj_error_SSIM = compute_SSIM_loss(curr_proj_image, tgt_image)
curr_proj_error_l1_SSIM=(1-opt.ssim_weight)*curr_proj_error_l1+opt.ssim_weight*curr_proj_error_SSIM
proj_error_stack.append(curr_proj_error_l1_SSIM)
pixel_loss_stack.append(curr_proj_error_l1_SSIM)
proj_min_error=tf.reduce_min(pixel_loss_stack, axis=[0])
pixel_loss += tf.reduce_mean(proj_min_error)
total_loss = pixel_loss + smooth_loss
with tf.name_scope("train_op"):
self.global_step = tf.Variable(0, name='global_step', trainable=False)
self.incr_global_step = tf.assign(self.global_step, self.global_step+1)
do_reduce_learning_rate = tf.greater(self.global_step, self.steps_per_epoch*opt.epoch_to_reduce_learning_rate)
learning_rate = tf.cond(do_reduce_learning_rate, lambda: opt.learning_rate, lambda: opt.learning_rate/10)
optim = tf.train.AdamOptimizer(learning_rate, opt.beta1)
self.train_op = optim.minimize(total_loss)
self.pred_depth = pred_depth
self.pred_poses = pred_poses
self.total_loss = total_loss
self.smooth_loss = smooth_loss
self.mean_inverse_depth = mean_inverse_depth
self.proj_image_stack = proj_image_stack
self.proj_error_stack = proj_error_stack
self.proj_min_error= proj_min_error
def preprocess_image(self, image):
# Assuming input image is uint8
image = tf.image.convert_image_dtype(image, dtype=tf.float32)
return image * 2. - 1.
def deprocess_image(self, image):
# Assuming input image is float32
image = (image + 1.)/2.
return tf.image.convert_image_dtype(image, dtype=tf.uint8)
def save(self, sess, checkpoint_dir, step):
model_name = 'model'
print(" [*] Saving checkpoint to %s..." % checkpoint_dir)
if step == 'latest':
self.saver.save(sess, os.path.join(checkpoint_dir, model_name + '.latest'))
else:
self.saver.save(sess, os.path.join(checkpoint_dir, model_name), global_step=step)
def build_summaries(self):
tf.summary.scalar("smooth_loss", self.smooth_loss)
tf.summary.scalar("mean_inverse_depth", self.mean_inverse_depth)
tf.summary.scalar("total_loss", self.total_loss)
tf.summary.image('scale0_disparity_image', 1./self.pred_depth[0])
tf.summary.image('scale0_disparity_image_RGB', depth_to_rgb_image_batch(self.pred_depth[0]))
tf.summary.image('scale0_target_image', self.deprocess_image(self.tgt_image))
tf.summary.image('scale0_source_image_0', self.deprocess_image(self.src_image_stack[:, :, :, :3]))
tf.summary.image('scale0_source_image_1', self.deprocess_image(self.src_image_stack[:, :, :, 3*(self.opt.num_source-1):]))
tf.summary.image('scale0_projected_image_0', self.deprocess_image(self.proj_image_stack[0]))
tf.summary.image('scale0_projected_image_1', self.deprocess_image(self.proj_image_stack[1]))
tf.summary.image('scale0__proj_error_image_0',self.deprocess_image(tf.clip_by_value(self.proj_error_stack[0] - 1, -1, 1)))
tf.summary.image('scale0__proj_error_image_1',self.deprocess_image(tf.clip_by_value(self.proj_error_stack[-1] - 1, -1, 1)))
tf.summary.image('scale0_proj_min_error', self.deprocess_image(tf.clip_by_value(self.proj_min_error- 1, -1, 1)))
def get_disp_model(self, input_shape):
input_layer = Input(input_shape)
input_uint8 = Lambda(self.preprocess_image)(input_layer)
disp, _, _, _ = disp_net(input_uint8)
return Model(inputs=input_layer, outputs=[disp])
def get_disp_model_for_export(self, input_shape):
input_layer = Input(input_shape,name="input_layer")
disp, _, _, _ = disp_net(input_layer)
return Model(inputs=input_layer, outputs=[disp])
def get_pose_model(self):
input_layer = Input(input_shape)
src_image_1, tgt_image, src_image_2 = Lambda(lambda x: tf.unstack(x, axis=1))(input_layer)
input_src_image_1_uint8 = Lambda(self.preprocess_image)(src_image_1)
input_tgt_image_uint8 = Lambda(self.preprocess_image)(tgt_image)
input_src_image_2_uint8 = Lambda(self.preprocess_image)(src_image_2)
pred_poses = pose_net_old(input_src_image_1_uint8, input_tgt_image_uint8, input_src_image_2_uint8)
return Model(inputs=[input_layer], outputs=[pred_poses])