coupe.DVSG/model.py at master · leejonghyeop/coupe.DVSG · GitHub

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
import tensorflow as tf
import tensorlayer as tl
import tensorflow.contrib.slim as slim
from tensorflow.contrib.slim.nets import resnet_v2

import collections

from warp_with_optical_flow import *
from spatial_transformer import *
from networks import *

class StabNet:
    def __init__(self, h, w):
        self.h = h
        self.w = w
        self.c = 3
        self.control_point_num = 5 # grid_num (4) + 1
        self.reuse = collections.OrderedDict()
        self.elastic_transformer = ElasticTransformer(out_size = [self.h, self.w], param_dim = 2 * self.control_point_num ** 2, param_dim_per_side = self.control_point_num)


    def init_pretrain_inputs(self, sample_num):
        self.sample_num = sample_num
        self.inputs_pretrain = self.init_train_inputs(self.sample_num)
        return self.inputs_pretrain

    def get_pretrain_model(self, is_train):
        self.inputs = self.inputs_pretrain
        return self.get_train_model(is_train)

    def init_train_inputs(self, sample_num):
        self.sample_num = sample_num
        with tf.variable_scope('input'):
            self.inputs = collections.OrderedDict()
            self.inputs['patches_t_1'] = tf.placeholder('float32', [None, None, None, 3 * self.sample_num], name = 'input_frames_t_1')
            self.inputs['patches_t'] = tf.placeholder('float32', [None, None, None, 3 * self.sample_num], name = 'input_frames_t')

            self.inputs['s_t_1_gt'] = tf.placeholder('float32', [None, None, None, 3], name = 'stable_frame_t_1_gt')
            self.inputs['s_t_gt'] = tf.placeholder('float32', [None, None, None, 3], name = 'stable_frame_t_gt')

            self.inputs['u_t_1'] = tf.placeholder('float32', [None, None, None, 3], name = 'unstable_frame_t_1')
            self.inputs['u_t'] = tf.placeholder('float32', [None, None, None, 3], name = 'unstable_frame_t')

            self.inputs['of_t'] = tf.placeholder('float32', [None, None, None, 2], name = 'optical_flow_t')

            self.inputs['surfs_t_1'] = tf.placeholder('float32', [None, 2, None, 2], name = 'surfs_t_1')
            self.inputs['surfs_t'] = tf.placeholder('float32', [None, 2, None, 2], name = 'surfs_t')

            self.inputs['surfs_dim_t_1'] = tf.placeholder('float32', [None], name = 'surfs_dim_t_1')
            self.inputs['surfs_dim_t'] = tf.placeholder('float32', [None], name = 'surfs_dim_t')

        return self.inputs

    def get_train_model(self, is_train):
        outputs = collections.OrderedDict()
        outputs['patches_masked_t_1'], outputs['random_masks_t_1'] = self.random_mask(self.inputs['patches_t_1'], [self.h, self.w], self.sample_num)
        outputs['patches_masked_t'], outputs['random_masks_t'] = self.random_mask(self.inputs['patches_t'], [self.h, self.w], self.sample_num)

        with tf.variable_scope('stabNet') as scope:
            ## Regressor
            outputs['F_t_1'] = localizationNet(outputs['patches_masked_t_1'], is_train, self.get_reuse('stabNet'), scope = scope)
            outputs['F_t'] = localizationNet(outputs['patches_masked_t'], is_train, self.get_reuse('stabNet'), scope = scope)

            ## STN
            stl_affine = ProjectiveTransformer([self.h, self.w])

            outputs['s_t_1_pred'] = self.elastic_transformer.transform(self.inputs['u_t_1'], outputs['F_t_1'])
            outputs['s_t_1_pred_mask'] = self.elastic_transformer.transform(tf.ones_like(self.inputs['u_t_1']), outputs['F_t_1'])

            outputs['s_t_pred'] = self.elastic_transformer.transform(self.inputs['u_t'], outputs['F_t'])
            outputs['s_t_pred_mask'] = self.elastic_transformer.transform(tf.ones_like(self.inputs['u_t']), outputs['F_t'])
            outputs['s_t_pred_warped'] = tf_warp(outputs['s_t_pred'], self.inputs['of_t'], self.h, self.w)
            outputs['s_t_pred_warped_mask'] = tf_warp(outputs['s_t_pred_mask'], self.inputs['of_t'], self.h, self.w)

        return outputs

    def get_evaluation_model(self, sample_num):
        self.sample_num = sample_num
        with tf.variable_scope('input'):
            ### PLACE HOlDERS ###
            inputs = collections.OrderedDict()
            inputs['patches'] = tf.placeholder('float32', [None, None, None, 3 * 2 * self.sample_num], name = 'input_frames_t')
            inputs['u'] = tf.placeholder('float32', [None, None, None, 3], name = 'unstable_frame_t')

            batch_size = tf.shape(inputs['u'])[0]
            h = tf.shape(inputs['u'])[1]
            w = tf.shape(inputs['u'])[2]

            patches_masked, _ = self.random_mask(inputs['patches'], [h, w], self.sample_num, True)

        with tf.variable_scope('main_net') as scope:
            with tf.variable_scope('stabNet') as scope:
                F = self.localizationNet(patches_masked, is_train = False, reuse = False, scope = scope)

            with tf.variable_scope('spatial_transformer') as scope:
                stl_affine = ProjectiveTransformer([h, w])
                s_pred = stl_affine.transform(inputs['u'], F)

        self.inputs = inputs
        self.output = s_pred
        return self.inputs, self.output

    def init_vars(self, sess):
        exclude_scope = 'stabNet/resnet_v1_50/conv1'
        variables_to_restore = collections.OrderedDict()
        for var in slim.get_model_variables():
            if var.op.name.startswith(exclude_scope) == False:
                variables_to_restore[var.op.name.replace('stabNet/', '')] = var
        init_function = slim.assign_from_checkpoint_fn('./pretrained/resnet_v1_50.ckpt', variables_to_restore, ignore_missing_vars=True)
        init_function(sess)

    def random_mask(self, patches, out_size, sample_num):
        mask_affine = ProjectiveTransformer(out_size)
        batch_size = tf.shape(patches)[0]
        mask = tf.ones_like(patches)
        H = tf.random_uniform([batch_size, 8], minval = -1, maxval = 1)
        H = H * tf.constant([0.1, 0.1, 0.5, 0.1, 0.1, 0.5, 0.1, 0.1])
        H = H + tf.constant([1.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0])
        mask = mask_affine.transform(mask, H)

        return patches * mask, mask

    def get_reuse(self, scope):
        if scope in list(self.reuse.keys()):
            self.reuse[scope] = True if self.reuse[scope] is False else True
        else:
            self.reuse[scope] = False

        return self.reuse[scope]

    def get_vars_train(self):
        return self._get_vars('stabNet', True, False)

    def get_save_vars_train(self):
        return self._get_vars('stabNet', False, False)

    def get_vars_pretrain(self):
        return self._get_vars('stabNet', True, False)

    def get_save_vars_pretrain(self):
        return self._get_vars('stabNet', False, False)

    def _get_vars(self, name, train_only, verbose, exclude = None):
        if exclude is None:
            return tl.layers.get_variables_with_name(name, train_only, verbose)
        else:
            return [var for var in tl.layers.get_variables_with_name(name, train_only, verbose) if exclude not in var.name]