-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathoutput_spatial_transform.py
217 lines (198 loc) · 9.1 KB
/
output_spatial_transform.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
import cv2
import numpy as np
from PIL import Image
import torch
import time
class Transform_output(torch.nn.Module):
"""Spatial Transformer Layer
Implements a spatial transformer layer as described in [1]_.
Based on [2]_ and edited by David Dao for Tensorflow.
Parameters
----------
U : float
The output of a convolutional net should have the
shape [num_batch, height, width, num_channels].
theta: float
The output of the
localisation network should be [num_batch, 6].
out_size: tuple of two ints
The size of the output of the network (height, width)
"""
def __init__(self):
super().__init__()
# self.grid = torch.reshape(torch.reshape(torch.unsqueeze(self._meshgrid(height, width),0), [-1]).repeat(batch_size),[batch_size, 3, -1]).to(device)
# base = _repeat(tf.range(num_batch) * dim1, out_height * out_width)
# self.base = self._repeat(torch.arange(batch_size) * (width * height), height * width).to(device)
def _repeat(self, x, n_repeats):
rep = torch.unsqueeze(torch.ones(int(n_repeats),), 1).permute([1, 0])
rep = torch.FloatTensor(rep)
x=x.float()
x = torch.matmul(torch.reshape(x, (-1, 1)), rep)
return torch.reshape(x, [-1])
def _interpolate(self,im, x, y, out_size):
# constants
batch_size = im.shape[0]
height = im.shape[1]
width = im.shape[2]
channels = im.shape[3]
x = x.float()
y = y.float()
height_f = height
width_f = width
out_height = out_size[0]
out_width = out_size[1]
zero=0
max_y = im.shape[1] - 1
max_x = im.shape[2] - 1
#scale indices from [-1, 1] to [0, width/height]
# x = (x + 1.0) * (width_f) / 2.0
# y = (y + 1.0) * (height_f) / 2.0
# do sampling
x0 = torch.floor(x)
x1 = x0 + 1
y0 = torch.floor(y)
y1 = y0 + 1
x0 = torch.clamp(x0, zero, max_x) #0-127 zuobiao
x1 = torch.clamp(x1, zero, max_x)
y0 = torch.clamp(y0, zero, max_y)
y1 = torch.clamp(y1, zero, max_y)
dim2 = width
dim1 = width * height
base = self._repeat(torch.arange(batch_size) * dim1, out_height * out_width).to(im.device)
# print(out_height*out_width)
base_y0 = base + y0 * dim2
base_y1 = base + y1 * dim2
idx_a = base_y0 + x0
idx_b = base_y1 + x0
idx_c = base_y0 + x1
idx_d = base_y1 + x1
# use indices to lookup pixels in the flat image and restore
# channels dim
# print(channels)
# print(im.shape)
# time.sleep(1000)
im_flat = torch.reshape(im, (-1, channels))
im_flat = im_flat.float()
Ia=im_flat[idx_a.type(torch.long)]
Ib=im_flat[idx_b.type(torch.long)]
Ic=im_flat[idx_c.type(torch.long)]
Id=im_flat[idx_d.type(torch.long)]
# and finally calculate interpolated values
x0_f = x0.float()
x1_f = x1.float()
y0_f = y0.float()
y1_f = y1.float()
wa = torch.unsqueeze(((x1_f - x) * (y1_f - y)), 1)
wb = torch.unsqueeze(((x1_f - x) * (y - y0_f)), 1)
wc = torch.unsqueeze(((x - x0_f) * (y1_f - y)), 1)
wd = torch.unsqueeze(((x - x0_f) * (y - y0_f)), 1)
output = sum([wa * Ia, wb * Ib, wc * Ic, wd * Id])
return output
def _transform(self,im, H, width_max, width_min, height_max, height_min):
num_batch = im.shape[0]
num_height = im.shape[2]
num_width = im.shape[3]
num_channels = im.shape[1]
out_width = width_max - width_min
out_height = height_max - height_min
grid = self._meshgrid(width_max, width_min, height_max, height_min)
grid = torch.unsqueeze(grid, 0)
grid = grid.view(-1)
grid = grid.repeat(num_batch)
# print(grid.shape)
# time.sleep(1000)
grid = grid.view(num_batch, 3, -1).to(im.device)
H = H.float()
T_g = torch.matmul(H, grid)
x_s = T_g[:, 0:1, :]
y_s = T_g[:, 1:2, :]
t_s = T_g[:, 2:3, :]
t_s_flat = torch.reshape(t_s, [-1])
one = torch.tensor(1, dtype=torch.float32)
small = torch.tensor(1e-7, dtype=torch.float32)
smallers = 1e-6 * (one - torch.gt(torch.abs(t_s_flat), small).float())
t_s_flat = t_s_flat + smallers
x_s_flat = torch.reshape(x_s, [-1]) / t_s_flat
y_s_flat = torch.reshape(y_s, [-1]) / t_s_flat
input_transformed = self._interpolate(im.permute(0,2,3,1), x_s_flat, y_s_flat, (out_height,out_width))
output = torch.reshape(input_transformed, (num_batch, out_height, out_width, num_channels)).permute(0,3,1,2)
# print(output.shape)
return output
return output
# def _meshgrid(self,height, width):
def _meshgrid(self,width_max, width_min, height_max, height_min):
width = width_max - width_min
height = height_max - height_min
# torch.linspace(width_min, width_max, width)
# torch.ones(shape=torch.stack([height, 1]))
# print(torch.ones(int(height), 1).shape)
# print("ohoho")
# print(torch.unsqueeze(torch.linspace(width_min, width_max, width), 1).permute([1, 0]).shape)
# x_t = torch.matmul(torch.ones(int(height), 1),
# torch.unsqueeze(torch.linspace(int(width_min, width_max, width), 1).permute([1, 0]))
x_t = torch.matmul(torch.ones(int(height), 1),
torch.unsqueeze(torch.linspace(float(width_min), float(width_max), int(width)), 1).permute([1, 0]))
# print("hahah")
y_t = torch.matmul(torch.unsqueeze(torch.linspace(float(height_min), float(height_max), int(height)), 1),
torch.ones(1, int(width)))
# print(height)
# print(width)
# print(height*width)
# print(x_t.shape)
# print(y_t.shape)
x_t_flat = torch.reshape(x_t, (1, -1))
y_t_flat = torch.reshape(y_t, (1, -1))
# print(x_t_flat.shape)
# print(y_t_flat.shape)
# time.sleep(1000)
ones = torch.ones_like(x_t_flat)
grid = torch.cat([x_t_flat, y_t_flat, ones], 0)
return grid
def forward(self, inputs, H,size,resized_shift):
# pts_1_tile = torch.repeat(size, [1, 4, 1])
pts_1_tile = torch.tile(size, (1, 4, 1))
tmp = torch.unsqueeze(torch.unsqueeze(torch.tensor([0., 0., 1., 0., 0., 1., 1., 1.], dtype=torch.float32, device=inputs.device), 0),-1)
# tmp = torch.unsqueeze(torch.tensor([0., 0., 1., 0., 0., 1., 1., 1.], shape=(8,1), dtype = tf.float32), [0])
pts_1 = pts_1_tile*tmp
pts_2 = resized_shift + pts_1
# pts1_list = torch.split(pts_1, 8, dim=1)
pts1_list = torch.split(pts_1, 1, dim=1)
pts2_list = torch.split(pts_2, 1, dim=1)
pts_list = pts1_list + pts2_list
width_list = [pts_list[i] for i in range(0, 16, 2)]
height_list = [pts_list[i] for i in range(1, 16, 2)]
width_list_tf = torch.cat(width_list, axis=1)
height_list_tf = torch.cat(height_list, axis=1)
width_max = int(torch.max(width_list_tf))
width_min = int(torch.min(width_list_tf))
height_max = int(torch.max(height_list_tf))
height_min = int(torch.min(height_list_tf))
out_width = int(width_max - width_min)
out_height = int(height_max - height_min)
batch_size=inputs.shape[0]
H_one = torch.eye(3)
H_one = torch.tile(torch.unsqueeze(H_one, 0), [batch_size, 1, 1]).to(inputs.device)
# pts_2 = tf.add(resized_shift, pts_1)
if inputs.shape[1]==2:
img1 = inputs[:, 0,...][:,None]
one = torch.ones_like(img1).float()
else:
img1 = inputs[:, 0:3,...]
one = torch.ones_like(img1).float()
img1 = self._transform(img1, H_one, width_max, width_min, height_max, height_min)
if inputs.shape[1]==2:
warp = inputs[:, 1,...][:,None]
else:
warp = inputs[:, 3:6,...]
warp = self._transform(warp, H, width_max, width_min, height_max, height_min)
# one = torch.ones_like(img1).float()
mask1 = self._transform(one, H_one, width_max, width_min, height_max, height_min)
mask2 = self._transform(one, H, width_max, width_min, height_max, height_min)
# resized_height = out_height - out_height%8
# resized_width = out_width - out_width%8
# img1 = torch.nn.functional.interpolate(img1, [resized_height, resized_width], method=0)
# warp = torch.nn.functional.interpolate(warp, [resized_height, resized_width], method=0)
# mask1 = torch.nn.functional.interpolate(mask1, [resized_height, resized_width], method=0)
# mask2 = torch.nn.functional.interpolate(mask2, [resized_height, resized_width], method=0)
output = torch.cat([img1, warp, mask1, mask2], axis=1)
return output