-
Notifications
You must be signed in to change notification settings - Fork 7
/
tta.py
174 lines (135 loc) · 6.93 KB
/
tta.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
from evaluate.reasoning_dataloader import background_transforms
from evaluate.mae_utils import *
import torch
from PIL import Image
import numpy as np
DEVICE = 'cuda'
h, w = 224, 224
class RowColShuffle(torch.nn.Module):
def __init__(self, shuffle_rows=False, shuffle_cols=False, num_rows=3):
super(RowColShuffle, self).__init__()
self.shuffle_rows = shuffle_rows
self.shuffle_cols = shuffle_cols
self.num_rows = num_rows
def forward(self, pairs):
background_image = Image.new('RGB', (224, 224), color='black')
canvas = background_transforms(background_image)
v_order = np.arange(0, self.num_rows)
if self.shuffle_rows:
np.random.shuffle(v_order)
shuffle_cols = False
if self.shuffle_cols:
shuffle_cols = np.random.choice([True, False])
padding = 1
figure_size = 74
for i in range(len(pairs)):
img, label = pairs[v_order[i]]
start_row = i * (figure_size + padding)
if shuffle_cols:
img, label = label, img
canvas[:, start_row:start_row + figure_size, 224 // 2 - figure_size:224 // 2] = img
canvas[:, start_row:start_row + figure_size, 224 // 2 + 1: 224 // 2 + 1 + figure_size] = label
pred_row_idx = np.where(v_order == 2)[0][0]
pred_col_idx = 1 if not shuffle_cols else 0
canvas = np.array(canvas)
# keep all but occluded part
mask_psuedo_gt = np.ones((14, 14))
row_mask_start = int(np.floor(14 * float(pred_row_idx) / 3))
row_mask_end = int(np.ceil(14 * float(pred_row_idx + 1) / 3)) + 1
mask_psuedo_gt[row_mask_start:row_mask_end, 2 + pred_col_idx * 5:2 + pred_col_idx * 5 + 5] = 0
# keep everything in 20% except for the occluded part
mask = np.round(np.random.uniform(0, 1, (14, 14)) >= 0.5)
mask[row_mask_start:row_mask_end, 2 + pred_col_idx * 5:2 + pred_col_idx * 5 + 5] = 0
_mask = obtain_values_from_mask(mask)
_mask_psuedo_gt = obtain_values_from_mask(mask_psuedo_gt)
return canvas, len(_mask), fill_to_full(_mask), mask, len(_mask_psuedo_gt), fill_to_full(
_mask_psuedo_gt), mask_psuedo_gt
def shuffle_cols(self, canvas, num_cols, h_order, fig_size, border_size):
new_canvas = np.zeros_like(canvas)
for i in range(num_cols):
col_start = 224 - fig_size + i * (fig_size + border_size)
original_col_start = 224 - fig_size + h_order[i] * (fig_size + border_size)
new_canvas[:, :, col_start:col_start + fig_size] = canvas[:, :,
original_col_start: original_col_start + fig_size]
return new_canvas
def shuffle_rows(self, canvas, num_rows, v_order, fig_size, border_size):
new_canvas = np.zeros_like(canvas)
for i in range(num_rows):
start_col = i * (fig_size + border_size)
old_start_col = v_order[i] * (fig_size + border_size)
new_canvas[:, start_col:start_col + fig_size] = canvas[:, old_start_col: old_start_col + fig_size]
return new_canvas
def reverse_trans(im_paste, v_order, shuffle_cols, transpose):
background_image = Image.new('RGB', (224, 224), color='black')
new_canvas = np.array(background_image)
if transpose:
im_paste = np.transpose(im_paste, [1, 0, 2])
padding = 1
figure_size = 74
for i in range(len(v_order)):
start_row = i * (figure_size + padding)
img = im_paste[start_row: start_row + figure_size, 224 // 2 - figure_size - 1: 224 // 2 - 1]
label = im_paste[start_row:start_row + figure_size, 224 // 2 + 1: 224 // 2 + 1 + figure_size]
if shuffle_cols:
img, label = label, img
start_row = v_order[i] * (figure_size + padding)
new_canvas[start_row:start_row + figure_size, 224 // 2 - figure_size - 1:224 // 2 - 1] = img
new_canvas[start_row:start_row + figure_size, 224 // 2 + 1: 224 // 2 + 1 + figure_size] = label
return new_canvas
class TTA(torch.nn.Module):
def __init__(self, shuffle_rows=False, shuffle_cols=False, transpose=False, num_rows=3):
super(TTA, self).__init__()
self.shuffle_rows = shuffle_rows
self.shuffle_cols = shuffle_cols
self.transpose = transpose
self.num_rows = num_rows
def forward(self, pairs):
background_image = Image.new('RGB', (224, 224), color='black')
canvas = background_transforms(background_image)
v_order = np.arange(0, self.num_rows)
if self.shuffle_rows:
v_order = [2, 0, 1]
shuffle_cols = False
if self.shuffle_cols:
shuffle_cols = True
padding = 1
figure_size = 74
for i in range(len(pairs)):
img, label = pairs[v_order[i]]
start_row = i * (figure_size + padding)
if shuffle_cols:
img, label = label, img
canvas[:, start_row:start_row + figure_size, 224 // 2 - figure_size - 1:224 // 2 - 1] = img
canvas[:, start_row:start_row + figure_size, 224 // 2 + 1: 224 // 2 + 1 + figure_size] = label
pred_row_idx = np.where(v_order == 2)[0][0]
pred_col_idx = 1 if not shuffle_cols else 0
canvas = np.array(canvas)
# keep all but occluded part
mask_psuedo_gt = np.ones((14, 14))
row_mask_start = int(np.floor(14 * float(pred_row_idx) / 3))
row_mask_end = int(np.ceil(14 * float(pred_row_idx + 1) / 3)) + 1
mask_psuedo_gt[row_mask_start:row_mask_end, 2 + pred_col_idx * 5:2 + pred_col_idx * 5 + 5] = 0
transpose_img = False
if self.transpose:
transpose_img = True
if transpose_img:
mask_psuedo_gt = np.transpose(mask_psuedo_gt, [1, 0])
canvas = np.transpose(canvas, [0, 2, 1])
_mask_psuedo_gt = obtain_values_from_mask(mask_psuedo_gt)
return canvas, len(_mask_psuedo_gt), fill_to_full(
_mask_psuedo_gt), mask_psuedo_gt, v_order, shuffle_cols, transpose_img
def shuffle_cols(self, canvas, num_cols, h_order, fig_size, border_size):
new_canvas = np.zeros_like(canvas)
for i in range(num_cols):
col_start = 224 - fig_size + i * (fig_size + border_size)
original_col_start = 224 - fig_size + h_order[i] * (fig_size + border_size)
new_canvas[:, :, col_start:col_start + fig_size] = canvas[:, :,
original_col_start: original_col_start + fig_size]
return new_canvas
def shuffle_rows(self, canvas, num_rows, v_order, fig_size, border_size):
new_canvas = np.zeros_like(canvas)
for i in range(num_rows):
start_col = i * (fig_size + border_size)
old_start_col = v_order[i] * (fig_size + border_size)
new_canvas[:, start_col:start_col + fig_size] = canvas[:, old_start_col: old_start_col + fig_size]
return new_canvas