DAMT/ops.py at main · jongdory/DAMT · 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
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
import random
import numpy as np
from numpy.random import randint
import matplotlib.pyplot as plt
import torch
from einops import rearrange


def patch_rand_drop(args, x, x_rep=None, max_drop=0.3, max_block_sz=0.25, tolr=0.05):
    c, h, w, z = x.size()
    n_drop_pix = np.random.uniform(0, max_drop) * h * w * z
    mx_blk_height = int(h * max_block_sz)
    mx_blk_width = int(w * max_block_sz)
    mx_blk_slices = int(z * max_block_sz)
    tolr = (int(tolr * h), int(tolr * w), int(tolr * z))
    total_pix = 0
    while total_pix < n_drop_pix:
        rnd_r = randint(0, h - tolr[0])
        rnd_c = randint(0, w - tolr[1])
        rnd_s = randint(0, z - tolr[2])
        rnd_h = min(randint(tolr[0], mx_blk_height) + rnd_r, h)
        rnd_w = min(randint(tolr[1], mx_blk_width) + rnd_c, w)
        rnd_z = min(randint(tolr[2], mx_blk_slices) + rnd_s, z)
        if x_rep is None:
            x_uninitialized = torch.empty(
                (c, rnd_h - rnd_r, rnd_w - rnd_c, rnd_z - rnd_s), dtype=x.dtype, device=args.local_rank
            ).normal_()
            x_uninitialized = (x_uninitialized - torch.min(x_uninitialized)) / (
                torch.max(x_uninitialized) - torch.min(x_uninitialized)
            )
            x[:, rnd_r:rnd_h, rnd_c:rnd_w, rnd_s:rnd_z] = x_uninitialized
        else:
            x[:, rnd_r:rnd_h, rnd_c:rnd_w, rnd_s:rnd_z] = x_rep[:, rnd_r:rnd_h, rnd_c:rnd_w, rnd_s:rnd_z]
        total_pix = total_pix + (rnd_h - rnd_r) * (rnd_w - rnd_c) * (rnd_z - rnd_s)
    return x


def rot_rand(args, x_s, a_s):
    img_n = x_s.size()[0]
    x_aug = x_s.detach().clone()
    a_aug = a_s.detach().clone()
    device = torch.device(f"cuda:{args.local_rank}")
    x_rot = torch.zeros(img_n).long().to(device)
    for i in range(img_n):
        x = x_s[i]
        a = a_s[i]
        orientation = np.random.randint(0, 10)
        if orientation == 0:
            pass
        elif orientation == 1:
            x = x.rot90(1, (2, 3))
            a = a.rot90(1, (2, 3))
        elif orientation == 2:
            x = x.rot90(2, (2, 3))
            a = a.rot90(2, (2, 3))
        elif orientation == 3:
            x = x.rot90(3, (2, 3))
            a = a.rot90(3, (2, 3))
        elif orientation == 4:
            x = x.rot90(1, (1, 3))
            a = a.rot90(1, (1, 3))
        elif orientation == 5:
            x = x.rot90(2, (1, 3))
            a = a.rot90(2, (1, 3))
        elif orientation == 6:
            x = x.rot90(3, (1, 3))
            a = a.rot90(3, (1, 3))
        elif orientation == 7:
            x = x.rot90(1, (1, 2))
            a = a.rot90(1, (1, 2))
        elif orientation == 8:
            x = x.rot90(2, (1, 2))
            a = a.rot90(2, (1, 2))
        elif orientation == 9:
            x = x.rot90(3, (1, 2))
            a = a.rot90(3, (1, 2))
        x_aug[i] = x
        a_aug[i] = a
        x_rot[i] = orientation

    return x_aug, a_aug, x_rot


def rot_rand_v2(args, x_s, permutations):
    img_n = x_s.size()[0]
    x_aug = torch.cat((x_s,x_s,x_s,x_s)).detach().clone()
    device = torch.device(f"cuda:{args.local_rank}")
    x_rot = torch.zeros(img_n*4).long().to(device)
    for i in range(img_n):
        x = x_s[i]
        perm = np.random.randint(len(permutations))
        for j, ori in enumerate(permutations[perm]):
            if ori == 0:
                pass
            elif ori == 1:
                x = x.rot90(1, dims=[1,2])
            elif ori == 2:
                x = x.rot90(2, dims=[1,2])
            elif ori == 3:
                x = x.rot90(3, dims=[1,2])
            x_aug[i*4 + j] = x
            x_rot[i*4 + j] = ori

    return x_aug, x_rot


def aug_rand(args, samples):
    img_n = samples.size()[0]
    x_aug = samples.detach().clone()
    for i in range(img_n):
        x_aug[i] = patch_rand_drop(args, x_aug[i])
        idx_rnd = randint(0, img_n)
        if idx_rnd != i:
            x_aug[i] = patch_rand_drop(args, x_aug[i], x_aug[idx_rnd])
    return x_aug


def aug_rand_v2(args, x_s):
    li = [0,1,2,3]
    img_n = x_s.size()[0]
    x_aug = x_s.detach().clone()
    device = torch.device(f"cuda:{args.local_rank}")
    x_rot = torch.zeros(img_n).long().to(device)

    for i in range(0, img_n, 4):
        li_a = random.sample(li, 4)
        li_b = list(set(li) - set(li_a))

        for j in li_a:
            x = x_s[i+j]
            orientation = np.random.randint(0, 4)
            if orientation == 0:
                pass
            elif orientation == 1:
                x = x.rot90(1, dims=[1,2])
            elif orientation == 2:
                x = x.rot90(2, dims=[1,2])
            elif orientation == 3:
                x = x.rot90(3, dims=[1,2])

            x_aug[i+j] = patch_rand_drop(args, x)
            idx_rnd = randint(0, img_n)
            if idx_rnd != (i+j):
                x_aug[i+j] = patch_rand_drop(args, x_aug[i+j], x_aug[idx_rnd])
            x_rot[i+j] = orientation


    return x_aug, x_rot


def jig_rand(args, x_s, permutations):

    x_s = rearrange(x_s, "(b p) c h w -> b p c h w", p=4)

    img_n = x_s.shape[0]
    x_jig = []

    for i in range(img_n):
        x = x_s[i]
        tiles = [x_tiles for x_tiles in x]

        perm = np.random.randint(len(permutations))
        data = [tiles[permutations[perm][t]] for t in range(4)]
        jig = torch.tensor([[permutations[perm][t]] for t in range(4)]).to(args.device)
        data = torch.stack(data, 0)

        x_s[i] = data
        x_jig.append(jig)

    x_jig = torch.stack(x_jig, 0)
    x_jig = rearrange(x_jig, "b p c -> (b p) c ", p=4).squeeze()
    x_aug = rearrange(x_s, "b p c h w -> (b p) c h w", p=4)

    return x_aug, x_jig

def get_atlas_mask(args, atlas):
    img_n = atlas.size()[0]
    device = torch.device(f"cuda:{args.local_rank}")
    contents = torch.zeros((img_n,120)).to(device)
    masks = torch.ones((img_n,120)).to(device)
    atlas_masks = torch.ones_like(atlas)
    for i in range(img_n):
        for j in range(120):
            contents[i,j] = torch.numel(atlas[atlas==j])
            if contents[i,j] < 100:
                masks[i,j] = 0
            if contents[i,0] == 1:
                masks[i] = masks[i] * 0
                atlas_masks[i] = atlas_masks[i] * 0
                break
    return atlas_masks, masks

def get_feature_mask(args, mask):
    img_n = mask.size()[0]
    device = torch.device(f"cuda:{args.local_rank}")
    fmask = torch.zeros((img_n,138))
    for i in range(img_n):
        fmask[i] = torch.tensor([1] + [w for w in mask[i][51:85] for i in range(2)] + \
                                [1] + [w for w in mask[i][86:] for i in range(2)])
    return fmask.to(device)