ADViSE/datasets.py at master · FHomps/ADViSE · 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
import random
import numpy as np
import collections

from torch.utils.data import Dataset
from PIL import Image
import torchvision.transforms as T
import torchvision.transforms.functional as TF

import torch
from torch import tensor

class LandingZoneDataset(Dataset):
    def __init__(self, img_tensor, gt_tensor, selection, out_res = 128,
                 min_zoom = 0.125, max_zoom = 8, padding_mode="wrap", rotate = True, twin_img_tensor = None):

        self.selection = tensor(selection, dtype=torch.int64)
        self.images = torch.index_select(img_tensor, 0, self.selection)
        self.ground_truth = torch.index_select(gt_tensor, 0, self.selection)
        if twin_img_tensor != None:
            self.images = torch.cat((self.images, torch.index_select(twin_img_tensor, 0, self.selection)), 0)
            self.ground_truth = self.ground_truth.repeat((2, 1, 1, 1))

        self.img_size = self.images.size()[-1]
        self.gt_size = self.ground_truth.size()[-1]

        self.out_res = out_res
        self.min_zoom = min_zoom
        self.max_zoom = max_zoom
        if rotate:
            self.transform = self.transform_rotate
        else:
            self.transform = self.transform_norotate
        self.padding_mode = padding_mode

    def transform_rotate(self, image, image_gt):
        if image.dim() == 3:
            channels_img, _, img_res = image.size()
            channels_gt, _, gt_res = image_gt.size()
        else:
            batchSize, channels_img, _, img_res = image.size()
            _, channels_gt, _, gt_res = image_gt.size()

        img_gt_ratio = round(img_res / gt_res)

        z = np.exp(random.uniform(np.log(self.min_zoom), np.log(self.max_zoom)))

        # Compute extended crop boundaries (with margin for rotate)

        # First, get standard crop boundaries.
        if z < 1:
            b_h = b_w = round(gt_res / z)
            b_t = b_l = round((gt_res - b_h) / 2)
        else:
            # RandomResizedCrop returns a crop within the bounds of the image
            RRC_scale = 1 / (z * z)
            b_t, b_l, b_h, b_w = T.RandomResizedCrop.get_params(image_gt, (RRC_scale, RRC_scale), (1., 1.))

        # The crop boundaries are extended by an optimal margin to include more data and minimize black corners after rotation
        angle = random.random() * 360
        alpha = (45 - abs((angle % 90) - 45)) * np.pi / 180
        expansionRatio = np.cos(alpha) + np.sin(alpha)
        crop_margin = int((expansionRatio - 1) * b_w / 2 + 1)

        eb_t = b_t - crop_margin
        eb_l = b_l - crop_margin
        eb_b = eb_t + b_h + 2 * crop_margin
        eb_r = eb_l + b_w + 2 * crop_margin

        # Get restricted bounds: intersection of extended bounds and image bounds
        rb_t = max(eb_t, 0)
        rb_l = max(eb_l, 0)
        rb_b = min(eb_b, gt_res)
        rb_r = min(eb_r, gt_res)

        gt = TF.crop(image_gt, rb_t, rb_l, rb_b - rb_t, rb_r - rb_l)
        img = TF.crop(image, *map(lambda x : round(x * img_gt_ratio), (rb_t, rb_l, rb_b - rb_t, rb_r - rb_l)))

        # Get the scaled down bounds
        crop_margin, eb_t, eb_l, eb_b, eb_r, rb_t, rb_l, rb_b, rb_r = map(lambda x : round(x * self.out_res / b_w),
       (crop_margin, eb_t, eb_l, eb_b, eb_r, rb_t, rb_l, rb_b, rb_r))

        # Scale down the crop
        img = TF.resize(img, (rb_b - rb_t, rb_r - rb_l), interpolation=Image.BILINEAR)
        gt = TF.resize(gt, (rb_b - rb_t, rb_r - rb_l), interpolation=Image.BILINEAR)

        # Reflect pad up to extended boundaries
        if image.dim() == 3:
            pad_size = ((0, 0), (rb_t - eb_t, eb_b - rb_b), (rb_l - eb_l, eb_r - rb_r))
        else:
            pad_size = ((0, 0), (0, 0), (rb_t - eb_t, eb_b - rb_b), (rb_l - eb_l, eb_r - rb_r))
        # PyTorch's pad doesn't support reflect padding on big areas, so numpy is used
        img = torch.from_numpy(np.pad(img.numpy(), pad_size, mode=self.padding_mode))
        gt = torch.from_numpy(np.pad(gt.numpy(), pad_size, mode=self.padding_mode))

        # Now, do the rotation and crop to the final size
        img = TF.rotate(img, angle, resample=Image.BILINEAR)
        gt = TF.rotate(gt, angle, resample=Image.BILINEAR)

        img = TF.crop(img, crop_margin, crop_margin, self.out_res, self.out_res)
        gt = TF.crop(gt, crop_margin, crop_margin, self.out_res, self.out_res)

        return (img.float() / 255, gt.float() / 255)

    def transform_norotate(self, image, image_gt):
        if image.dim() == 3:
            channels_img, _, img_res = image.size()
            channels_gt, _, gt_res = image_gt.size()
        else:
            batchSize, channels_img, _, img_res = image.size()
            _, channels_gt, _, gt_res = image_gt.size()

        img_gt_ratio = round(img_res / gt_res)

        z = np.exp(random.uniform(np.log(self.min_zoom), np.log(self.max_zoom)))

        if z >= 1:
            # RandomResizedCrop returns a crop within the bounds of the image
            RRC_scale = 1 / (z * z)
            b_t, b_l, b_h, b_w = T.RandomResizedCrop.get_params(image_gt, (RRC_scale, RRC_scale), (1., 1.))
            img = TF.crop(image, *map(lambda x : round(x * img_gt_ratio), (b_t, b_l, b_h, b_w)))
            gt = TF.crop(image_gt, b_t, b_l, b_h, b_w)
            img = TF.resize(img, self.out_res)
            gt = TF.resize(gt, self.out_res)
            return (img.float() / 255, gt.float() / 255)

        else:
            # Get scaled down crop boundaries
            b_h = b_w = round(gt_res / z)
            b_t = b_l = round((gt_res - b_h) / 2)

            zone_res, b_h, b_w, b_t, b_l = map(lambda x : round(x * self.out_res / b_w),
           (gt_res,   b_h, b_w, b_t, b_l))

            # Scale down the image
            img = TF.resize(image, (zone_res, zone_res), interpolation=Image.BILINEAR)
            gt = TF.resize(image_gt, (zone_res, zone_res), interpolation=Image.BILINEAR)

            # Reflect pad up to extended boundaries
            lpad = round((self.out_res - zone_res) / 2)
            rpad = self.out_res - zone_res - lpad
            if image.dim() == 3:
                pad_size = ((0, 0), (lpad, rpad), (lpad, rpad))
            else:
                pad_size = ((0, 0), (0, 0), (lpad, rpad), (lpad, rpad))
            # PyTorch's pad doesn't support reflect padding on big areas, so numpy is used
            img = torch.from_numpy(np.pad(img.numpy(), pad_size, mode=self.padding_mode))
            gt = torch.from_numpy(np.pad(gt.numpy(), pad_size, mode=self.padding_mode))

            return (img.float() / 255, gt.float() / 255)

    def __len__(self):
        return self.images.size(0)

    def __getitem__(self, idx):
        return self.transform(self.images[idx], self.ground_truth[idx])

    def getAll(self, shuffle=True):
        if (shuffle):
            indices = torch.randperm(len(self.selection))
            return self.transform(self.images[indices], self.ground_truth[indices])
        else:
            return self.transform(self.images, self.ground_truth)