BigGANsAreWatching/postprocessing.py at master · RyougiJarvis/BigGANsAreWatching · 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
import numpy as np
import torch
from skimage.measure import label
from torchvision.transforms import ToPILImage, ToTensor, Resize


def resize(x, target_shape):
    x = ToPILImage()(x.cpu().to(torch.float32))
    x = Resize(target_shape)(x)
    x = ToTensor()(x)
    return x.cuda()


def resize_min_edge(x, size):
    img_shape = x.shape[-2:]
    if img_shape[0] > img_shape[1]:
        x = resize((x[0] + 1.0) / 2.0, (size * img_shape[0] // img_shape[1], size))
    else:
        x = resize((x[0] + 1.0) / 2.0, (size, size * img_shape[1] // img_shape[0]))

    return x.unsqueeze(0) * 2.0 - 1.0


class SegmentationInference(object):
    def __init__(self, model=None, resize_to=None):
        self.model = model
        self.resize_to = resize_to

    @torch.no_grad()
    def __call__(self, img, mask=None):
        return self.apply(img, mask=None)

    @torch.no_grad()
    def apply(self, img, mask=None):
        img_shape = img.shape[-2:]
        if mask is None:
            if self.model is not None:
                if self.resize_to is not None:
                    img = resize_min_edge(img, self.resize_to)
                mask = self.model(img)
            else:
                raise Exception(
                    'Eithr both (img, mask) should be provided or self.model is not None')

        if len(mask.shape) == 4:
            mask = (1.0 - torch.softmax(mask, dim=1))[:, 0]

        if self.resize_to is not None and mask.shape[-2:] != img_shape:
            mask = resize(mask, img_shape)

        return mask


class Threshold(SegmentationInference):
    def __init__(self, model=None, thr=0.5, resize_to=None):
        super(Threshold, self).__init__(model, resize_to)
        self.thr = thr

    @torch.no_grad()
    def __call__(self, img, mask=None):
        mask = self.apply(img, mask)
        return mask >= self.thr


def connected_components_filter(*args):
    mask = args[-1].cpu().numpy()
    for i in range(len(mask)):
        component, num = label(mask[i], return_num=True, background=0)

        stats = np.zeros([num + 1])
        for comp in range(1, num + 1, 1):
            stats[comp] = np.sum(component == comp)

        max_component = np.argmax(stats)
        max_component_area = stats[max_component]

        mask[i] *= 0
        for comp in range(1, num + 1, 1):
            area = stats[comp]
            if float(area) / max_component_area > 0.2:
                mask[i][component == comp] = True

    return torch.from_numpy(mask).cuda()