transform.py 9.02 KB
Newer Older
1
import random
2
3
import math
import torch
eellison's avatar
eellison committed
4
from torch import nn, Tensor
5
import torchvision
eellison's avatar
eellison committed
6
from torch.jit.annotations import List, Tuple, Dict, Optional
7
8
9
10
11
12
13

from torchvision.ops import misc as misc_nn_ops
from .image_list import ImageList
from .roi_heads import paste_masks_in_image


class GeneralizedRCNNTransform(nn.Module):
14
15
16
17
18
19
20
21
22
23
24
    """
    Performs input / target transformation before feeding the data to a GeneralizedRCNN
    model.

    The transformations it perform are:
        - input normalization (mean subtraction and std division)
        - input / target resizing to match min_size / max_size

    It returns a ImageList for the inputs, and a List[Dict[Tensor]] for the targets
    """

25
26
    def __init__(self, min_size, max_size, image_mean, image_std):
        super(GeneralizedRCNNTransform, self).__init__()
27
28
29
30
        if not isinstance(min_size, (list, tuple)):
            min_size = (min_size,)
        self.min_size = min_size
        self.max_size = max_size
31
32
33
34
        self.image_mean = image_mean
        self.image_std = image_std

    def forward(self, images, targets=None):
eellison's avatar
eellison committed
35
        # type: (List[Tensor], Optional[List[Dict[str, Tensor]]])
36
        images = [img for img in images]
37
38
        for i in range(len(images)):
            image = images[i]
eellison's avatar
eellison committed
39
40
            target_index = targets[i] if targets is not None else None

41
42
43
44
            if image.dim() != 3:
                raise ValueError("images is expected to be a list of 3d tensors "
                                 "of shape [C, H, W], got {}".format(image.shape))
            image = self.normalize(image)
eellison's avatar
eellison committed
45
            image, target_index = self.resize(image, target_index)
46
            images[i] = image
eellison's avatar
eellison committed
47
48
            if targets is not None and target_index is not None:
                targets[i] = target_index
49

50
51
        image_sizes = [img.shape[-2:] for img in images]
        images = self.batch_images(images)
eellison's avatar
eellison committed
52
53
54
55
56
57
        image_sizes_list = torch.jit.annotate(List[Tuple[int, int]], [])
        for image_size in image_sizes:
            assert len(image_size) == 2
            image_sizes_list.append((image_size[0], image_size[1]))

        image_list = ImageList(images, image_sizes_list)
58
59
60
61
62
63
64
65
        return image_list, targets

    def normalize(self, image):
        dtype, device = image.dtype, image.device
        mean = torch.as_tensor(self.image_mean, dtype=dtype, device=device)
        std = torch.as_tensor(self.image_std, dtype=dtype, device=device)
        return (image - mean[:, None, None]) / std[:, None, None]

eellison's avatar
eellison committed
66
67
68
69
70
71
72
73
74
75
    def torch_choice(self, l):
        # type: (List[int])
        """
        Implements `random.choice` via torch ops so it can be compiled with
        TorchScript. Remove if https://github.com/pytorch/pytorch/issues/25803
        is fixed.
        """
        index = int(torch.empty(1).uniform_(0., float(len(l))).item())
        return l[index]

76
    def resize(self, image, target):
eellison's avatar
eellison committed
77
        # type: (Tensor, Optional[Dict[str, Tensor]])
78
        h, w = image.shape[-2:]
79
80
81
        im_shape = torch.tensor(image.shape[-2:])
        min_size = float(torch.min(im_shape))
        max_size = float(torch.max(im_shape))
82
        if self.training:
eellison's avatar
eellison committed
83
            size = float(self.torch_choice(self.min_size))
84
85
        else:
            # FIXME assume for now that testing uses the largest scale
eellison's avatar
eellison committed
86
            size = float(self.min_size[-1])
87
        scale_factor = size / min_size
88
89
90
        if max_size * scale_factor > self.max_size:
            scale_factor = self.max_size / max_size
        image = torch.nn.functional.interpolate(
91
92
            image[None], scale_factor=scale_factor, mode='bilinear',
            align_corners=False)[0]
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111

        if target is None:
            return image, target

        bbox = target["boxes"]
        bbox = resize_boxes(bbox, (h, w), image.shape[-2:])
        target["boxes"] = bbox

        if "masks" in target:
            mask = target["masks"]
            mask = misc_nn_ops.interpolate(mask[None].float(), scale_factor=scale_factor)[0].byte()
            target["masks"] = mask

        if "keypoints" in target:
            keypoints = target["keypoints"]
            keypoints = resize_keypoints(keypoints, (h, w), image.shape[-2:])
            target["keypoints"] = keypoints
        return image, target

112
113
    # _onnx_batch_images() is an implementation of
    # batch_images() that is supported by ONNX tracing.
eellison's avatar
eellison committed
114
    @torch.jit.unused
115
    def _onnx_batch_images(self, images, size_divisible=32):
eellison's avatar
eellison committed
116
        # type: (List[Tensor], int) -> Tensor
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
        max_size = []
        for i in range(images[0].dim()):
            max_size_i = torch.max(torch.stack([img.shape[i] for img in images]).to(torch.float32)).to(torch.int64)
            max_size.append(max_size_i)
        stride = size_divisible
        max_size[1] = (torch.ceil((max_size[1].to(torch.float32)) / stride) * stride).to(torch.int64)
        max_size[2] = (torch.ceil((max_size[2].to(torch.float32)) / stride) * stride).to(torch.int64)
        max_size = tuple(max_size)

        # work around for
        # pad_img[: img.shape[0], : img.shape[1], : img.shape[2]].copy_(img)
        # which is not yet supported in onnx
        padded_imgs = []
        for img in images:
            padding = [(s1 - s2) for s1, s2 in zip(max_size, tuple(img.shape))]
132
            padded_img = torch.nn.functional.pad(img, (0, padding[2], 0, padding[1], 0, padding[0]))
133
134
135
136
            padded_imgs.append(padded_img)

        return torch.stack(padded_imgs)

eellison's avatar
eellison committed
137
138
139
140
141
142
143
144
    def max_by_axis(self, the_list):
        # type: (List[List[int]]) -> List[int]
        maxes = the_list[0]
        for sublist in the_list[1:]:
            for index, item in enumerate(sublist):
                maxes[index] = max(maxes[index], item)
        return maxes

145
    def batch_images(self, images, size_divisible=32):
eellison's avatar
eellison committed
146
        # type: (List[Tensor], int)
147
148
149
150
        if torchvision._is_tracing():
            # batch_images() does not export well to ONNX
            # call _onnx_batch_images() instead
            return self._onnx_batch_images(images, size_divisible)
151

eellison's avatar
eellison committed
152
153
        max_size = self.max_by_axis([list(img.shape) for img in images])
        stride = float(size_divisible)
154
        max_size = list(max_size)
155
156
        max_size[1] = int(math.ceil(float(max_size[1]) / stride) * stride)
        max_size[2] = int(math.ceil(float(max_size[2]) / stride) * stride)
157

eellison's avatar
eellison committed
158
159
        batch_shape = [len(images)] + max_size
        batched_imgs = images[0].new_full(batch_shape, 0)
160
161
162
163
164
165
        for img, pad_img in zip(images, batched_imgs):
            pad_img[: img.shape[0], : img.shape[1], : img.shape[2]].copy_(img)

        return batched_imgs

    def postprocess(self, result, image_shapes, original_image_sizes):
eellison's avatar
eellison committed
166
        # type: (List[Dict[str, Tensor]], List[Tuple[int, int]], List[Tuple[int, int]])
167
168
169
170
171
172
        if self.training:
            return result
        for i, (pred, im_s, o_im_s) in enumerate(zip(result, image_shapes, original_image_sizes)):
            boxes = pred["boxes"]
            boxes = resize_boxes(boxes, im_s, o_im_s)
            result[i]["boxes"] = boxes
173
174
            if "masks" in pred:
                masks = pred["masks"]
175
                masks = paste_masks_in_image(masks, boxes, o_im_s)
176
                result[i]["masks"] = masks
177
178
179
180
181
182
            if "keypoints" in pred:
                keypoints = pred["keypoints"]
                keypoints = resize_keypoints(keypoints, im_s, o_im_s)
                result[i]["keypoints"] = keypoints
        return result

183
184
185
186
187
188
189
190
191
    def __repr__(self):
        format_string = self.__class__.__name__ + '('
        _indent = '\n    '
        format_string += "{0}Normalize(mean={1}, std={2})".format(_indent, self.image_mean, self.image_std)
        format_string += "{0}Resize(min_size={1}, max_size={2}, mode='bilinear')".format(_indent, self.min_size,
                                                                                         self.max_size)
        format_string += '\n)'
        return format_string

192
193

def resize_keypoints(keypoints, original_size, new_size):
eellison's avatar
eellison committed
194
    # type: (Tensor, List[int], List[int])
195
196
197
198
199
    ratios = [
        torch.tensor(s, dtype=torch.float32, device=keypoints.device) /
        torch.tensor(s_orig, dtype=torch.float32, device=keypoints.device)
        for s, s_orig in zip(new_size, original_size)
    ]
200
201
    ratio_h, ratio_w = ratios
    resized_data = keypoints.clone()
202
203
204
205
206
207
208
    if torch._C._get_tracing_state():
        resized_data_0 = resized_data[:, :, 0] * ratio_w
        resized_data_1 = resized_data[:, :, 1] * ratio_h
        resized_data = torch.stack((resized_data_0, resized_data_1, resized_data[:, :, 2]), dim=2)
    else:
        resized_data[..., 0] *= ratio_w
        resized_data[..., 1] *= ratio_h
209
210
211
212
    return resized_data


def resize_boxes(boxes, original_size, new_size):
eellison's avatar
eellison committed
213
    # type: (Tensor, List[int], List[int])
214
215
216
217
218
    ratios = [
        torch.tensor(s, dtype=torch.float32, device=boxes.device) /
        torch.tensor(s_orig, dtype=torch.float32, device=boxes.device)
        for s, s_orig in zip(new_size, original_size)
    ]
219
220
    ratio_height, ratio_width = ratios
    xmin, ymin, xmax, ymax = boxes.unbind(1)
221

222
223
224
225
226
    xmin = xmin * ratio_width
    xmax = xmax * ratio_width
    ymin = ymin * ratio_height
    ymax = ymax * ratio_height
    return torch.stack((xmin, ymin, xmax, ymax), dim=1)