File size: 10,071 Bytes
2366e36
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
# Copyright (c) OpenMMLab. All rights reserved.
import torch
import torch.nn.functional as F
from mmdet.core import BitmapMasks
from torch import nn

from mmocr.models.builder import LOSSES
from mmocr.utils import check_argument


@LOSSES.register_module()
class DRRGLoss(nn.Module):
    """The class for implementing DRRG loss. This is partially adapted from
    https://github.com/GXYM/DRRG licensed under the MIT license.

    DRRG: `Deep Relational Reasoning Graph Network for Arbitrary Shape Text
    Detection <https://arxiv.org/abs/1908.05900>`_.

    Args:
        ohem_ratio (float): The negative/positive ratio in ohem.
    """

    def __init__(self, ohem_ratio=3.0):
        super().__init__()
        self.ohem_ratio = ohem_ratio

    def balance_bce_loss(self, pred, gt, mask):
        """Balanced Binary-CrossEntropy Loss.

        Args:
            pred (Tensor): Shape of :math:`(1, H, W)`.
            gt (Tensor): Shape of :math:`(1, H, W)`.
            mask (Tensor): Shape of :math:`(1, H, W)`.

        Returns:
            Tensor: Balanced bce loss.
        """
        assert pred.shape == gt.shape == mask.shape
        assert torch.all(pred >= 0) and torch.all(pred <= 1)
        assert torch.all(gt >= 0) and torch.all(gt <= 1)
        positive = gt * mask
        negative = (1 - gt) * mask
        positive_count = int(positive.float().sum())
        gt = gt.float()
        if positive_count > 0:
            loss = F.binary_cross_entropy(pred, gt, reduction='none')
            positive_loss = torch.sum(loss * positive.float())
            negative_loss = loss * negative.float()
            negative_count = min(
                int(negative.float().sum()),
                int(positive_count * self.ohem_ratio))
        else:
            positive_loss = torch.tensor(0.0, device=pred.device)
            loss = F.binary_cross_entropy(pred, gt, reduction='none')
            negative_loss = loss * negative.float()
            negative_count = 100
        negative_loss, _ = torch.topk(negative_loss.view(-1), negative_count)

        balance_loss = (positive_loss + torch.sum(negative_loss)) / (
            float(positive_count + negative_count) + 1e-5)

        return balance_loss

    def gcn_loss(self, gcn_data):
        """CrossEntropy Loss from gcn module.

        Args:
            gcn_data (tuple(Tensor, Tensor)): The first is the
                prediction with shape :math:`(N, 2)` and the
                second is the gt label with shape :math:`(m, n)`
                where :math:`m * n = N`.

        Returns:
            Tensor: CrossEntropy loss.
        """
        gcn_pred, gt_labels = gcn_data
        gt_labels = gt_labels.view(-1).to(gcn_pred.device)
        loss = F.cross_entropy(gcn_pred, gt_labels)

        return loss

    def bitmasks2tensor(self, bitmasks, target_sz):
        """Convert Bitmasks to tensor.

        Args:
            bitmasks (list[BitmapMasks]): The BitmapMasks list. Each item is
                for one img.
            target_sz (tuple(int, int)): The target tensor of size
                :math:`(H, W)`.

        Returns:
            list[Tensor]: The list of kernel tensors. Each element stands for
            one kernel level.
        """
        assert check_argument.is_type_list(bitmasks, BitmapMasks)
        assert isinstance(target_sz, tuple)

        batch_size = len(bitmasks)
        num_masks = len(bitmasks[0])

        results = []

        for level_inx in range(num_masks):
            kernel = []
            for batch_inx in range(batch_size):
                mask = torch.from_numpy(bitmasks[batch_inx].masks[level_inx])
                # hxw
                mask_sz = mask.shape
                # left, right, top, bottom
                pad = [
                    0, target_sz[1] - mask_sz[1], 0, target_sz[0] - mask_sz[0]
                ]
                mask = F.pad(mask, pad, mode='constant', value=0)
                kernel.append(mask)
            kernel = torch.stack(kernel)
            results.append(kernel)

        return results

    def forward(self, preds, downsample_ratio, gt_text_mask,
                gt_center_region_mask, gt_mask, gt_top_height_map,
                gt_bot_height_map, gt_sin_map, gt_cos_map):
        """Compute Drrg loss.

        Args:
            preds (tuple(Tensor)): The first is the prediction map
                with shape :math:`(N, C_{out}, H, W)`.
                The second is prediction from GCN module, with
                shape :math:`(N, 2)`.
                The third is ground-truth label with shape :math:`(N, 8)`.
            downsample_ratio (float): The downsample ratio.
            gt_text_mask (list[BitmapMasks]): Text mask.
            gt_center_region_mask (list[BitmapMasks]): Center region mask.
            gt_mask (list[BitmapMasks]): Effective mask.
            gt_top_height_map (list[BitmapMasks]): Top height map.
            gt_bot_height_map (list[BitmapMasks]): Bottom height map.
            gt_sin_map (list[BitmapMasks]): Sinusoid map.
            gt_cos_map (list[BitmapMasks]): Cosine map.

        Returns:
            dict:  A loss dict with ``loss_text``, ``loss_center``,
            ``loss_height``, ``loss_sin``, ``loss_cos``, and ``loss_gcn``.
        """
        assert isinstance(preds, tuple)
        assert isinstance(downsample_ratio, float)
        assert check_argument.is_type_list(gt_text_mask, BitmapMasks)
        assert check_argument.is_type_list(gt_center_region_mask, BitmapMasks)
        assert check_argument.is_type_list(gt_mask, BitmapMasks)
        assert check_argument.is_type_list(gt_top_height_map, BitmapMasks)
        assert check_argument.is_type_list(gt_bot_height_map, BitmapMasks)
        assert check_argument.is_type_list(gt_sin_map, BitmapMasks)
        assert check_argument.is_type_list(gt_cos_map, BitmapMasks)

        pred_maps, gcn_data = preds
        pred_text_region = pred_maps[:, 0, :, :]
        pred_center_region = pred_maps[:, 1, :, :]
        pred_sin_map = pred_maps[:, 2, :, :]
        pred_cos_map = pred_maps[:, 3, :, :]
        pred_top_height_map = pred_maps[:, 4, :, :]
        pred_bot_height_map = pred_maps[:, 5, :, :]
        feature_sz = pred_maps.size()
        device = pred_maps.device

        # bitmask 2 tensor
        mapping = {
            'gt_text_mask': gt_text_mask,
            'gt_center_region_mask': gt_center_region_mask,
            'gt_mask': gt_mask,
            'gt_top_height_map': gt_top_height_map,
            'gt_bot_height_map': gt_bot_height_map,
            'gt_sin_map': gt_sin_map,
            'gt_cos_map': gt_cos_map
        }
        gt = {}
        for key, value in mapping.items():
            gt[key] = value
            if abs(downsample_ratio - 1.0) < 1e-2:
                gt[key] = self.bitmasks2tensor(gt[key], feature_sz[2:])
            else:
                gt[key] = [item.rescale(downsample_ratio) for item in gt[key]]
                gt[key] = self.bitmasks2tensor(gt[key], feature_sz[2:])
                if key in ['gt_top_height_map', 'gt_bot_height_map']:
                    gt[key] = [item * downsample_ratio for item in gt[key]]
            gt[key] = [item.to(device) for item in gt[key]]

        scale = torch.sqrt(1.0 / (pred_sin_map**2 + pred_cos_map**2 + 1e-8))
        pred_sin_map = pred_sin_map * scale
        pred_cos_map = pred_cos_map * scale

        loss_text = self.balance_bce_loss(
            torch.sigmoid(pred_text_region), gt['gt_text_mask'][0],
            gt['gt_mask'][0])

        text_mask = (gt['gt_text_mask'][0] * gt['gt_mask'][0]).float()
        negative_text_mask = ((1 - gt['gt_text_mask'][0]) *
                              gt['gt_mask'][0]).float()
        loss_center_map = F.binary_cross_entropy(
            torch.sigmoid(pred_center_region),
            gt['gt_center_region_mask'][0].float(),
            reduction='none')
        if int(text_mask.sum()) > 0:
            loss_center_positive = torch.sum(
                loss_center_map * text_mask) / torch.sum(text_mask)
        else:
            loss_center_positive = torch.tensor(0.0, device=device)
        loss_center_negative = torch.sum(
            loss_center_map *
            negative_text_mask) / torch.sum(negative_text_mask)
        loss_center = loss_center_positive + 0.5 * loss_center_negative

        center_mask = (gt['gt_center_region_mask'][0] *
                       gt['gt_mask'][0]).float()
        if int(center_mask.sum()) > 0:
            map_sz = pred_top_height_map.size()
            ones = torch.ones(map_sz, dtype=torch.float, device=device)
            loss_top = F.smooth_l1_loss(
                pred_top_height_map / (gt['gt_top_height_map'][0] + 1e-2),
                ones,
                reduction='none')
            loss_bot = F.smooth_l1_loss(
                pred_bot_height_map / (gt['gt_bot_height_map'][0] + 1e-2),
                ones,
                reduction='none')
            gt_height = (
                gt['gt_top_height_map'][0] + gt['gt_bot_height_map'][0])
            loss_height = torch.sum(
                (torch.log(gt_height + 1) *
                 (loss_top + loss_bot)) * center_mask) / torch.sum(center_mask)

            loss_sin = torch.sum(
                F.smooth_l1_loss(
                    pred_sin_map, gt['gt_sin_map'][0], reduction='none') *
                center_mask) / torch.sum(center_mask)
            loss_cos = torch.sum(
                F.smooth_l1_loss(
                    pred_cos_map, gt['gt_cos_map'][0], reduction='none') *
                center_mask) / torch.sum(center_mask)
        else:
            loss_height = torch.tensor(0.0, device=device)
            loss_sin = torch.tensor(0.0, device=device)
            loss_cos = torch.tensor(0.0, device=device)

        loss_gcn = self.gcn_loss(gcn_data)

        results = dict(
            loss_text=loss_text,
            loss_center=loss_center,
            loss_height=loss_height,
            loss_sin=loss_sin,
            loss_cos=loss_cos,
            loss_gcn=loss_gcn)

        return results