File size: 7,055 Bytes
e1b51e5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
print("Importing standard...")
from abc import ABC, abstractmethod

print("Importing external...")
import torch
from torch.nn.functional import binary_cross_entropy

# from matplotlib import pyplot as plt

print("Importing internal...")
from utils import preprocess_masks_features, get_row_col, symlog, calculate_iou


######### BINARY LOSSES ###############
def my_lovasz_hinge(logits, gt, downsample=False):
    if downsample:
        offset = int(torch.randint(downsample - 1, (1,)))
        logits, gt = logits[:, offset::downsample], gt[:, offset::downsample]
        # B, HW
    gt = 1.0 * gt  # go float
    areas = gt.sum(dim=1, keepdims=True)  # B, 1
    # per_image = True, ignore = None
    signs = 2 * gt - 1
    errors = 1 - logits * signs
    errors_sorted, perm = torch.sort(errors, dim=1, descending=True)
    gt_sorted = torch.gather(gt, 1, perm)  # B, HW
    # lovasz grad
    intersection = areas - gt_sorted.cumsum(dim=1)  # B, HW
    union = areas + (1 - gt_sorted).cumsum(dim=1)  # B, HW
    jaccard = 1 - intersection / union  # B, HW
    jaccard[:, 1:] = jaccard[:, 1:] - jaccard[:, :-1]
    loss = (torch.relu(errors_sorted) * jaccard).sum(dim=1)  # B,
    return torch.nanmean(loss)


def focal_loss(scores, targets, alpha=0.25, gamma=2):
    p = scores
    ce_loss = binary_cross_entropy(p, targets, reduction="none")
    p_t = p * targets + (1 - p) * (1 - targets)
    loss = ce_loss * ((1 - p_t) ** gamma)

    if alpha >= 0:
        alpha_t = alpha * targets + (1 - alpha) * (1 - targets)
        loss = alpha_t * loss

    return loss


# also binary_cross_entropy and lovasz


########## SUBFUNCTIONS ######################3
def get_distances(features, refs, sigma, norm_p, square_distances, H, W):
    # features: B, 1, F, HW
    # refs: B, M, F, 1
    # sigma: B, M, 1, 1
    B, M = refs.shape[0], refs.shape[1]
    distances = torch.norm(
        features - refs, dim=2, p=norm_p, keepdim=True
    )  # B, M, 1, H*W
    distances = distances**2 if square_distances else distances
    distances = (distances / (2 * sigma**2)).reshape(B, M, H * W)
    return distances


def activate(features, masks, activation, use_sigma, offset_pos, ret_prediction):
    # sigmoid is very similar to exp
    # prepare features
    assert activation in ["sigmoid", "symlog"]
    if masks is None:  # when inferencing
        B, M = 1, 1
        F, N = sorted(features.shape)
        H, W = [int(N ** (0.5))] * 2
        features = features.reshape(1, 1, -1, H * W)
    else:
        masks, features, M, B, H, W, F = preprocess_masks_features(masks, features)
    # features: B, 1, F, H*W
    # masks: B, M, 1, H*W
    if use_sigma:
        sigma = torch.nn.functional.softplus(features)[:, :, -1:]  # B, 1, 1, H*W
        features = features[:, :, :-1]
        F = features.shape[2]
    else:
        sigma = 1
    features = symlog(features) if activation == "symlog" else torch.sigmoid(features)
    if offset_pos:
        assert F >= 2
        row, col = get_row_col(H, W, features.device)
        row = row.reshape(1, 1, 1, H, 1).expand(B, 1, 1, H, W).reshape(B, 1, 1, H * W)
        col = col.reshape(1, 1, 1, 1, W).expand(B, 1, 1, H, W).reshape(B, 1, 1, H * W)
        positional_features = torch.cat([row, col], dim=2)  # B, 1, 2, H*W
        features[:, :, :2] = features[:, :, :2] + positional_features
    prediction = features.reshape(B, 1, -1, H, W) if ret_prediction else None
    if masks is None:
        features = features.reshape(-1, H * W)
        sigma = sigma.reshape(-1, H * W) if use_sigma else 1
        return features, sigma, H, W
    return features, masks, sigma, prediction, B, M, F, H, W


class AbstractLoss(ABC):
    @staticmethod
    @abstractmethod
    def loss(features, masks, ret_prediction=False, **kwargs):
        pass

    @staticmethod
    @abstractmethod
    def get_mask_from_query(features, sindex, **kwargs):
        pass


class IISLoss(AbstractLoss):
    @staticmethod
    def loss(features, masks, ret_prediction=False, K=3, logger=None):
        features, masks, sigma, prediction, B, M, F, H, W = activate(
            features, masks, "symlog", False, False, ret_prediction
        )
        rindices = torch.randperm(H * W, device=masks.device)
        # the following should work if all masks have more than K pixels
        sindices = torch.stack(
            [
                torch.stack([rindices[masks[b, m, 0, rindices]][:K] for m in range(M)])
                for b in range(B)
            ]
        )  # B, M, K
        feats_at_sindices = torch.gather(
            features.permute(0, 3, 1, 2).expand(B, H * W, K, F),
            dim=1,
            index=sindices.reshape(B, M, K, 1).expand(B, M, K, F),
        )  # B, M, K, F
        feats_at_sindices = feats_at_sindices.reshape(B, M, K, F, 1)  # B, M, K, F, 1
        dists = get_distances(
            features, feats_at_sindices.reshape(B, M * K, F, 1), sigma, 2, True, H, W
        )
        score = torch.exp(-dists)  # B, M*K, H*W [0, 1]
        targets = (
            masks.expand(B, M, K, H * W).reshape(B, M * K, H * W).float()
        )  # B, M, K, H*W
        floss = focal_loss(score, targets).mean()
        lloss = my_lovasz_hinge(
            score.view(B * M * K, H * W) * 2 - 1,
            targets.view(B * M * K, H * W),
        )
        loss = floss + lloss
        return loss, prediction

    @staticmethod
    def get_mask_from_query(features, sindex):
        features, _, H, W = activate(features, None, "symlog", False, False, False)
        F = features.shape[0]
        query_feat = features[:, sindex]
        dists = get_distances(
            features.reshape(1, 1, F, H * W),
            query_feat.reshape(1, 1, F, 1),
            1,
            2,
            True,
            H,
            W,
        )
        score = torch.exp(-dists)  # 1, H*W
        pred = score > 0.5
        return pred


def iis_iou(features, masks, get_mask_from_query, K=20):
    masks, features, M, B, H, W, F = preprocess_masks_features(masks, features)
    # features: B, 1, F, H*W
    # masks: B, M, 1, H*W
    rindices = torch.randperm(H * W).to(masks.device)
    sindices = torch.stack(
        [
            torch.stack([rindices[masks[b, m, 0, rindices]][:K] for m in range(M)])
            for b in range(B)
        ]
    )  # B, M, K
    cum_iou, n_samples = 0, 0
    for b in range(B):
        for m in range(M):
            for k in range(K):
                sindex = sindices[b, m, k]
                pred = get_mask_from_query(features[b, 0], sindex)
                iou = calculate_iou(pred, masks[b, m, 0, :])
                cum_iou += iou
                n_samples += 1

    return cum_iou / n_samples


losses_names = [
    "iis",
]
#


def get_loss_class(loss_name):
    if loss_name == "iis":
        return IISLoss
    else:
        raise NotImplementedError


def get_get_mask_from_query(loss_name):
    loss_class = get_loss_class(loss_name)
    return loss_class.get_mask_from_query


def get_loss(loss_name):
    loss_class = get_loss_class(loss_name)
    return loss_class.loss