File size: 10,417 Bytes
684943d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
#
# Copyright (C) 2023, Inria
# GRAPHDECO research group, https://team.inria.fr/graphdeco
# All rights reserved.
#
# This software is free for non-commercial, research and evaluation use 
# under the terms of the LICENSE.md file.
#
# For inquiries contact  [email protected]
#

from math import exp

import numpy as np
import torch
import torch.nn.functional as F
from torch.autograd import Variable


def l1_loss(network_output, gt):
    return torch.abs((network_output - gt)).mean()


def l2_loss(network_output, gt):
    return ((network_output - gt) ** 2).mean()


def cos_loss(network_output, gt):
    return 1 - F.cosine_similarity(network_output, gt, dim=0).mean()


def gaussian(window_size, sigma):
    gauss = torch.Tensor([exp(-(x - window_size // 2) ** 2 / float(2 * sigma ** 2)) for x in range(window_size)])
    return gauss / gauss.sum()


def create_window(window_size, channel):
    _1D_window = gaussian(window_size, 1.5).unsqueeze(1)
    _2D_window = _1D_window.mm(_1D_window.t()).float().unsqueeze(0).unsqueeze(0)
    window = Variable(_2D_window.expand(channel, 1, window_size, window_size).contiguous())
    return window


def ssim(img1, img2, window_size=11, size_average=True):
    channel = img1.size(-3)
    window = create_window(window_size, channel)

    if img1.is_cuda:
        window = window.cuda(img1.get_device())
    window = window.type_as(img1)

    return _ssim(img1, img2, window, window_size, channel, size_average)


def _ssim(img1, img2, window, window_size, channel, size_average=True):
    mu1 = F.conv2d(img1, window, padding=window_size // 2, groups=channel)
    mu2 = F.conv2d(img2, window, padding=window_size // 2, groups=channel)

    mu1_sq = mu1.pow(2)
    mu2_sq = mu2.pow(2)
    mu1_mu2 = mu1 * mu2

    sigma1_sq = F.conv2d(img1 * img1, window, padding=window_size // 2, groups=channel) - mu1_sq
    sigma2_sq = F.conv2d(img2 * img2, window, padding=window_size // 2, groups=channel) - mu2_sq
    sigma12 = F.conv2d(img1 * img2, window, padding=window_size // 2, groups=channel) - mu1_mu2

    C1 = 0.01 ** 2
    C2 = 0.03 ** 2

    ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) / ((mu1_sq + mu2_sq + C1) * (sigma1_sq + sigma2_sq + C2))

    if size_average:
        return ssim_map.mean()
    else:
        return ssim_map.mean(1).mean(1).mean(1)


def ssim2(img1, img2, window_size=11):
    channel = img1.size(-3)
    window = create_window(window_size, channel)

    if img1.is_cuda:
        window = window.cuda(img1.get_device())
    window = window.type_as(img1)

    mu1 = F.conv2d(img1, window, padding=window_size // 2, groups=channel)
    mu2 = F.conv2d(img2, window, padding=window_size // 2, groups=channel)

    mu1_sq = mu1.pow(2)
    mu2_sq = mu2.pow(2)
    mu1_mu2 = mu1 * mu2

    sigma1_sq = F.conv2d(img1 * img1, window, padding=window_size // 2, groups=channel) - mu1_sq
    sigma2_sq = F.conv2d(img2 * img2, window, padding=window_size // 2, groups=channel) - mu2_sq
    sigma12 = F.conv2d(img1 * img2, window, padding=window_size // 2, groups=channel) - mu1_mu2

    C1 = 0.01 ** 2
    C2 = 0.03 ** 2

    ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) / ((mu1_sq + mu2_sq + C1) * (sigma1_sq + sigma2_sq + C2))

    return ssim_map.mean(0)


def get_img_grad_weight(img, beta=2.0):
    _, hd, wd = img.shape
    bottom_point = img[..., 2:hd, 1:wd - 1]
    top_point = img[..., 0:hd - 2, 1:wd - 1]
    right_point = img[..., 1:hd - 1, 2:wd]
    left_point = img[..., 1:hd - 1, 0:wd - 2]
    grad_img_x = torch.mean(torch.abs(right_point - left_point), 0, keepdim=True)
    grad_img_y = torch.mean(torch.abs(top_point - bottom_point), 0, keepdim=True)
    grad_img = torch.cat((grad_img_x, grad_img_y), dim=0)
    grad_img, _ = torch.max(grad_img, dim=0)
    grad_img = (grad_img - grad_img.min()) / (grad_img.max() - grad_img.min())
    grad_img = torch.nn.functional.pad(grad_img[None, None], (1, 1, 1, 1), mode='constant', value=1.0).squeeze()
    return grad_img


def lncc(ref, nea):
    # ref_gray: [batch_size, total_patch_size]
    # nea_grays: [batch_size, total_patch_size]
    bs, tps = nea.shape
    patch_size = int(np.sqrt(tps))

    ref_nea = ref * nea
    ref_nea = ref_nea.view(bs, 1, patch_size, patch_size)
    ref = ref.view(bs, 1, patch_size, patch_size)
    nea = nea.view(bs, 1, patch_size, patch_size)
    ref2 = ref.pow(2)
    nea2 = nea.pow(2)

    # sum over kernel
    filters = torch.ones(1, 1, patch_size, patch_size, device=ref.device)
    padding = patch_size // 2
    ref_sum = F.conv2d(ref, filters, stride=1, padding=padding)[:, :, padding, padding]
    nea_sum = F.conv2d(nea, filters, stride=1, padding=padding)[:, :, padding, padding]
    ref2_sum = F.conv2d(ref2, filters, stride=1, padding=padding)[:, :, padding, padding]
    nea2_sum = F.conv2d(nea2, filters, stride=1, padding=padding)[:, :, padding, padding]
    ref_nea_sum = F.conv2d(ref_nea, filters, stride=1, padding=padding)[:, :, padding, padding]

    # average over kernel
    ref_avg = ref_sum / tps
    nea_avg = nea_sum / tps

    cross = ref_nea_sum - nea_avg * ref_sum
    ref_var = ref2_sum - ref_avg * ref_sum
    nea_var = nea2_sum - nea_avg * nea_sum

    cc = cross * cross / (ref_var * nea_var + 1e-8)
    ncc = 1 - cc
    ncc = torch.clamp(ncc, 0.0, 2.0)
    ncc = torch.mean(ncc, dim=1, keepdim=True)
    mask = (ncc < 0.9)
    return ncc, mask


def loss_cls_3d(features, predictions, k=5, lambda_val=2.0, max_points=200000, sample_size=800):
    # Randomly downsample
    if features.size(0) > max_points:
        indices = torch.randperm(features.size(0))[:max_points]
        features = features[indices]
        predictions = predictions[indices]
    # Normalize predictions to [0, 1] range
    min_value = predictions.min()
    max_value = predictions.max()
    if max_value > min_value:
        predictions = (predictions - min_value) / (max_value - min_value)

    # Randomly sample points for which we'll compute the loss
    indices = torch.randperm(features.size(0))[:sample_size]
    sample_features = features[indices]
    sample_preds = predictions[indices]

    # Compute top-k nearest neighbors directly in PyTorch
    dists = torch.cdist(sample_features, features)  # Compute pairwise distances
    _, neighbor_indices_tensor = dists.topk(k, largest=False)  # Get top-k smallest distances

    # Fetch neighbor predictions using indexing
    neighbor_preds = predictions[neighbor_indices_tensor]

    # Compute KL divergence
    kl = sample_preds.unsqueeze(1) * (torch.log(sample_preds.unsqueeze(1) + 1e-10) - torch.log(neighbor_preds + 1e-10))
    loss = torch.abs(kl).mean()

    return lambda_val * loss


def get_loss_semantic_group(gt_seg, language_feature, num=10000):
    # Randomly select num indices from gt_seg
    if gt_seg.size(0) < num:
        indices = torch.randperm(gt_seg.size(0))
        num = gt_seg.size(0)
    else:
        indices = torch.randperm(gt_seg.size(0))[:num]
    input_id1 = input_id2 = gt_seg[indices]
    language_feature = language_feature[indices]

    # Expand labels, create masks for valid positive pairs, excluding self-pairs.
    labels1_expanded = input_id1.unsqueeze(1).expand(-1, input_id1.shape[0])
    labels2_expanded = input_id2.unsqueeze(0).expand(input_id2.shape[0], -1)
    mask_full_positive = labels1_expanded == labels2_expanded
    block_mask = torch.ones(num, num, dtype=torch.bool, device=gt_seg.device)
    block_mask = torch.triu(block_mask, diagonal=0)
    diag_mask = torch.eye(block_mask.shape[0], device=gt_seg.device, dtype=torch.bool)
    # Compute semantic loss for positive pairs
    total_loss = 0
    mask = torch.where(mask_full_positive * block_mask * (~diag_mask))
    semantic_loss = torch.norm(
        language_feature[mask[0]] - language_feature[mask[1]], p=2, dim=-1
    ).nansum()
    total_loss += semantic_loss
    total_loss = total_loss / torch.sum(block_mask).float()
    return 2 * total_loss


def get_loss_instance_group(sam_seg, instance_feature, language_feature, num=1000):
    # Randomly select num indices from gt_seg
    margin = 1.0
    if sam_seg.size(0) < num:
        indices = torch.randperm(sam_seg.size(0))
        num = sam_seg.size(0)
    else:
        indices = torch.randperm(sam_seg.size(0))[:num]
    instance_feature = instance_feature[indices]
    input_id1 = input_id2 = sam_seg[indices]
    language_feature = language_feature[indices]

    # Expand labels, create masks for valid positive pairs, excluding self-pairs.
    labels1_expanded = input_id1.unsqueeze(1).expand(-1, input_id1.shape[0])
    labels2_expanded = input_id2.unsqueeze(0).expand(input_id2.shape[0], -1)
    mask_full_positive = labels1_expanded == labels2_expanded
    mask_full_negative = ~mask_full_positive
    block_mask = torch.ones(num, num, dtype=torch.bool, device=sam_seg.device)
    block_mask = torch.triu(block_mask, diagonal=0)
    diag_mask = torch.eye(block_mask.shape[0], device=sam_seg.device, dtype=torch.bool)

    # Compute instance loss for positive pairs
    total_loss = 0
    mask = torch.where(mask_full_positive * block_mask * (~diag_mask))
    instance_loss_1 = torch.norm(
        instance_feature[mask[0]] - instance_feature[mask[1]], p=2, dim=-1
    ).nansum()
    total_loss += instance_loss_1

    # Create mask for negative pairs and compute language similarity using cosine similarity
    mask = torch.where(mask_full_negative * block_mask)
    language_similarity = torch.nn.functional.cosine_similarity(
        language_feature[mask[0]], language_feature[mask[1]], dim=-1
    )

    # Compute instance loss for negative pairs with margin and language similarity
    instance_loss_2 = (
            torch.relu(
                margin - torch.norm(instance_feature[mask[0]] - instance_feature[mask[1]], p=2, dim=-1)
            ) * (1 + language_similarity)
    ).nansum()
    total_loss += instance_loss_2
    total_loss = total_loss / torch.sum(block_mask).float()
    return 2 * total_loss

def ranking_loss(error, penalize_ratio=1.0, type="mean"):
    sorted_error, _ = torch.sort(error.flatten(), descending=True)
    k = int(penalize_ratio * len(sorted_error))
    if k == 0:
        return torch.tensor(0.0, device=error.device)
    selected_error = sorted_error[:k]
    if type == "mean":
        return torch.mean(selected_error)
    elif type == "sum":
        return torch.sum(selected_error)
    else:
        raise ValueError(f"Unsupported type: {type}")