File size: 8,632 Bytes
5df2892
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright 2024 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================

import torch
import einops
import torch.nn.functional as F


def get_mask(idx, array):
    '''

    array: b m, records # of elements to be masked

    '''
    b, m = array.shape
    n = idx.size(-1)
    A = torch.arange(n, dtype=idx.dtype, device=idx.device).unsqueeze(0).unsqueeze(0).expand(b, m, n)  # 1 1 n -> b m n
    mask = A < array.unsqueeze(-1)
    return mask


def alloc(var, rest, budget, tp, maximum, times=0, fast=False):
    '''

    var: (b m) variance of each pixel POSITIVE VALUE

    rest: (b m) list of already allocated budgets

    budget: (b) remaining to be allocated

    tp: mean type, plain/softmax

    maximum: maximum budget for each pixel

    '''
    b, m = var.shape
    if tp == 'plain':
        var_p = var * (rest < maximum)
        var_sum = var_p.sum(dim=-1, keepdim=True)  # b 1
        proportion = var_p / var_sum  # b m
    elif tp == 'softmax':
        var_p = var.clone()
        var_p[rest >= maximum] = -float('inf')  # maximum
        proportion = torch.nn.functional.softmax(var_p, dim=-1)  # b m
    allocation = torch.round(proportion * budget.unsqueeze(1))  # b m
    new_rest = torch.clamp(rest + allocation, 0, maximum)  # b m
    remain_budget = budget - (new_rest - rest).sum(dim=-1)  # b m allocated
    negative_remain = (remain_budget < 0)
    while negative_remain.sum() > 0:
        offset = torch.eye(m, device=rest.device)[
            torch.randint(m, (negative_remain.sum().int().item(),), device=rest.device)]
        new_rest[negative_remain] = torch.clamp(new_rest[negative_remain] - offset, 1, maximum)  # reduce by one

        # update remain budget
        remain_budget = budget - (new_rest - rest).sum(dim=-1)  # b m allocated
        negative_remain = (remain_budget < 0)

    if (remain_budget > 0).sum() > 0:
        if times < 3:
            new_rest[remain_budget > 0] = alloc(var[remain_budget > 0], new_rest[remain_budget > 0],
                                                remain_budget[remain_budget > 0], tp, maximum, times + 1, fast=fast)
        elif not fast:  # precise budget allocation
            positive_remain = (remain_budget > 0)
            while positive_remain.sum() > 0:
                offset = torch.eye(m, device=rest.device)[
                    torch.randint(m, (positive_remain.sum().int().item(),), device=rest.device)]
                new_rest[positive_remain] = torch.clamp(new_rest[positive_remain] + offset, 1, maximum)  # add by one
                # update remain budget
                remain_budget = budget - (new_rest - rest).sum(dim=-1)  # b m allocated
                positive_remain = (remain_budget > 0)
    return new_rest


def flex(D_: torch.Tensor, X: torch.Tensor, idx: torch.Tensor, flex_type, topk_, current_iter, total_iters, X_diff,

         fast=False, return_maskarray=False):
    '''

    D: (b m n) Gram matrix, sorted on last dim, descending

    X: (b numh numw he) c (sh sw) X_data

    idx: (b m n) sorted index of D

    x_size: (h, w) 2-tuple tensor

    OUT: (b m n) Binary mask

    '''
    b, m, n = D_.shape
    if flex_type is None or flex_type == 'none':
        mask_array = topk_ * torch.ones((b, m), dtype=torch.int, device=D_.device)

    elif flex_type == 'gsort':
        D = D_.clone()
        D -= (D == D.max(dim=-1, keepdim=True)) * 100000  # neglect max position
        val, g_idx = torch.sort(D.view(b, -1), dim=-1, descending=True)  # global sort
        # g_idx: (b m*n)
        g_idx += m * n * torch.arange(b, dtype=g_idx.dtype, device=g_idx.device).unsqueeze(-1)  # b 1
        non_topk_idx = g_idx[:, topk_ * (m - 1):]  # select top k, neglect max

        mask_ = torch.ones_like(D).bool()
        mask_.view(-1)[non_topk_idx.reshape(-1)] = False  # set to negative value
        mask_array = mask_.sum(dim=-1)
        mask_array += 1  # include max, ensure each pixel has at least one match

    elif flex_type == 'interdiff_plain':  # interpolate and diff

        rest = torch.ones_like(X_diff)
        budget = torch.ones(b, dtype=torch.int, device=idx.device) * (topk_ - 1) * idx.size(1)
        mask_array = alloc(X_diff, rest, budget, tp='plain', maximum=idx.size(-1), fast=fast)
    else:
        raise NotImplementedError(f'Graph type {flex_type} not implemented...')

    if return_maskarray:
        return mask_array

    mask = ~get_mask(idx, mask_array)  # negated

    return mask


def cossim(X_sample, Y_sample, graph=None):
    if graph is not None:
        return torch.einsum('a b m c, a b n c -> a b m n', F.normalize(X_sample, dim=-1),
                            F.normalize(Y_sample, dim=-1)) + (-100.) * (~graph)
    return torch.einsum('a b m c, a b n c -> a b m n', F.normalize(X_sample, dim=-1), F.normalize(Y_sample, dim=-1))


def local_sampling(x, group_size, unfold_dict, output=0, tp='bhwc'):
    '''

        output:

        x (grouped) [B, nn, c]

        x_unfold [B, NN, C]

        0/1/2: grouped, sampled, both

    '''
    if isinstance(group_size, int):
        group_size = (group_size, group_size)

    if output != 1:
        if tp == 'bhwc':
            x_grouped = einops.rearrange(x, 'b (numh sh) (numw sw) c-> (b numh numw) (sh sw) c', sh=group_size[0],
                                         sw=group_size[1])
        elif tp == 'bchw':
            x_grouped = einops.rearrange(x, 'b c (numh sh) (numw sw)-> (b numh numw) (sh sw) c', sh=group_size[0],
                                         sw=group_size[1])

        if output == 0:
            return x_grouped

    if tp == 'bhwc':
        x = einops.rearrange(x, 'b h w c -> b c h w')

    x_sampled = einops.rearrange(F.unfold(x, **unfold_dict), 'b (c k0 k1) l -> (b l) (k0 k1) c',
                                 k0=unfold_dict['kernel_size'][0], k1=unfold_dict['kernel_size'][1])

    if output == 1:
        return x_sampled

    assert x_grouped.size(0) == x_sampled.size(0)
    return x_grouped, x_sampled


def global_sampling(x, group_size, sample_size, output=0, tp='bhwc'):
    '''

        output:

        x (grouped) [B, nn, c]

        x_unfold [B, NN, C]

    '''
    if isinstance(group_size, int):
        group_size = (group_size, group_size)
    if isinstance(sample_size, int):
        sample_size = (sample_size, sample_size)

    if output != 1:
        if tp == 'bchw':
            x_grouped = einops.rearrange(x, 'b c (sh numh) (sw numw) -> (b numh numw) (sh sw) c', sh=group_size[0],
                                         sw=group_size[1])
        elif tp == 'bhwc':
            x_grouped = einops.rearrange(x, 'b (sh numh) (sw numw) c -> (b numh numw) (sh sw) c', sh=group_size[0],
                                         sw=group_size[1])

        if output == 0:
            return x_grouped

    if tp == 'bchw':
        x_sampled = einops.rearrange(x, 'b c (sh extrah numh) (sw extraw numw) -> b extrah numh extraw numw c sh sw',
                                     sh=sample_size[0], sw=sample_size[1], extrah=1, extraw=1)
    elif tp == 'bhwc':
        x_sampled = einops.rearrange(x, 'b (sh extrah numh) (sw extraw numw) c -> b extrah numh extraw numw c sh sw',
                                     sh=sample_size[0], sw=sample_size[1], extrah=1, extraw=1)
    b_y, _, numh, _, numw, c_y, sh_y, sw_y = x_sampled.shape
    ratio_h, ratio_w = sample_size[0] // group_size[0], sample_size[1] // group_size[1]
    x_sampled = x_sampled.expand(b_y, ratio_h, numh, ratio_w, numw, c_y, sh_y, sw_y).reshape(-1, c_y,
                                                                                             sh_y * sw_y).permute(0, 2,
                                                                                                                  1)

    if output == 1:
        return x_sampled

    assert x_grouped.size(0) == x_sampled.size(0)
    return x_grouped, x_sampled