File size: 7,327 Bytes
4d0eb62
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) OpenMMLab. All rights reserved.
from typing import List, Optional, Union

import numpy as np
import torch
import torch.distributed as dist
import torch.nn as nn
from mmengine.dist import all_reduce
from mmengine.model import BaseModule

from mmpretrain.registry import MODELS


@torch.no_grad()
def distributed_sinkhorn(out: torch.Tensor, sinkhorn_iterations: int,
                         world_size: int, epsilon: float) -> torch.Tensor:
    """Apply the distributed sinknorn optimization on the scores matrix to find
    the assignments.

    This function is modified from
    https://github.com/facebookresearch/swav/blob/main/main_swav.py

    Args:
        out (torch.Tensor): The scores matrix
        sinkhorn_iterations (int): Number of iterations in Sinkhorn-Knopp
            algorithm.
        world_size (int): The world size of the process group.
        epsilon (float): regularization parameter for Sinkhorn-Knopp algorithm.

    Returns:
        torch.Tensor: Output of sinkhorn algorithm.
    """
    eps_num_stab = 1e-12
    Q = torch.exp(out / epsilon).t(
    )  # Q is K-by-B for consistency with notations from our paper
    B = Q.shape[1] * world_size  # number of samples to assign
    K = Q.shape[0]  # how many prototypes

    # make the matrix sums to 1
    sum_Q = torch.sum(Q)
    all_reduce(sum_Q)
    Q /= sum_Q

    for it in range(sinkhorn_iterations):
        # normalize each row: total weight per prototype must be 1/K
        u = torch.sum(Q, dim=1, keepdim=True)
        if len(torch.nonzero(u == 0)) > 0:
            Q += eps_num_stab
            u = torch.sum(Q, dim=1, keepdim=True, dtype=Q.dtype)
            all_reduce(u)
        Q /= u
        Q /= K

        # normalize each column: total weight per sample must be 1/B
        Q /= torch.sum(Q, dim=0, keepdim=True)
        Q /= B

    Q *= B  # the columns must sum to 1 so that Q is an assignment
    return Q.t()


class MultiPrototypes(BaseModule):
    """Multi-prototypes for SwAV head.

    Args:
        output_dim (int): The output dim from SwAV neck.
        num_prototypes (List[int]): The number of prototypes needed.
        init_cfg (dict or List[dict], optional): Initialization config dict.
            Defaults to None.
    """

    def __init__(self,
                 output_dim: int,
                 num_prototypes: List[int],
                 init_cfg: Optional[Union[List[dict], dict]] = None) -> None:
        super().__init__(init_cfg=init_cfg)
        assert isinstance(num_prototypes, list)
        self.num_heads = len(num_prototypes)
        for i, k in enumerate(num_prototypes):
            self.add_module('prototypes' + str(i),
                            nn.Linear(output_dim, k, bias=False))

    def forward(self, x: torch.Tensor) -> List[torch.Tensor]:
        """Run forward for every prototype."""
        out = []
        for i in range(self.num_heads):
            out.append(getattr(self, 'prototypes' + str(i))(x))
        return out


@MODELS.register_module()
class SwAVLoss(BaseModule):
    """The Loss for SwAV.

    This Loss contains clustering and sinkhorn algorithms to compute Q codes.
    Part of the code is borrowed from `script
    <https://github.com/facebookresearch/swav>`_.
    The queue is built in `engine/hooks/swav_hook.py`.

    Args:
        feat_dim (int): feature dimension of the prototypes.
        sinkhorn_iterations (int): number of iterations in Sinkhorn-Knopp
            algorithm. Defaults to 3.
        epsilon (float): regularization parameter for Sinkhorn-Knopp algorithm.
            Defaults to 0.05.
        temperature (float): temperature parameter in training loss.
            Defaults to 0.1.
        crops_for_assign (List[int]): list of crops id used for computing
            assignments. Defaults to [0, 1].
        num_crops (List[int]): list of number of crops. Defaults to [2].
        num_prototypes (int): number of prototypes. Defaults to 3000.
        init_cfg (dict or List[dict], optional): Initialization config dict.
            Defaults to None.
    """

    def __init__(self,
                 feat_dim: int,
                 sinkhorn_iterations: int = 3,
                 epsilon: float = 0.05,
                 temperature: float = 0.1,
                 crops_for_assign: List[int] = [0, 1],
                 num_crops: List[int] = [2],
                 num_prototypes: int = 3000,
                 init_cfg: Optional[Union[List[dict], dict]] = None):
        super().__init__(init_cfg=init_cfg)
        self.sinkhorn_iterations = sinkhorn_iterations
        self.epsilon = epsilon
        self.temperature = temperature
        self.crops_for_assign = crops_for_assign
        self.num_crops = num_crops
        self.use_queue = False
        self.queue = None
        self.world_size = dist.get_world_size() if dist.is_initialized() else 1

        # prototype layer
        self.prototypes = None
        if isinstance(num_prototypes, list):
            self.prototypes = MultiPrototypes(feat_dim, num_prototypes)
        elif num_prototypes > 0:
            self.prototypes = nn.Linear(feat_dim, num_prototypes, bias=False)
        assert self.prototypes is not None

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """Forward function of SwAV loss.

        Args:
            x (torch.Tensor): NxC input features.
        Returns:
            torch.Tensor: The returned loss.
        """
        # normalize the prototypes
        with torch.no_grad():
            w = self.prototypes.weight.data.clone()
            w = nn.functional.normalize(w, dim=1, p=2)
            self.prototypes.weight.copy_(w)

        embedding, output = x, self.prototypes(x)
        embedding = embedding.detach()

        bs = int(embedding.size(0) / sum(self.num_crops))
        loss = 0
        for i, crop_id in enumerate(self.crops_for_assign):
            with torch.no_grad():
                out = output[bs * crop_id:bs * (crop_id + 1)].detach()
                # time to use the queue
                if self.queue is not None:
                    if self.use_queue or not torch.all(self.queue[i,
                                                                  -1, :] == 0):
                        self.use_queue = True
                        out = torch.cat(
                            (torch.mm(self.queue[i],
                                      self.prototypes.weight.t()), out))
                    # fill the queue
                    self.queue[i, bs:] = self.queue[i, :-bs].clone()
                    self.queue[i, :bs] = embedding[crop_id * bs:(crop_id + 1) *
                                                   bs]

                # get assignments (batch_size * num_prototypes)
                q = distributed_sinkhorn(out, self.sinkhorn_iterations,
                                         self.world_size, self.epsilon)[-bs:]

            # cluster assignment prediction
            subloss = 0
            for v in np.delete(np.arange(np.sum(self.num_crops)), crop_id):
                x = output[bs * v:bs * (v + 1)] / self.temperature
                subloss -= torch.mean(
                    torch.sum(q * nn.functional.log_softmax(x, dim=1), dim=1))
            loss += subloss / (np.sum(self.num_crops) - 1)
        loss /= len(self.crops_for_assign)
        return loss