File size: 8,209 Bytes
cc0dd3c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) OpenMMLab. All rights reserved.
import math
from functools import reduce
from operator import mul
from typing import Dict, List, Optional, Union

import torch
import torch.nn as nn
from torch.nn.modules.batchnorm import _BatchNorm

from mmpretrain.models.backbones import VisionTransformer
from mmpretrain.models.utils import (build_2d_sincos_position_embedding,
                                     to_2tuple)
from mmpretrain.registry import MODELS
from mmpretrain.structures import DataSample
from ..utils import CosineEMA
from .base import BaseSelfSupervisor


@MODELS.register_module()
class MoCoV3ViT(VisionTransformer):
    """Vision Transformer for MoCoV3 pre-training.

    A pytorch implement of: `An Images is Worth 16x16 Words: Transformers for
    Image Recognition at Scale <https://arxiv.org/abs/2010.11929>`_.

    Part of the code is modified from:
    `<https://github.com/facebookresearch/moco-v3/blob/main/vits.py>`_.

    Args:
        stop_grad_conv1 (bool): whether to stop the gradient of
            convolution layer in `PatchEmbed`. Defaults to False.
        frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
            -1 means not freezing any parameters. Defaults to -1.
        norm_eval (bool): Whether to set norm layers to eval mode, namely,
            freeze running stats (mean and var). Note: Effect on Batch Norm
            and its variants only. Defaults to False.
        init_cfg (dict or list[dict], optional): Initialization config dict.
            Defaults to None.
    """

    def __init__(self,
                 stop_grad_conv1: bool = False,
                 frozen_stages: int = -1,
                 norm_eval: bool = False,
                 init_cfg: Optional[Union[dict, List[dict]]] = None,
                 **kwargs) -> None:

        # add MoCoV3 ViT-small arch
        self.arch_zoo.update(
            dict.fromkeys(
                ['mocov3-s', 'mocov3-small'], {
                    'embed_dims': 384,
                    'num_layers': 12,
                    'num_heads': 12,
                    'feedforward_channels': 1536,
                }))

        super().__init__(init_cfg=init_cfg, **kwargs)
        self.patch_size = kwargs['patch_size']
        self.frozen_stages = frozen_stages
        self.norm_eval = norm_eval
        self.init_cfg = init_cfg

        if stop_grad_conv1:
            self.patch_embed.projection.weight.requires_grad = False
            self.patch_embed.projection.bias.requires_grad = False

        self._freeze_stages()

    def init_weights(self) -> None:
        """Initialize position embedding, patch embedding, qkv layers and cls
        token."""
        super().init_weights()

        if not (isinstance(self.init_cfg, dict)
                and self.init_cfg['type'] == 'Pretrained'):

            # Use fixed 2D sin-cos position embedding
            pos_emb = build_2d_sincos_position_embedding(
                patches_resolution=self.patch_resolution,
                embed_dims=self.embed_dims,
                cls_token=True)
            self.pos_embed.data.copy_(pos_emb)
            self.pos_embed.requires_grad = False

            # xavier_uniform initialization for PatchEmbed
            val = math.sqrt(
                6. / float(3 * reduce(mul, to_2tuple(self.patch_size), 1) +
                           self.embed_dims))
            nn.init.uniform_(self.patch_embed.projection.weight, -val, val)
            nn.init.zeros_(self.patch_embed.projection.bias)

            # initialization for linear layers
            for name, m in self.named_modules():
                if isinstance(m, nn.Linear):
                    if 'qkv' in name:
                        # treat the weights of Q, K, V separately
                        val = math.sqrt(
                            6. /
                            float(m.weight.shape[0] // 3 + m.weight.shape[1]))
                        nn.init.uniform_(m.weight, -val, val)
                    else:
                        nn.init.xavier_uniform_(m.weight)
                    nn.init.zeros_(m.bias)
            nn.init.normal_(self.cls_token, std=1e-6)

    def _freeze_stages(self) -> None:
        """Freeze patch_embed layer, some parameters and stages."""
        if self.frozen_stages >= 0:
            self.patch_embed.eval()
            for param in self.patch_embed.parameters():
                param.requires_grad = False

            self.cls_token.requires_grad = False
            self.pos_embed.requires_grad = False

        for i in range(1, self.frozen_stages + 1):
            m = self.layers[i - 1]
            m.eval()
            for param in m.parameters():
                param.requires_grad = False

            if i == (self.num_layers) and self.final_norm:
                for param in getattr(self, 'norm1').parameters():
                    param.requires_grad = False

    def train(self, mode: bool = True) -> None:
        super().train(mode)
        self._freeze_stages()
        if mode and self.norm_eval:
            for m in self.modules():
                # trick: eval have effect on BatchNorm only
                if isinstance(m, _BatchNorm):
                    m.eval()


@MODELS.register_module()
class MoCoV3(BaseSelfSupervisor):
    """MoCo v3.

    Implementation of `An Empirical Study of Training Self-Supervised Vision
    Transformers <https://arxiv.org/abs/2104.02057>`_.

    Args:
        backbone (dict): Config dict for module of backbone
        neck (dict): Config dict for module of deep features to compact feature
            vectors.
        head (dict): Config dict for module of head functions.
        base_momentum (float): Momentum coefficient for the momentum-updated
            encoder. Defaults to 0.01.
        pretrained (str, optional): The pretrained checkpoint path, support
            local path and remote path. Defaults to None.
        data_preprocessor (dict, optional): The config for preprocessing
            input data. If None or no specified type, it will use
            "SelfSupDataPreprocessor" as type.
            See :class:`SelfSupDataPreprocessor` for more details.
            Defaults to None.
        init_cfg (Union[List[dict], dict], optional): Config dict for weight
            initialization. Defaults to None.
    """

    def __init__(self,
                 backbone: dict,
                 neck: dict,
                 head: dict,
                 base_momentum: float = 0.01,
                 pretrained: Optional[str] = None,
                 data_preprocessor: Optional[dict] = None,
                 init_cfg: Optional[Union[List[dict], dict]] = None) -> None:
        super().__init__(
            backbone=backbone,
            neck=neck,
            head=head,
            pretrained=pretrained,
            data_preprocessor=data_preprocessor,
            init_cfg=init_cfg)

        # create momentum model
        self.momentum_encoder = CosineEMA(
            nn.Sequential(self.backbone, self.neck), momentum=base_momentum)

    def loss(self, inputs: List[torch.Tensor], data_samples: List[DataSample],
             **kwargs) -> Dict[str, torch.Tensor]:
        """The forward function in training.

        Args:
            inputs (List[torch.Tensor]): The input images.
            data_samples (List[DataSample]): All elements required
                during the forward function.

        Returns:
            Dict[str, torch.Tensor]: A dictionary of loss components.
        """
        assert isinstance(inputs, list)
        view_1 = inputs[0]
        view_2 = inputs[1]

        # compute query features, [N, C] each
        q1 = self.neck(self.backbone(view_1))[0]
        q2 = self.neck(self.backbone(view_2))[0]

        # compute key features, [N, C] each, no gradient
        with torch.no_grad():
            # update momentum encoder
            self.momentum_encoder.update_parameters(
                nn.Sequential(self.backbone, self.neck))

            k1 = self.momentum_encoder(view_1)[0]
            k2 = self.momentum_encoder(view_2)[0]

        loss = self.head.loss(q1, k2) + self.head.loss(q2, k1)

        losses = dict(loss=loss)
        return losses