File size: 6,584 Bytes
d670799
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) OpenMMLab. All rights reserved.
from typing import Any, Dict, List, Tuple

import torch
from mmengine.dist import all_gather, get_rank
from mmengine.model import BaseModel
from mmengine.structures import InstanceData

from mmaction.registry import MODELS
from mmaction.utils import ForwardResults, OptSampleList


class GatherLayer(torch.autograd.Function):
    """Gather tensors from all process, supporting backward propagation."""

    @staticmethod
    def forward(ctx: Any, input: torch.Tensor) -> Tuple[List]:
        ctx.save_for_backward(input)
        output = all_gather(input)
        return tuple(output)

    @staticmethod
    def backward(ctx: Any, *grads: torch.Tensor) -> torch.Tensor:
        input, = ctx.saved_tensors
        grad_out = torch.zeros_like(input)
        grad_out[:] = grads[get_rank()]
        return grad_out


@MODELS.register_module()
class CLIPSimilarity(BaseModel):
    """CLIP-based similarity model.



    Args:

        clip_arch (str): The architecture of the clip model.

            Supported choices are `'ViT-B/32'`, `'ViT-B/16'`,

            `'ViT-L/14'` and `'ViT-L/14@336px'`.

        data_preprocessor (dict): The pre-process config.

        adapter (dict): The 3D adapter config.

        to_float32 (bool): Whether to convert the dtype of params of clip

            model to float32.

        frozen_layers: Layers to be frozen (all params fixed). -1 means

            not freezing any parameters. Defaults to -1.

        loss (dict): The config of loss. Defaults to

            `dict(type='CrossEntropyLoss', loss_weight=0.5)`.

    """

    def __init__(

        self,

        clip_arch: str,

        data_preprocessor: Dict[str, Dict],

        adapter: Dict,

        to_float32: bool = False,

        frozen_layers: int = -1,

        loss: Dict = dict(type='CrossEntropyLoss', loss_weight=0.5)

    ) -> None:
        super(CLIPSimilarity,
              self).__init__(data_preprocessor=data_preprocessor)

        try:
            import clip
        except ImportError:
            raise ImportError('Please run `pip install '
                              'git+https://github.com/openai/CLIP.git` '
                              'to install clip first. ')

        self.clip = clip.load(clip_arch, device='cpu')[0]
        if to_float32:
            self.clip.float()
        self.loss = MODELS.build(loss)
        self.adapter = MODELS.build(adapter)
        self.frozen_layers = frozen_layers
        self._freeze_stages()

    def encode_video(self, video: torch.Tensor) -> torch.Tensor:
        """Encode video."""
        b, n, c, h, w = video.shape
        video = video.view(-1, c, h, w)
        frames_features = self.encode_image(video)
        frames_features = frames_features.view(b, n, -1)
        video_features = self.adapter(frames_features)
        return video_features

    def encode_image(self, image: torch.Tensor) -> torch.Tensor:
        """Encode image."""
        return self.clip.encode_image(image)

    def encode_text(self, text: torch.Tensor) -> torch.Tensor:
        """Encode text."""
        return self.clip.encode_text(text)

    def extract_feat(self,

                     inputs: Dict[str, torch.Tensor],

                     norm: bool = True) -> Tuple:
        """Extract features."""
        text_inputs = inputs['text']
        video_inputs = inputs['imgs']
        text_features = self.encode_text(text_inputs)
        video_features = self.encode_video(video_inputs)

        if norm:
            text_features = text_features / text_features.norm(
                dim=-1, keepdim=True)
            video_features = video_features / video_features.norm(
                dim=-1, keepdim=True)

        return video_features, text_features

    def forward(self,

                inputs: Dict[str, torch.Tensor],

                data_samples: OptSampleList = None,

                mode: str = 'tensor') -> ForwardResults:
        """Forward function."""

        if mode == 'tensor':
            return self.extract_feat(inputs, norm=False)

        elif mode == 'loss':
            video_features, text_features = self.extract_feat(inputs)
            video_features = torch.cat(
                GatherLayer.apply(video_features), dim=0)
            text_features = torch.cat(GatherLayer.apply(text_features), dim=0)

            logit_scale = self.clip.logit_scale.exp()
            logits_per_video = logit_scale * video_features @ text_features.t()
            logits_per_text = logits_per_video.t()

            labels = torch.arange(logits_per_video.shape[0]).to(
                logit_scale.device)

            sim_loss_v2t = self.loss(logits_per_video, labels)
            sim_loss_t2v = self.loss(logits_per_text, labels)

            losses = dict()
            losses['sim_loss_v2t'] = sim_loss_v2t
            losses['sim_loss_t2v'] = sim_loss_t2v
            return losses

        elif mode == 'predict':
            video_features, text_features = self.extract_feat(inputs)
            for ds, vf, tf in zip(data_samples, video_features, text_features):
                features = InstanceData(video_feature=vf, text_feature=tf)
                ds.features = features
            return data_samples

        else:
            raise RuntimeError(f'Invalid mode "{mode}". '
                               'Only supports loss, predict and tensor mode')

    def train(self, mode: bool = True) -> None:
        """Set the optimization status when training."""
        super().train(mode)
        self._freeze_stages()

    def _freeze_stages(self) -> None:
        """Prevent all the parameters from being optimized before

        ``self.frozen_layers``."""

        if self.frozen_layers >= 0:
            top_layers = [
                'ln_final', 'text_projection', 'logit_scale', 'visual.ln_post',
                'visual.proj'
            ]
            mid_layers = [
                'visual.transformer.resblocks', 'transformer.resblocks'
            ]

            for name, param in self.clip.named_parameters():
                if any(name.find(n) == 0 for n in top_layers):
                    continue
                elif any(name.find(n) == 0 for n in mid_layers):
                    layer_n = int(name.split('.resblocks.')[1].split('.')[0])
                    if layer_n >= self.frozen_layers:
                        continue
                param.requires_grad = False