File size: 8,778 Bytes
9ddc7da
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""
================================================
@author: Jaron
@time: 2024/02/20 16:21:56
@email: fjjth98@163.com
@description: QFormer projector, convert image and video into fixed-length tokens
================================================
"""

import math
import torch
import torch.nn as nn
from torch.nn.functional import interpolate
from transformers.models.blip_2.modeling_blip_2 import Blip2QFormerModel, Blip2QFormerEncoder

from .configuration_ccam_projector import CCAMConfig


class SimpleQFormerOutput(nn.Module):
    # replace last residual MLP with normal MLP
    def __init__(self, config):
        super().__init__()
        self.dense = nn.Linear(config.intermediate_size, config.output_size)

    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor = None) -> torch.Tensor:
        return self.dense(hidden_states)


class SimpleQFormerIdentity(nn.Module):
    # just to replace the first attention module with identity, since it is useless
 
    def forward(self, hidden_states: torch.Tensor, *args, **kwargs) -> torch.Tensor:
        return hidden_states,


class CCAMModel(Blip2QFormerModel):
    _auto_class = 'AutoModel'
    config_class = CCAMConfig
    base_model_prefix = 'model'
    supports_gradient_checkpointing = True
 
    def __init__(self, config: CCAMConfig):
        super(Blip2QFormerModel, self).__init__(config)
        self.gradient_checkpointing = False
        self.config = config
        self.num_query_tokens = config.num_query_tokens
        self.visual_attn_mask_type = config.visual_attn_mask_type

        self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
        self.dropout = nn.Dropout(config.hidden_dropout_prob)
        self.encoder = Blip2QFormerEncoder(config)
        self.encoder.layer[0].attention = SimpleQFormerIdentity()   # replace the 1st attention module with identity
        self.encoder.layer[-1].output_query = SimpleQFormerOutput(config)

        # initialize query tokens
        self.query_tokens = nn.Parameter(torch.zeros(1, config.num_query_tokens, config.hidden_size))

        # initialize pos embed
        self.spatial_pos_embed = self._create_pos_embed(*config.spatial_resolution, type=config.spatial_pos_embed_type)     # (H, W, C)
        self.temporal_pos_embed = self._create_pos_embed(config.temporal_resolution, type=config.temporal_pos_embed_type)   # (T, C)

        # initialize query attn mask
        if config.query_attn_mask_type == 'full':
            self.query_attn_mask = None
        elif config.query_attn_mask_type == 'causal':
            query_attn_mask = torch.ones(self.num_query_tokens, self.num_query_tokens)
            q = torch.arange(self.num_query_tokens)
            query_attn_mask.masked_fill_(q > q[:, None], 0)
            self.query_attn_mask = query_attn_mask[None]
        else:
            raise NotImplementedError(f'Do not support {self.query_attn_mask} query_attn_mask')

        self.post_init()

    def _create_pos_embed(self, *size: int, type: str = 'none') -> torch.Tensor:
        C = self.config.encoder_hidden_size
        if type == 'none':
            pos_embed = None
        elif type == 'learnable':
            pos_embed = nn.Parameter(.02 * torch.randn(*size, C))
        elif type == 'cosine':
            total_len = 1
            for i in size:
                total_len *= i
            raw = torch.outer(torch.arange(total_len), torch.exp(torch.arange(0, C, 2) * (-math.log(10000.) / C)))
            pos_embed = nn.Parameter(torch.stack((raw.sin(), raw.cos()), dim=-1).view(*size, C), requires_grad=False)
        else:
            raise NotImplementedError(f'Do not support {type} position embeddings')
        return pos_embed

    def get_attn_mask(self, embeddings: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
        """Get visual_attn_mask and query_attn_mask if needed 
            embeddings (torch.Tensor): (B, T, L, C)
        """
        B, T, L, _ = embeddings.size()
        device = embeddings.device

        # visual attn mask only work for videos
        if T > 1:
            if self.visual_attn_mask_type == 'ccam':
                base_attn_mask = torch.ones(T, T, device=device)
                t = torch.arange(T, device=device)
                base_attn_mask.masked_fill_(t > t[:, None], 0)
                visual_attn_mask = torch.cat((
                    torch.kron(
                        base_attn_mask,
                        torch.ones(self.num_query_tokens // T, L, device=device)
                    ),
                    torch.ones(self.num_query_tokens % T, T * L, device=device)
                ), dim=0)[None].expand(B, -1, -1)
            elif self.visual_attn_mask_type == 'full':
                visual_attn_mask = None
            else:
                raise NotImplementedError(f'Do not support {self.visual_attn_mask_type} attn_mask')
        else:
            visual_attn_mask = None

        if self.query_attn_mask is None:
            query_attn_mask = None
        else:
            query_attn_mask = self.query_attn_mask.expand(B, -1, -1)

        return visual_attn_mask, query_attn_mask

    def batch_forward_no_spatial(self, visual_embeds: torch.Tensor) -> torch.Tensor:
        """Batch forward without spatial mask position embeddings

        Args:
            visual_embeds (torch.Tensor): (B, T, L, C)

        Returns:
            torch.Tensor: (B, Q, C)
        """
        B, T, _, C = visual_embeds.size()
        query_embeds = self.query_tokens.expand(B, -1, -1)
        visual_attn_mask, query_attn_mask = self.get_attn_mask(visual_embeds)

        # add temporal position embeddings
        if self.temporal_pos_embed is not None:
            if T == self.temporal_pos_embed.size(0):
                pos_embed = self.temporal_pos_embed
            elif T == 1:
                pos_embed = 0. * self.temporal_pos_embed[:1]     # for deepspeed 
            else:
                pos_embed = interpolate(
                    self.temporal_pos_embed.T[None],    # (1, C, t)
                    size=(T,),
                    mode='linear',
                    align_corners=False
                )[0].T      # (T, C)
            visual_embeds = visual_embeds + pos_embed.view(1, T, 1, C)
        visual_embeds = visual_embeds.flatten(1, 2)

        return super().forward(
            query_embeds=query_embeds,
            attention_mask=query_attn_mask,
            encoder_hidden_states=visual_embeds,
            encoder_attention_mask=visual_attn_mask
        )[0]

    def forward(self, visual_embeds: torch.Tensor, split_sizes: list[int], unmasked_ids: torch.LongTensor = None):
        """
            visual_embeds (torch.Tensor): (T, L, C)
            split_sizes (list[int]): [t0, t1, ...] sum_i ti=T
            unmasked_ids (torch.LongTensor): If provided, should be in the shape of (T, L) whose value v 0<=v<=HW-1
            output_attentions (_type_, optional): _description_. Defaults to None.
            output_hidden_states (_type_, optional): _description_. Defaults to None.
            return_dict (_type_, optional): _description_. Defaults to None.
        """
        _, L, C = visual_embeds.size()

        # add spatial position embeddings
        if self.spatial_pos_embed is not None:
            pos_embed = self.spatial_pos_embed.view(-1, C)  # (H*W, C)
            if unmasked_ids is None:
                pos_embed = pos_embed.view(1, L, C)     # if not provided, L must equals to H*W
            else:
                pos_embed = pos_embed[unmasked_ids]     # (T, L, C)
            visual_embeds = visual_embeds + pos_embed

        # all inputs in this batch has the same t
        if len(set(split_sizes)) == 1:
            visual_embeds = visual_embeds.view(len(split_sizes), split_sizes[0], L, C)
            output = self.batch_forward_no_spatial(visual_embeds)
        else:
            visual_embeds = visual_embeds.split(split_sizes, dim=0)
            # group visual_embeds accoding to the number of frames
            output, group_visual_embeds = [None] * len(split_sizes), {}
            for idx, (embed, t) in enumerate(zip(visual_embeds, split_sizes)):
                if t in group_visual_embeds:
                    group_visual_embeds[t][0].append(idx)
                    group_visual_embeds[t][1].append(embed)
                else:
                    group_visual_embeds[t] = [[idx], [embed]]
            for idx, embeds in group_visual_embeds.values():
                cur_output = self.batch_forward_no_spatial(torch.stack(embeds, dim=0))
                for i, j in enumerate(idx):
                    output[j] = cur_output[i]
            output = torch.stack(output, dim=0)

        return output