File size: 4,726 Bytes
b90b5f6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import re


class IdentityMap(nn.Module):
    def __init__(self):
        super().__init__()

    def forward(self, x, *args, **kwargs):
        return x

    @property
    def config(self):
        return {"mm_projector_type": "identity"}


class SimpleResBlock(nn.Module):
    def __init__(self, channels):
        super().__init__()
        self.pre_norm = nn.LayerNorm(channels)

        self.proj = nn.Sequential(
            nn.Linear(channels, channels), nn.GELU(), nn.Linear(channels, channels)
        )

    def forward(self, x):
        x = self.pre_norm(x)
        return x + self.proj(x)

class MlpGeLUPostCrossAttProjector(nn.Module):
    def __init__(self, mlp_gelu_cross_att_match, config):
        super(MlpGeLUPostCrossAttProjector, self).__init__()

        mlp_depth = int(mlp_gelu_cross_att_match.group(1))
        modules = [nn.Linear(config.mm_hidden_size, config.hidden_size)]
        for _ in range(1, mlp_depth):
            modules.append(nn.GELU())
            modules.append(nn.Linear(config.hidden_size, config.hidden_size))

        self.projector = nn.Sequential(*modules)
        query_num = int(mlp_gelu_cross_att_match.group(3))
        self.query = nn.Parameter(torch.zeros(1, query_num, config.hidden_size))
        self.query.data.normal_(mean=0.0, std=0.02)

        att_layer_num = int(mlp_gelu_cross_att_match.group(2))
        decoder_layer = nn.TransformerDecoderLayer(
            config.hidden_size,
            config.num_attention_heads,
            dim_feedforward=config.hidden_size * 4,
            dropout=0,
            activation="gelu",
            batch_first=True,
        )
        self.cross_att_layers = nn.TransformerDecoder(
            decoder_layer, att_layer_num, norm=None
        )

    def forward(self, vision_embedding):
        projected_embedding = self.projector(vision_embedding)
        batch_size = vision_embedding.shape[0]
        query = self.query.expand(batch_size, -1, -1)
        output = self.cross_att_layers(query, projected_embedding)
        return output

class MlpGeLUPreCrossAttProjector(nn.Module):
    def __init__(self, mlp_gelu_cross_att_match, config):
        super(MlpGeLUPreCrossAttProjector, self).__init__()

        mlp_depth = int(mlp_gelu_cross_att_match.group(1))
        modules = [nn.Linear(config.mm_hidden_size, config.hidden_size)]
        for _ in range(1, mlp_depth):
            modules.append(nn.GELU())
            modules.append(nn.Linear(config.hidden_size, config.hidden_size))
        self.projector = nn.Sequential(*modules)

        query_num = int(mlp_gelu_cross_att_match.group(3))
        self.query = nn.Parameter(torch.zeros(1, query_num, config.mm_hidden_size))
        self.query.data.normal_(mean=0.0, std=0.02)
        att_layer_num = int(mlp_gelu_cross_att_match.group(2))
        decoder_layer = nn.TransformerDecoderLayer(
            config.mm_hidden_size,
            8,
            dim_feedforward=config.mm_hidden_size * 4,
            dropout=0,
            activation="gelu",
            batch_first=True,
        )
        self.cross_att_layers = nn.TransformerDecoder(
            decoder_layer, att_layer_num, norm=None
        )

    def forward(self, vision_embedding):
        batch_size = vision_embedding.shape[0]
        query = self.query.expand(batch_size, -1, -1)
        output = self.cross_att_layers(query, vision_embedding)
        output = self.projector(output)
        return output


def build_vision_projector(config, delay_load=False, **kwargs):
    projector_type = getattr(config, "mm_projector_type", "linear")

    if projector_type == "linear":
        return nn.Linear(config.mm_hidden_size, config.hidden_size)

    mlp_gelu_match = re.match(r"^mlp(\d+)x_gelu$", projector_type)
    if mlp_gelu_match:
        mlp_depth = int(mlp_gelu_match.group(1))
        modules = [nn.Linear(config.mm_hidden_size, config.hidden_size)]
        for _ in range(1, mlp_depth):
            modules.append(nn.GELU())
            modules.append(nn.Linear(config.hidden_size, config.hidden_size))
        return nn.Sequential(*modules)

    mlp_gelu_pre_cross_att_match = re.match(
        r"^mlp(\d+)x_gelu_pre_(\d+)att_(\d+)q$", projector_type
    )
    if mlp_gelu_pre_cross_att_match:
        return MlpGeLUPreCrossAttProjector(mlp_gelu_pre_cross_att_match, config)

    mlp_gelu_post_cross_att_match = re.match(
        r"^mlp(\d+)x_gelu_post_(\d+)att_(\d+)q$", projector_type
    )
    if mlp_gelu_post_cross_att_match:
        return MlpGeLUPostCrossAttProjector(mlp_gelu_post_cross_att_match, config)

    if projector_type == "identity":
        return IdentityMap()

    raise ValueError(f"Unknown projector type: {projector_type}")