File size: 5,470 Bytes
d9bb75c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This software may be used and distributed in accordance with
# the terms of the DINOv3 License Agreement.

import logging
from typing import Optional

import torch

from dinov3.eval.text.text_transformer import TextTransformer
from dinov3.layers import CausalSelfAttentionBlock
from torch import nn

logger = logging.getLogger("dinov3")


class TextHead(nn.Module):
    def __init__(
        self,
        input_dim: int,
        embed_dim: int,
        num_heads: int,
        num_blocks: int,
        block_drop_prob: float,
        is_causal: bool,
        use_linear_projection: bool,
    ):
        super().__init__()
        block_list = [nn.Identity()]
        self.ln_final = nn.Identity()
        if num_blocks > 0:
            logger.info(f"Adding {num_blocks} text tower transformer head blocks")
            block_list = [
                CausalSelfAttentionBlock(
                    dim=input_dim,
                    num_heads=num_heads,
                    is_causal=is_causal,
                    dropout_prob=block_drop_prob,
                )
                for _ in range(num_blocks)
            ]
            self.ln_final = nn.LayerNorm(input_dim)
        self.blocks = nn.ModuleList(block_list)
        self.num_blocks = num_blocks
        self.linear_projection = nn.Identity()
        if input_dim != embed_dim or use_linear_projection:
            logger.info(
                f"Text tower : Using a linear projection from {input_dim} to {embed_dim}"
            )
            self.linear_projection = nn.Linear(input_dim, embed_dim, bias=False)

    def init_weights(self):
        if self.num_blocks > 0:
            for i in range(self.num_blocks):
                self.blocks[i].init_weights()
            self.ln_final.reset_parameters()
        if isinstance(self.linear_projection, nn.Linear):
            nn.init.normal_(
                self.linear_projection.weight,
                std=self.linear_projection.in_features**-0.5,
            )

    def forward(self, text_tokens: torch.Tensor) -> torch.Tensor:
        for block in self.blocks:
            text_tokens = block(text_tokens)
        text_tokens = self.ln_final(text_tokens)
        return self.linear_projection(text_tokens)


class TextTower(nn.Module):
    def __init__(
        self,
        backbone: nn.Module,
        freeze_backbone: bool,
        embed_dim: int,
        num_head_blocks: int,
        head_blocks_is_causal: bool,
        head_blocks_block_drop_prob: float,
        tokens_pooler_type: str,
        use_linear_projection: bool,
    ):
        super().__init__()
        self.backbone = backbone
        self.freeze_backbone = freeze_backbone
        backbone_out_dim = backbone.embed_dim
        logger.info(f"Text backbone embedding dimension: {backbone_out_dim}")
        self.backbone = backbone
        self.head = TextHead(
            backbone_out_dim,
            embed_dim,
            self.backbone.num_heads,
            num_head_blocks,
            head_blocks_block_drop_prob,
            head_blocks_is_causal,
            use_linear_projection,
        )
        self.tokens_pooler_type = tokens_pooler_type

    def init_weights(self):
        self.backbone.init_weights()
        self.head.init_weights()

    def forward(self, token_indices: torch.Tensor) -> torch.Tensor:
        text_tokens = self.backbone(token_indices)
        text_tokens = self.head(text_tokens)
        if self.tokens_pooler_type == "first":
            features = text_tokens[:, 0]
        elif self.tokens_pooler_type == "last":
            features = text_tokens[:, -1]
        elif self.tokens_pooler_type == "argmax":
            assert token_indices is not None
            features = text_tokens[
                torch.arange(text_tokens.shape[0]), token_indices.argmax(dim=-1)
            ]
        else:
            raise ValueError(f"Unknown text tokens pooler type: {self.pooler_type}")
        return features


def build_text_backbone(
    cfg,
) -> torch.nn.Module:
    logger.info("Setting up a text transformer")
    model = TextTransformer(
        context_length=cfg.context_length,
        vocab_size=cfg.vocab_size,
        dim=cfg.dim,
        num_heads=cfg.num_heads,
        num_layers=cfg.num_layers,
        ffn_ratio=cfg.ffn_ratio,
        is_causal=cfg.is_causal,
        ls_init_value=cfg.ls_init_value,
        dropout_prob=cfg.dropout_prob,
    )
    logger.info(f"Setting upa custom text transformer {cfg.model_name}")
    return model


def build_text_model(
    embed_dim: int,
    backbone_model_config: str,
    freeze_backbone: bool,
    num_head_blocks: int,
    head_blocks_is_causal: bool,
    head_blocks_drop_prob: float,
    tokens_pooler_type: str,
    use_linear_projection: bool,
    backbone: Optional[nn.Module] = None,
):
    if backbone is None:
        if backbone_model_config is not None:
            from omegaconf import OmegaConf

            cfg = OmegaConf.load(backbone_model_config)
            backbone = build_text_backbone(cfg)
        else:
            raise RuntimeError(
                "Failed to create, text backbone, either backbone or backbone_model_config should be not None"
            )
    return TextTower(
        backbone,
        freeze_backbone,
        embed_dim,
        num_head_blocks,
        head_blocks_is_causal,
        head_blocks_drop_prob,
        tokens_pooler_type,
        use_linear_projection,
    )