File size: 6,302 Bytes
2690807
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from typing import Optional, Tuple, Union

import torch
from .configuration_aimv2 import AIMv2Config
from torch import nn
from torch.nn import functional as F
from transformers.modeling_outputs import BaseModelOutputWithNoAttention
from transformers.modeling_utils import PreTrainedModel

__all__ = ["AIMv2Model"]


class RMSNorm(nn.Module):
    def __init__(self, dim: int, eps: float = 1e-6):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(dim))
        self.eps = eps

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        output = self._norm(x.float()).type_as(x)
        return output * self.weight

    def extra_repr(self) -> str:
        return f"{tuple(self.weight.shape)}, eps={self.eps}"

    def _norm(self, x: torch.Tensor) -> torch.Tensor:
        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)


class AIMv2SwiGLUFFN(nn.Module):
    def __init__(self, config: AIMv2Config):
        super().__init__()
        hidden_features = config.intermediate_size
        in_features = config.hidden_size
        bias = config.use_bias

        self.fc1 = nn.Linear(in_features, hidden_features, bias=bias)
        self.fc2 = nn.Linear(hidden_features, in_features, bias=bias)
        self.fc3 = nn.Linear(in_features, hidden_features, bias=bias)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = F.silu(self.fc1(x)) * self.fc3(x)
        x = self.fc2(x)
        return x


class AIMv2PatchEmbed(nn.Module):
    def __init__(self, config: AIMv2Config):
        super().__init__()
        self.proj = nn.Conv2d(
            config.num_channels,
            config.hidden_size,
            kernel_size=(config.patch_size, config.patch_size),
            stride=(config.patch_size, config.patch_size),
        )
        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = self.proj(x).flatten(2).transpose(1, 2)
        x = self.norm(x)
        return x


class AIMv2ViTPreprocessor(nn.Module):
    def __init__(self, config: AIMv2Config):
        super().__init__()
        num_patches = (config.image_size // config.patch_size) ** 2

        self.patchifier = AIMv2PatchEmbed(config)
        self.pos_embed = nn.Parameter(torch.zeros((1, num_patches, config.hidden_size)))

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        tokens = self.patchifier(x)
        _, N, _ = tokens.shape
        pos_embed = self.pos_embed.to(tokens.device)
        tokens = tokens + pos_embed[:, :N]
        return tokens


class AIMv2Attention(nn.Module):
    def __init__(self, config: AIMv2Config):
        super().__init__()
        dim = config.hidden_size

        self.num_heads = config.num_attention_heads
        self.qkv = nn.Linear(dim, dim * 3, bias=config.qkv_bias)
        self.attn_drop = nn.Dropout(config.attention_dropout)
        self.proj = nn.Linear(dim, dim, bias=config.use_bias)
        self.proj_drop = nn.Dropout(config.projection_dropout)

    def forward(
        self, x: torch.Tensor, mask: Optional[torch.Tensor] = None
    ) -> torch.Tensor:
        B, N, C = x.shape
        qkv = (
            self.qkv(x)
            .reshape(B, N, 3, self.num_heads, C // self.num_heads)
            .permute(2, 0, 3, 1, 4)
        )
        q, k, v = qkv.unbind(0)

        x = F.scaled_dot_product_attention(q, k, v, attn_mask=mask)
        x = x.transpose(1, 2).contiguous().reshape(B, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x


class AIMv2Block(nn.Module):
    def __init__(self, config: AIMv2Config):
        super().__init__()
        self.attn = AIMv2Attention(config)
        self.norm_1 = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
        self.mlp = AIMv2SwiGLUFFN(config)
        self.norm_2 = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(
        self, x: torch.Tensor, mask: Optional[torch.Tensor] = None
    ) -> torch.Tensor:
        x = x + self.attn(self.norm_1(x), mask)
        x = x + self.mlp(self.norm_2(x))
        return x


class AIMv2Transformer(nn.Module):
    def __init__(self, config: AIMv2Config):
        super().__init__()
        self.blocks = nn.ModuleList(
            [AIMv2Block(config) for _ in range(config.num_hidden_layers)]
        )
        self.post_trunk_norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)

    def forward(
        self,
        tokens: torch.Tensor,
        mask: Optional[torch.Tensor] = None,
        output_hidden_states: bool = False,
    ) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor, ...]]]:
        hidden_states = () if output_hidden_states else None
        for block in self.blocks:
            tokens = block(tokens, mask)
            if output_hidden_states:
                hidden_states += (tokens,)
        tokens = self.post_trunk_norm(tokens)
        return tokens, hidden_states


class AIMv2PretrainedModel(PreTrainedModel):
    config_class = AIMv2Config
    base_model_prefix = "aimv2"
    main_input_name = "pixel_values"
    _supports_sdpa = True


class AIMv2Model(AIMv2PretrainedModel):
    def __init__(self, config: AIMv2Config):
        super().__init__(config)
        self.preprocessor = AIMv2ViTPreprocessor(config)
        self.trunk = AIMv2Transformer(config)

    def forward(
        self,
        pixel_values: torch.Tensor,
        mask: Optional[torch.Tensor] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[
        Tuple[torch.Tensor],
        Tuple[torch.Tensor, Tuple[torch.Tensor, ...]],
        BaseModelOutputWithNoAttention,
    ]:
        if output_hidden_states is None:
            output_hidden_states = self.config.output_hidden_states
        if return_dict is None:
            return_dict = self.config.use_return_dict

        x = self.preprocessor(pixel_values)
        x, hidden_states = self.trunk(
            x, mask, output_hidden_states=output_hidden_states
        )

        if not return_dict:
            res = (x,)
            res += (hidden_states,) if output_hidden_states else ()
            return res

        return BaseModelOutputWithNoAttention(
            last_hidden_state=x,
            hidden_states=hidden_states,
        )