File size: 5,865 Bytes
db6a3b7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from typing import *
import torch
import torch.nn as nn
from ..attention import MultiHeadAttention
from ..norm import LayerNorm32


class AbsolutePositionEmbedder(nn.Module):
    """
    Embeds spatial positions into vector representations.
    """
    def __init__(self, channels: int, in_channels: int = 3):
        super().__init__()
        self.channels = channels
        self.in_channels = in_channels
        self.freq_dim = channels // in_channels // 2
        self.freqs = torch.arange(self.freq_dim, dtype=torch.float32) / self.freq_dim
        self.freqs = 1.0 / (10000 ** self.freqs)
        
    def _sin_cos_embedding(self, x: torch.Tensor) -> torch.Tensor:
        """
        Create sinusoidal position embeddings.

        Args:
            x: a 1-D Tensor of N indices

        Returns:
            an (N, D) Tensor of positional embeddings.
        """
        self.freqs = self.freqs.to(x.device)
        out = torch.outer(x, self.freqs)
        out = torch.cat([torch.sin(out), torch.cos(out)], dim=-1)
        return out

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Args:
            x (torch.Tensor): (N, D) tensor of spatial positions
        """
        N, D = x.shape
        assert D == self.in_channels, "Input dimension must match number of input channels"
        embed = self._sin_cos_embedding(x.reshape(-1))
        embed = embed.reshape(N, -1)
        if embed.shape[1] < self.channels:
            embed = torch.cat([embed, torch.zeros(N, self.channels - embed.shape[1], device=embed.device)], dim=-1)
        return embed


class FeedForwardNet(nn.Module):
    def __init__(self, channels: int, mlp_ratio: float = 4.0):
        super().__init__()
        self.mlp = nn.Sequential(
            nn.Linear(channels, int(channels * mlp_ratio)),
            nn.GELU(approximate="tanh"),
            nn.Linear(int(channels * mlp_ratio), channels),
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self.mlp(x)


class TransformerBlock(nn.Module):
    """
    Transformer block (MSA + FFN).
    """
    def __init__(
        self,
        channels: int,
        num_heads: int,
        mlp_ratio: float = 4.0,
        attn_mode: Literal["full", "windowed"] = "full",
        window_size: Optional[int] = None,
        shift_window: Optional[int] = None,
        use_checkpoint: bool = False,
        use_rope: bool = False,
        qk_rms_norm: bool = False,
        qkv_bias: bool = True,
        ln_affine: bool = False,
    ):
        super().__init__()
        self.use_checkpoint = use_checkpoint
        self.norm1 = LayerNorm32(channels, elementwise_affine=ln_affine, eps=1e-6)
        self.norm2 = LayerNorm32(channels, elementwise_affine=ln_affine, eps=1e-6)
        self.attn = MultiHeadAttention(
            channels,
            num_heads=num_heads,
            attn_mode=attn_mode,
            window_size=window_size,
            shift_window=shift_window,
            qkv_bias=qkv_bias,
            use_rope=use_rope,
            qk_rms_norm=qk_rms_norm,
        )
        self.mlp = FeedForwardNet(
            channels,
            mlp_ratio=mlp_ratio,
        )

    def _forward(self, x: torch.Tensor) -> torch.Tensor:
        h = self.norm1(x)
        h = self.attn(h)
        x = x + h
        h = self.norm2(x)
        h = self.mlp(h)
        x = x + h
        return x

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        if self.use_checkpoint:
            return torch.utils.checkpoint.checkpoint(self._forward, x, use_reentrant=False)
        else:
            return self._forward(x)


class TransformerCrossBlock(nn.Module):
    """
    Transformer cross-attention block (MSA + MCA + FFN).
    """
    def __init__(
        self,
        channels: int,
        ctx_channels: int,
        num_heads: int,
        mlp_ratio: float = 4.0,
        attn_mode: Literal["full", "windowed"] = "full",
        window_size: Optional[int] = None,
        shift_window: Optional[Tuple[int, int, int]] = None,
        use_checkpoint: bool = False,
        use_rope: bool = False,
        qk_rms_norm: bool = False,
        qk_rms_norm_cross: bool = False,
        qkv_bias: bool = True,
        ln_affine: bool = False,
    ):
        super().__init__()
        self.use_checkpoint = use_checkpoint
        self.norm1 = LayerNorm32(channels, elementwise_affine=ln_affine, eps=1e-6)
        self.norm2 = LayerNorm32(channels, elementwise_affine=ln_affine, eps=1e-6)
        self.norm3 = LayerNorm32(channels, elementwise_affine=ln_affine, eps=1e-6)
        self.self_attn = MultiHeadAttention(
            channels,
            num_heads=num_heads,
            type="self",
            attn_mode=attn_mode,
            window_size=window_size,
            shift_window=shift_window,
            qkv_bias=qkv_bias,
            use_rope=use_rope,
            qk_rms_norm=qk_rms_norm,
        )
        self.cross_attn = MultiHeadAttention(
            channels,
            ctx_channels=ctx_channels,
            num_heads=num_heads,
            type="cross",
            attn_mode="full",
            qkv_bias=qkv_bias,
            qk_rms_norm=qk_rms_norm_cross,
        )
        self.mlp = FeedForwardNet(
            channels,
            mlp_ratio=mlp_ratio,
        )

    def _forward(self, x: torch.Tensor, context: torch.Tensor):
        h = self.norm1(x)
        h = self.self_attn(h)
        x = x + h
        h = self.norm2(x)
        h = self.cross_attn(h, context)
        x = x + h
        h = self.norm3(x)
        h = self.mlp(h)
        x = x + h
        return x

    def forward(self, x: torch.Tensor, context: torch.Tensor):
        if self.use_checkpoint:
            return torch.utils.checkpoint.checkpoint(self._forward, x, context, use_reentrant=False)
        else:
            return self._forward(x, context)