File size: 5,577 Bytes
434b0b0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) 2023-2024, Zexin He
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.


import torch.nn as nn

from .modulate import ModLN


class BasicBlock(nn.Module):
    """
    Transformer block that is in its simplest form.
    Designed for PF-LRM architecture.
    """

    # Block contains a self-attention layer and an MLP
    def __init__(
        self,
        inner_dim: int,
        num_heads: int,
        eps: float,
        attn_drop: float = 0.0,
        attn_bias: bool = False,
        mlp_ratio: float = 4.0,
        mlp_drop: float = 0.0,
    ):
        super().__init__()
        self.norm1 = nn.LayerNorm(inner_dim, eps=eps)
        self.self_attn = nn.MultiheadAttention(
            embed_dim=inner_dim,
            num_heads=num_heads,
            dropout=attn_drop,
            bias=attn_bias,
            batch_first=True,
        )
        self.norm2 = nn.LayerNorm(inner_dim, eps=eps)
        self.mlp = nn.Sequential(
            nn.Linear(inner_dim, int(inner_dim * mlp_ratio)),
            nn.GELU(),
            nn.Dropout(mlp_drop),
            nn.Linear(int(inner_dim * mlp_ratio), inner_dim),
            nn.Dropout(mlp_drop),
        )

    def forward(self, x):
        # x: [N, L, D]
        before_sa = self.norm1(x)
        x = x + self.self_attn(before_sa, before_sa, before_sa, need_weights=False)[0]
        x = x + self.mlp(self.norm2(x))
        return x


class ConditionBlock(nn.Module):
    """
    Transformer block that takes in a cross-attention condition.
    Designed for SparseLRM architecture.
    """

    # Block contains a cross-attention layer, a self-attention layer, and an MLP
    def __init__(
        self,
        inner_dim: int,
        cond_dim: int,
        num_heads: int,
        eps: float,
        attn_drop: float = 0.0,
        attn_bias: bool = False,
        mlp_ratio: float = 4.0,
        mlp_drop: float = 0.0,
    ):
        super().__init__()
        self.norm1 = nn.LayerNorm(inner_dim, eps=eps)
        self.cross_attn = nn.MultiheadAttention(
            embed_dim=inner_dim,
            num_heads=num_heads,
            kdim=cond_dim,
            vdim=cond_dim,
            dropout=attn_drop,
            bias=attn_bias,
            batch_first=True,
        )
        self.norm2 = nn.LayerNorm(inner_dim, eps=eps)
        self.self_attn = nn.MultiheadAttention(
            embed_dim=inner_dim,
            num_heads=num_heads,
            dropout=attn_drop,
            bias=attn_bias,
            batch_first=True,
        )
        self.norm3 = nn.LayerNorm(inner_dim, eps=eps)
        self.mlp = nn.Sequential(
            nn.Linear(inner_dim, int(inner_dim * mlp_ratio)),
            nn.GELU(),
            nn.Dropout(mlp_drop),
            nn.Linear(int(inner_dim * mlp_ratio), inner_dim),
            nn.Dropout(mlp_drop),
        )

    def forward(self, x, cond):
        # x: [N, L, D]
        # cond: [N, L_cond, D_cond]
        x = x + self.cross_attn(self.norm1(x), cond, cond, need_weights=False)[0]
        before_sa = self.norm2(x)
        x = x + self.self_attn(before_sa, before_sa, before_sa, need_weights=False)[0]
        x = x + self.mlp(self.norm3(x))
        return x


class ConditionModulationBlock(nn.Module):
    """
    Transformer block that takes in a cross-attention condition and another modulation vector applied to sub-blocks.
    Designed for raw LRM architecture.
    """

    # Block contains a cross-attention layer, a self-attention layer, and an MLP
    def __init__(
        self,
        inner_dim: int,
        cond_dim: int,
        mod_dim: int,
        num_heads: int,
        eps: float,
        attn_drop: float = 0.0,
        attn_bias: bool = False,
        mlp_ratio: float = 4.0,
        mlp_drop: float = 0.0,
    ):
        super().__init__()
        self.norm1 = ModLN(inner_dim, mod_dim, eps)
        self.cross_attn = nn.MultiheadAttention(
            embed_dim=inner_dim,
            num_heads=num_heads,
            kdim=cond_dim,
            vdim=cond_dim,
            dropout=attn_drop,
            bias=attn_bias,
            batch_first=True,
        )
        self.norm2 = ModLN(inner_dim, mod_dim, eps)
        self.self_attn = nn.MultiheadAttention(
            embed_dim=inner_dim,
            num_heads=num_heads,
            dropout=attn_drop,
            bias=attn_bias,
            batch_first=True,
        )
        self.norm3 = ModLN(inner_dim, mod_dim, eps)
        self.mlp = nn.Sequential(
            nn.Linear(inner_dim, int(inner_dim * mlp_ratio)),
            nn.GELU(),
            nn.Dropout(mlp_drop),
            nn.Linear(int(inner_dim * mlp_ratio), inner_dim),
            nn.Dropout(mlp_drop),
        )

    def forward(self, x, cond, mod):
        # x: [N, L, D]
        # cond: [N, L_cond, D_cond]
        # mod: [N, D_mod]
        x = x + self.cross_attn(self.norm1(x, mod), cond, cond, need_weights=False)[0]
        before_sa = self.norm2(x, mod)
        x = x + self.self_attn(before_sa, before_sa, before_sa, need_weights=False)[0]
        x = x + self.mlp(self.norm3(x, mod))
        return x