File size: 7,329 Bytes
801501a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
import torch
import torch.nn as nn
import torch.nn.functional as F
from dotmap import DotMap
from salad.model_components.simple_module import TimePointWiseEncoder, TimestepEmbedder


from salad.model_components.transformer import (
    PositionalEncoding,
    TimeTransformerDecoder,
    TimeTransformerEncoder,
)

class UnCondDiffNetwork(nn.Module):
    def __init__(self, input_dim, residual, **kwargs):
        """
        Transformer Encoder.
        """
        super().__init__()
        self.input_dim = input_dim
        self.residual = residual
        self.__dict__.update(kwargs)
        self.hparams = DotMap(self.__dict__)

        self._build_model()

    def _build_model(self):
        self.act = F.leaky_relu
        if self.hparams.get("use_timestep_embedder"):
            self.time_embedder = TimestepEmbedder(self.hparams.timestep_embedder_dim)
            dim_ctx = self.hparams.timestep_embedder_dim
        else:
            dim_ctx = 3

        """
        Encoder part
        """
        enc_dim = self.hparams.embedding_dim
        self.embedding = nn.Linear(self.hparams.input_dim, enc_dim)
        if not self.hparams.get("encoder_type"):
            self.encoder = TimeTransformerEncoder(
                enc_dim,
                dim_ctx=dim_ctx,
                num_heads=self.hparams.num_heads
                if self.hparams.get("num_heads")
                else 4,
                use_time=True,
                num_layers=self.hparams.enc_num_layers,
                last_fc=True,
                last_fc_dim_out=self.hparams.input_dim,
            )
        else:
            if self.hparams.encoder_type == "transformer":
                self.encoder = TimeTransformerEncoder(
                    enc_dim,
                    dim_ctx=dim_ctx,
                    num_heads=self.hparams.num_heads
                    if self.hparams.get("num_heads")
                    else 4,
                    use_time=True,
                    num_layers=self.hparams.enc_num_layers,
                    last_fc=True,
                    last_fc_dim_out=self.hparams.input_dim,
                    dropout=self.hparams.get("attn_dropout", 0.0)
                )
            else:
                raise ValueError

    def forward(self, x, beta):
        """
        Input:
            x: [B,G,D] latent
            beta: B
        Output:
            eta: [B,G,D]
        """
        B, G = x.shape[:2]
        if self.hparams.get("use_timestep_embedder"):
            time_emb = self.time_embedder(beta).unsqueeze(1)
        else:
            beta = beta.view(B, 1, 1)
            time_emb = torch.cat(
                [beta, torch.sin(beta), torch.cos(beta)], dim=-1
            )  # [B,1,3]

        ctx = time_emb
        x_emb = self.embedding(x)

        out = self.encoder(x_emb, ctx=ctx)

        if self.hparams.residual:
            out = out + x
        return out


class CondDiffNetwork(nn.Module):
    def __init__(self, input_dim, residual, **kwargs):
        """
        Transformer Encoder + Decoder.
        """
        super().__init__()
        self.input_dim = input_dim
        self.residual = residual
        self.__dict__.update(kwargs)
        self.hparams = DotMap(self.__dict__)

        self._build_model()

    def _build_model(self):
        self.act = F.leaky_relu
        if self.hparams.get("use_timestep_embedder"):
            self.time_embedder = TimestepEmbedder(self.hparams.timestep_embedder_dim)
            dim_ctx = self.hparams.timestep_embedder_dim
        else:
            dim_ctx = 3
        """
        Encoder part
        """
        enc_dim = self.hparams.context_embedding_dim
        self.context_embedding = nn.Linear(self.hparams.context_dim, enc_dim)
        if self.hparams.encoder_type == "transformer":
            self.encoder = TimeTransformerEncoder(
                enc_dim,
                3,
                num_heads=4,
                use_time=self.hparams.encoder_use_time,
                num_layers=self.hparams.enc_num_layers
                if self.hparams.get("enc_num_layers")
                else 3,
                last_fc=False,
            )

        elif self.hparams.encoder_type == "pointwise":
            self.encoder = TimePointWiseEncoder(
                enc_dim,
                dim_ctx=None,
                use_time=self.hparams.encoder_use_time,
                num_layers=self.hparams.enc_num_layers,
            )
        else:
            raise ValueError

        """ 
        Decoder part
        """
        dec_dim = self.hparams.embedding_dim
        input_dim = self.hparams.input_dim
        self.query_embedding = nn.Linear(self.hparams.input_dim, dec_dim)
        if self.hparams.decoder_type == "transformer_decoder":
            self.decoder = TimeTransformerDecoder(
                dec_dim,
                enc_dim,
                dim_ctx=dim_ctx,
                num_heads=4,
                last_fc=True,
                last_fc_dim_out=input_dim,
                num_layers=self.hparams.dec_num_layers
                if self.hparams.get("dec_num_layers")
                else 3,
            )
        elif self.hparams.decoder_type == "transformer_encoder":
            self.decoder = TimeTransformerEncoder(
                dec_dim,
                dim_ctx=enc_dim + dim_ctx,
                num_heads=4,
                last_fc=True,
                last_fc_dim_out=input_dim,
                num_layers=self.hparams.dec_num_layers
                if self.hparams.get("dec_num_layers")
                else 3,
            )
        else:
            raise ValueError

    def forward(self, x, beta, context):
        """
        Input:
            x: [B,G,D] intrinsic
            beta: B
            context: [B,G,D2] or [B, D2] condition
        Output:
            eta: [B,G,D]
        """
        # print(f"x: {x.shape} context: {context.shape} beta: {beta.shape}")
        B, G = x.shape[:2]

        if self.hparams.get("use_timestep_embedder"):
            time_emb = self.time_embedder(beta).unsqueeze(1)
        else:
            beta = beta.view(B, 1, 1)
            time_emb = torch.cat(
                [beta, torch.sin(beta), torch.cos(beta)], dim=-1
            )  # [B,1,3]
        ctx = time_emb
        """
        Encoding
        """
        cout = self.context_embedding(context)
        cout = self.encoder(cout, ctx=ctx if self.hparams.encoder_use_time else None)

        if cout.ndim == 2:
            cout = cout.unsqueeze(1).expand(-1, G, -1)

        """
        Decoding
        """
        out = self.query_embedding(x)
        if self.hparams.get("use_pos_encoding"):
            out = self.pos_encoding(out)

        if self.hparams.decoder_type == "transformer_encoder":
            try:
                ctx = ctx.expand(-1, G, -1)
                if cout.ndim == 2:
                    cout = cout.unsqueeze(1)
                cout = cout.expand(-1, G, -1)
                ctx = torch.cat([ctx, cout], -1)
            except Exception as e:
                print(e, G, ctx.shape, cout.shape)
            out = self.decoder(out, ctx=ctx)
        else:
            out = self.decoder(out, cout, ctx=ctx)

        # if hasattr(self, "last_fc"):
        # out = self.last_fc(out)

        if self.hparams.residual:
            out = out + x
        return out