File size: 9,737 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
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
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
from typing import *
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from ..modules.utils import zero_module, convert_module_to_f16, convert_module_to_f32
from ..modules.transformer import AbsolutePositionEmbedder
from ..modules.norm import LayerNorm32
from ..modules import sparse as sp
from ..modules.sparse.transformer import ModulatedSparseTransformerCrossBlock
from .sparse_structure_flow import TimestepEmbedder


class SparseResBlock3d(nn.Module):
    def __init__(
        self,
        channels: int,
        emb_channels: int,
        out_channels: Optional[int] = None,
        downsample: bool = False,
        upsample: bool = False,
    ):
        super().__init__()
        self.channels = channels
        self.emb_channels = emb_channels
        self.out_channels = out_channels or channels
        self.downsample = downsample
        self.upsample = upsample
        
        assert not (downsample and upsample), "Cannot downsample and upsample at the same time"

        self.norm1 = LayerNorm32(channels, elementwise_affine=True, eps=1e-6)
        self.norm2 = LayerNorm32(self.out_channels, elementwise_affine=False, eps=1e-6)
        self.conv1 = sp.SparseConv3d(channels, self.out_channels, 3)
        self.conv2 = zero_module(sp.SparseConv3d(self.out_channels, self.out_channels, 3))
        self.emb_layers = nn.Sequential(
            nn.SiLU(),
            nn.Linear(emb_channels, 2 * self.out_channels, bias=True),
        )
        self.skip_connection = sp.SparseLinear(channels, self.out_channels) if channels != self.out_channels else nn.Identity()
        self.updown = None
        if self.downsample:
            self.updown = sp.SparseDownsample(2)
        elif self.upsample:
            self.updown = sp.SparseUpsample(2)

    def _updown(self, x: sp.SparseTensor) -> sp.SparseTensor:
        if self.updown is not None:
            x = self.updown(x)
        return x

    def forward(self, x: sp.SparseTensor, emb: torch.Tensor) -> sp.SparseTensor:
        emb_out = self.emb_layers(emb).type(x.dtype)
        scale, shift = torch.chunk(emb_out, 2, dim=1)

        x = self._updown(x)
        h = x.replace(self.norm1(x.feats))
        h = h.replace(F.silu(h.feats))
        h = self.conv1(h)
        h = h.replace(self.norm2(h.feats)) * (1 + scale) + shift
        h = h.replace(F.silu(h.feats))
        h = self.conv2(h)
        h = h + self.skip_connection(x)

        return h
    

class SLatFlowModel(nn.Module):
    def __init__(
        self,
        resolution: int,
        in_channels: int,
        model_channels: int,
        cond_channels: int,
        out_channels: int,
        num_blocks: int,
        num_heads: Optional[int] = None,
        num_head_channels: Optional[int] = 64,
        mlp_ratio: float = 4,
        patch_size: int = 2,
        num_io_res_blocks: int = 2,
        io_block_channels: List[int] = None,
        pe_mode: Literal["ape", "rope"] = "ape",
        use_fp16: bool = False,
        use_checkpoint: bool = False,
        use_skip_connection: bool = True,
        share_mod: bool = False,
        qk_rms_norm: bool = False,
        qk_rms_norm_cross: bool = False,
    ):
        super().__init__()
        self.resolution = resolution
        self.in_channels = in_channels
        self.model_channels = model_channels
        self.cond_channels = cond_channels
        self.out_channels = out_channels
        self.num_blocks = num_blocks
        self.num_heads = num_heads or model_channels // num_head_channels
        self.mlp_ratio = mlp_ratio
        self.patch_size = patch_size
        self.num_io_res_blocks = num_io_res_blocks
        self.io_block_channels = io_block_channels
        self.pe_mode = pe_mode
        self.use_fp16 = use_fp16
        self.use_checkpoint = use_checkpoint
        self.use_skip_connection = use_skip_connection
        self.share_mod = share_mod
        self.qk_rms_norm = qk_rms_norm
        self.qk_rms_norm_cross = qk_rms_norm_cross
        self.dtype = torch.float16 if use_fp16 else torch.float32

        assert int(np.log2(patch_size)) == np.log2(patch_size), "Patch size must be a power of 2"
        assert np.log2(patch_size) == len(io_block_channels), "Number of IO ResBlocks must match the number of stages"

        self.t_embedder = TimestepEmbedder(model_channels)
        if share_mod:
            self.adaLN_modulation = nn.Sequential(
                nn.SiLU(),
                nn.Linear(model_channels, 6 * model_channels, bias=True)
            )

        if pe_mode == "ape":
            self.pos_embedder = AbsolutePositionEmbedder(model_channels)

        self.input_layer = sp.SparseLinear(in_channels, io_block_channels[0])
        self.input_blocks = nn.ModuleList([])
        for chs, next_chs in zip(io_block_channels, io_block_channels[1:] + [model_channels]):
            self.input_blocks.extend([
                SparseResBlock3d(
                    chs,
                    model_channels,
                    out_channels=chs,
                )
                for _ in range(num_io_res_blocks-1)
            ])
            self.input_blocks.append(
                SparseResBlock3d(
                    chs,
                    model_channels,
                    out_channels=next_chs,
                    downsample=True,
                )
            )
            
        self.blocks = nn.ModuleList([
            ModulatedSparseTransformerCrossBlock(
                model_channels,
                cond_channels,
                num_heads=self.num_heads,
                mlp_ratio=self.mlp_ratio,
                attn_mode='full',
                use_checkpoint=self.use_checkpoint,
                use_rope=(pe_mode == "rope"),
                share_mod=self.share_mod,
                qk_rms_norm=self.qk_rms_norm,
                qk_rms_norm_cross=self.qk_rms_norm_cross,
            )
            for _ in range(num_blocks)
        ])

        self.out_blocks = nn.ModuleList([])
        for chs, prev_chs in zip(reversed(io_block_channels), [model_channels] + list(reversed(io_block_channels[1:]))):
            self.out_blocks.append(
                SparseResBlock3d(
                    prev_chs * 2 if self.use_skip_connection else prev_chs,
                    model_channels,
                    out_channels=chs,
                    upsample=True,
                )
            )
            self.out_blocks.extend([
                SparseResBlock3d(
                    chs * 2 if self.use_skip_connection else chs,
                    model_channels,
                    out_channels=chs,
                )
                for _ in range(num_io_res_blocks-1)
            ])
        self.out_layer = sp.SparseLinear(io_block_channels[0], out_channels)

        self.initialize_weights()
        if use_fp16:
            self.convert_to_fp16()

    @property
    def device(self) -> torch.device:
        """
        Return the device of the model.
        """
        return next(self.parameters()).device

    def convert_to_fp16(self) -> None:
        """
        Convert the torso of the model to float16.
        """
        self.input_blocks.apply(convert_module_to_f16)
        self.blocks.apply(convert_module_to_f16)
        self.out_blocks.apply(convert_module_to_f16)

    def convert_to_fp32(self) -> None:
        """
        Convert the torso of the model to float32.
        """
        self.input_blocks.apply(convert_module_to_f32)
        self.blocks.apply(convert_module_to_f32)
        self.out_blocks.apply(convert_module_to_f32)

    def initialize_weights(self) -> None:
        # Initialize transformer layers:
        def _basic_init(module):
            if isinstance(module, nn.Linear):
                torch.nn.init.xavier_uniform_(module.weight)
                if module.bias is not None:
                    nn.init.constant_(module.bias, 0)
        self.apply(_basic_init)

        # Initialize timestep embedding MLP:
        nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02)
        nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02)

        # Zero-out adaLN modulation layers in DiT blocks:
        if self.share_mod:
            nn.init.constant_(self.adaLN_modulation[-1].weight, 0)
            nn.init.constant_(self.adaLN_modulation[-1].bias, 0)
        else:
            for block in self.blocks:
                nn.init.constant_(block.adaLN_modulation[-1].weight, 0)
                nn.init.constant_(block.adaLN_modulation[-1].bias, 0)

        # Zero-out output layers:
        nn.init.constant_(self.out_layer.weight, 0)
        nn.init.constant_(self.out_layer.bias, 0)

    def forward(self, x: sp.SparseTensor, t: torch.Tensor, cond: torch.Tensor) -> sp.SparseTensor:
        h = self.input_layer(x).type(self.dtype)
        t_emb = self.t_embedder(t)
        if self.share_mod:
            t_emb = self.adaLN_modulation(t_emb)
        t_emb = t_emb.type(self.dtype)
        cond = cond.type(self.dtype)

        skips = []
        # pack with input blocks
        for block in self.input_blocks:
            h = block(h, t_emb)
            skips.append(h.feats)
        
        if self.pe_mode == "ape":
            h = h + self.pos_embedder(h.coords[:, 1:]).type(self.dtype)
        for block in self.blocks:
            h = block(h, t_emb, cond)

        # unpack with output blocks
        for block, skip in zip(self.out_blocks, reversed(skips)):
            if self.use_skip_connection:
                h = block(h.replace(torch.cat([h.feats, skip], dim=1)), t_emb)
            else:
                h = block(h, t_emb)

        h = h.replace(F.layer_norm(h.feats, h.feats.shape[-1:]))
        h = self.out_layer(h.type(x.dtype))
        return h