File size: 11,130 Bytes
11e6f7b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
import torch.nn as nn
from inspect import isfunction
import math
import torch
import torch.nn.functional as F
from torch import nn, einsum
from einops import rearrange, repeat
from pdb import set_trace as st

from ldm.modules.attention import MemoryEfficientCrossAttention
from .dit_models_xformers import DiT, get_2d_sincos_pos_embed, DiTBlock, FinalLayer, t2i_modulate, PixelArtTextCondDiTBlock, T2IFinalLayer
# from .dit_models_xformers import CaptionEmbedder, approx_gelu, ImageCondDiTBlockPixelArt, t2i_modulate
# from fairscale.nn.model_parallel.layers import ColumnParallelLinear

from apex.normalization import FusedLayerNorm as LayerNorm

class DiT_TriLatent(DiT):
    # DiT with 3D_aware operations
    def __init__(
        self,
        input_size=32,
        patch_size=2,
        in_channels=4,
        hidden_size=1152,
        depth=28,
        num_heads=16,
        mlp_ratio=4,
        class_dropout_prob=0.1,
        num_classes=1000,
        learn_sigma=True,
        mixing_logit_init=-3,
        mixed_prediction=True,
        context_dim=False,
        roll_out=False,
        vit_blk=DiTBlock,
        final_layer_blk=FinalLayer,
    ):
        super().__init__(input_size, patch_size, in_channels, hidden_size,
                         depth, num_heads, mlp_ratio, class_dropout_prob,
                         num_classes, learn_sigma, mixing_logit_init,
                         mixed_prediction, context_dim, roll_out, vit_blk,
                         final_layer_blk)

        assert self.roll_out

    def init_PE_3D_aware(self):

        self.pos_embed = nn.Parameter(torch.zeros(
            1, self.plane_n * self.x_embedder.num_patches, self.embed_dim),
                                      requires_grad=False)

        # Initialize (and freeze) pos_embed by sin-cos embedding:
        p = int(self.x_embedder.num_patches**0.5)
        D = self.pos_embed.shape[-1]
        grid_size = (self.plane_n, p * p)  # B n HW C

        pos_embed = get_2d_sincos_pos_embed(D, grid_size).reshape(
            self.plane_n * p * p, D)  # H*W, D

        self.pos_embed.data.copy_(
            torch.from_numpy(pos_embed).float().unsqueeze(0))

    def initialize_weights(self):
        super().initialize_weights()

        # ! add 3d-aware PE
        self.init_PE_3D_aware()

    def forward(self,
                x,
                timesteps=None,
                context=None,
                y=None,
                get_attr='',
                **kwargs):
        """
        Forward pass of DiT.
        x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
        t: (N,) tensor of diffusion timesteps
        y: (N,) tensor of class labels
        """
        # t = timesteps
        assert context is not None

        t = self.t_embedder(timesteps)  # (N, D)

        # if self.roll_out:  # !
        x = rearrange(x, 'b (c n) h w->(b n) c h w',
                      n=3)  # downsample with same conv
        x = self.x_embedder(x)  # (b n) c h/f w/f

        x = rearrange(x, '(b n) l c -> b (n l) c', n=3)
        x = x + self.pos_embed  # (N, T, D), where T = H * W / patch_size ** 2

        # if self.roll_out:  # ! roll-out in the L dim, not B dim. add condition to all tokens.
        # x = rearrange(x, '(b n) l c ->b (n l) c', n=3)

        # assert context.ndim == 2
        if isinstance(context, dict):
            context = context['crossattn']  # sgm conditioner compat

        context = self.clip_text_proj(context)

        # c = t + context
        # else:
        # c = t  # BS 1024

        for blk_idx, block in enumerate(self.blocks):
            # if self.roll_out:
            if False:
                if blk_idx % 2 == 0:  # with-in plane self attention
                    x = rearrange(x, 'b (n l) c -> (b n) l c', n=3)
                    x = block(x, repeat(t, 'b c -> (b n) c ', n=3), # TODO, calculate once
                              repeat(context, 'b l c -> (b n) l c ', n=3))  # (N, T, D)

                else:  # global attention
                    x = rearrange(x, '(b n) l c -> b (n l) c ', n=self.plane_n)
                    x = block(x, t, context)  # (N, T, D)
            else:
                x = block(x, t, context)  # (N, T, D)

        # todo later
        x = self.final_layer(x, t)  # (N, T, patch_size ** 2 * out_channels)

        if self.roll_out:  # move n from L to B axis
            x = rearrange(x, 'b (n l) c ->(b n) l c', n=3)

        x = self.unpatchify(x)  # (N, out_channels, H, W)

        if self.roll_out:  # move n from L to B axis
            x = rearrange(x, '(b n) c h w -> b (c n) h w', n=3)
            # x = rearrange(x, 'b n) c h w -> b (n c) h w', n=3)

        # cast to float32 for better accuracy
        x = x.to(torch.float32).contiguous()
        # st()

        return x


class DiT_TriLatent_PixelArt(DiT_TriLatent):
    # DiT with 3D_aware operations
    def __init__(
        self,
        input_size=32,
        patch_size=2,
        in_channels=4,
        hidden_size=1152,
        depth=28,
        num_heads=16,
        mlp_ratio=4,
        class_dropout_prob=0.1,
        num_classes=1000,
        learn_sigma=True,
        mixing_logit_init=-3,
        mixed_prediction=True,
        context_dim=False,
        roll_out=False,
        vit_blk=DiTBlock,
        final_layer_blk=FinalLayer,
    ):
        super().__init__(input_size, patch_size, in_channels, hidden_size,
                         depth, num_heads, mlp_ratio, class_dropout_prob,
                         num_classes, learn_sigma, mixing_logit_init,
                         mixed_prediction, context_dim, roll_out, PixelArtTextCondDiTBlock,
                         final_layer_blk)

        self.adaLN_modulation = nn.Sequential(
            nn.SiLU(), nn.Linear(hidden_size, 6 * hidden_size, bias=True))
        del self.clip_text_proj
        self.cap_embedder = nn.Sequential( # TODO, init with zero here.
            LayerNorm(context_dim),
            nn.Linear(
                context_dim,
                hidden_size,
            ),
        )
        nn.init.constant_(self.cap_embedder[-1].weight, 0)
        nn.init.constant_(self.cap_embedder[-1].bias, 0)


    def forward(self,
                x,
                timesteps=None,
                context=None,
                y=None,
                get_attr='',
                **kwargs):
        """
        Forward pass of DiT.
        x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
        t: (N,) tensor of diffusion timesteps
        y: (N,) tensor of class labels
        """
        # t = timesteps
        assert context is not None

        clip_cls_token = self.cap_embedder(context['vector']) # pooled
        t = self.t_embedder(timesteps) + clip_cls_token  # (N, D)
        t0 = self.adaLN_modulation(t) # single-adaLN, B 6144

        # if self.roll_out:  # !
        x = rearrange(x, 'b (c n) h w->(b n) c h w',
                      n=3)  # downsample with same conv
        x = self.x_embedder(x)  # (b n) c h/f w/f

        x = rearrange(x, '(b n) l c -> b (n l) c', n=3)
        x = x + self.pos_embed  # (N, T, D), where T = H * W / patch_size ** 2

        # if self.roll_out:  # ! roll-out in the L dim, not B dim. add condition to all tokens.
        # x = rearrange(x, '(b n) l c ->b (n l) c', n=3)

        # assert context.ndim == 2
        if isinstance(context, dict):
            context = context['crossattn']  # sgm conditioner compat

        # context = self.clip_text_proj(context) # ! with rmsnorm here for 

        # c = t + context
        # else:
        # c = t  # BS 1024

        for blk_idx, block in enumerate(self.blocks):
            x = block(x, t0, context)  # (N, T, D)

        # todo later
        x = self.final_layer(x, t)  # (N, T, patch_size ** 2 * out_channels)

        if self.roll_out:  # move n from L to B axis
            x = rearrange(x, 'b (n l) c ->(b n) l c', n=3)

        x = self.unpatchify(x)  # (N, out_channels, H, W)

        if self.roll_out:  # move n from L to B axis
            x = rearrange(x, '(b n) c h w -> b (c n) h w', n=3)
            # x = rearrange(x, 'b n) c h w -> b (n c) h w', n=3)

        # cast to float32 for better accuracy
        x = x.to(torch.float32).contiguous()
        # st()

        return x

    # ! compat issue
    def forward_with_cfg(self, x, t, context, cfg_scale):
        """
        Forward pass of SiT, but also batches the unconSiTional forward pass for classifier-free guidance.
        """
        # https://github.com/openai/glide-text2im/blob/main/notebooks/text2im.ipynb
        # half = x[: len(x) // 2]
        # combined = torch.cat([half, half], dim=0)
        eps = self.forward(x, t, context)
        # eps, rest = model_out[:, :self.in_channels], model_out[:, self.in_channels:]
        # eps, rest = model_out[:, :3], model_out[:, 3:]
        cond_eps, uncond_eps = torch.split(eps, len(eps) // 2, dim=0)
        half_eps = uncond_eps + cfg_scale * (cond_eps - uncond_eps)
        eps = torch.cat([half_eps, half_eps], dim=0)
        return eps



#################################################################################
#                                   DiT_TriLatent Configs                                  #
#################################################################################


def DiT_XL_2(**kwargs):
    return DiT_TriLatent(depth=28,
                         hidden_size=1152,
                         patch_size=2,
                         num_heads=16,
                         **kwargs)


def DiT_L_2(**kwargs):
    return DiT_TriLatent(depth=24,
                         hidden_size=1024,
                         patch_size=2,
                         num_heads=16,
                         **kwargs)


def DiT_B_2(**kwargs):
    return DiT_TriLatent(depth=12,
                         hidden_size=768,
                         patch_size=2,
                         num_heads=12,
                         **kwargs)


def DiT_B_1(**kwargs):
    return DiT_TriLatent(depth=12,
                         hidden_size=768,
                         patch_size=1,
                         num_heads=12,
                         **kwargs)
def DiT_B_Pixelart_2(**kwargs):
    return DiT_TriLatent_PixelArt(depth=12,
                         hidden_size=768,
                         patch_size=2,
                         num_heads=12,
                        #  vit_blk=PixelArtTextCondDiTBlock,
                         final_layer_blk=T2IFinalLayer,
                         **kwargs)

def DiT_L_Pixelart_2(**kwargs):
    return DiT_TriLatent_PixelArt(depth=24,
                         hidden_size=1024,
                         patch_size=2,
                         num_heads=16,
                        #  vit_blk=PixelArtTextCondDiTBlock,
                         final_layer_blk=T2IFinalLayer,
                         **kwargs)

DiT_models = {
    'DiT-XL/2': DiT_XL_2,
    'DiT-L/2': DiT_L_2,
    'DiT-PixelArt-L/2': DiT_L_Pixelart_2,
    'DiT-PixelArt-B/2': DiT_B_Pixelart_2,
    'DiT-B/2': DiT_B_2,
    'DiT-B/1': DiT_B_1,
}