File size: 10,872 Bytes
7f51798
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
import torch
import torch.nn as nn
import numpy as np
import math

from einops import rearrange
from pdb import set_trace as st

# from .dit_models import DiT, DiTBlock, DiT_models, get_2d_sincos_pos_embed, modulate, FinalLayer

from .dit_models_xformers import DiT, DiTBlock, DiT_models, get_2d_sincos_pos_embed, modulate, FinalLayer
# from .dit_models import DiT, DiTBlock, DiT_models, get_2d_sincos_pos_embed, modulate, FinalLayer


def modulate2(x, shift, scale):
    return x * (1 + scale) + shift


class DiTBlock2(DiTBlock):
    """
    A DiT block with adaptive layer norm zero (adaLN-Zero) conditioning.
    """

    def __init__(self, hidden_size, num_heads, mlp_ratio=4, **block_kwargs):
        super().__init__(hidden_size, num_heads, mlp_ratio, **block_kwargs)

    def forward(self, x, c):
        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(
            c).chunk(6, dim=-1)
        # st()
        x = x + gate_msa * self.attn(
            modulate2(self.norm1(x), shift_msa, scale_msa))
        x = x + gate_mlp * self.mlp(
            modulate2(self.norm2(x), shift_mlp, scale_mlp))
        return x


class FinalLayer2(FinalLayer):
    """
    The final layer of DiT, basically the decoder_pred in MAE with adaLN.
    """

    def __init__(self, hidden_size, patch_size, out_channels):
        super().__init__(hidden_size, patch_size, out_channels)

    def forward(self, x, c):
        shift, scale = self.adaLN_modulation(c).chunk(2, dim=-1)
        x = modulate2(self.norm_final(x), shift, scale)
        x = self.linear(x)
        return x


class DiT2(DiT):
    # a conditional ViT
    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,
                 plane_n=3,
                 return_all_layers=False,
                 in_plane_attention=True,
                 vit_blk=...):
        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=DiTBlock2,
                         final_layer_blk=FinalLayer2)
        # st()

        # no t and x embedder
        del self.x_embedder
        del self.t_embedder
        del self.final_layer
        torch.cuda.empty_cache()
        self.clip_text_proj = None
        self.plane_n = plane_n
        self.return_all_layers = return_all_layers
        self.in_plane_attention = in_plane_attention

    def forward(self, c, *args, **kwargs):
        # return super().forward(x, timesteps, context, y, get_attr, **kwargs)
        """
        Forward pass of DiT.
        c: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
        """

        x = self.pos_embed.repeat(
            c.shape[0], 1, 1).to(c.dtype)  # (N, T, D), where T = H * W / patch_size ** 2

        if self.return_all_layers:
            all_layers = []

        # if context is not None:
        # c = context # B 3HW C

        for blk_idx, block in enumerate(self.blocks):
            if self.roll_out:
                if self.in_plane_attention: # plane-wise output
                    if blk_idx % 2 == 0:  # with-in plane self attention
                        x = rearrange(x, 'b (n l) c -> (b n) l c ', n=self.plane_n)
                        x = block(x,
                                rearrange(c,
                                            'b (n l) c -> (b n) l c ',
                                            n=self.plane_n))  # (N, T, D)
                        # st()
                        if self.return_all_layers:

                            all_layers.append(
                                rearrange(x,
                                        '(b n) l c -> b (n l) c',
                                        n=self.plane_n))

                            # all_layers.append(x)
                    else:  # global attention
                        x = rearrange(x, '(b n) l c -> b (n l) c ', n=self.plane_n)
                        x = block(x, c)  # (N, T, D)
                        # st()
                        if self.return_all_layers:
                            #  all merged into B dim
                            all_layers.append(x)
                            # all_layers.append(
                            #     rearrange(x,
                            #               'b (n l) c -> (b n) l c',
                            #               n=self.plane_n))
                else:

                    # ! already b (n l) c
                    # if blk_idx == 0: # rearrange once
                    #     x = rearrange(x, '(b n) l c -> b (n l) c ', n=self.plane_n)
                    x = block(x, c)  # (N, T, D)
                    if self.return_all_layers:
                        #  all merged into B dim
                        all_layers.append(x)

            else:
                x = block(x, c)  # (N, T, D)

        # x = self.final_layer(x, c)  # (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 (n c) h w', n=3)
        # st()

        if self.return_all_layers:
            return all_layers
        else:
            # return x.to(torch.float32)
            return x


# class DiT2_DPT(DiT2):
#     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, plane_n=3, vit_blk=...):
#         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, plane_n, vit_blk)
#         self.return_all_layers = True

#################################################################################
#                                   DiT2 Configs                                  #
#################################################################################


def DiT2_XL_2(**kwargs):
    return DiT2(depth=28,
                hidden_size=1152,
                patch_size=2,
                num_heads=16,
                **kwargs)


def DiT2_XL_2_half(**kwargs):
    return DiT2(depth=28 // 2,
                hidden_size=1152,
                patch_size=2,
                num_heads=16,
                **kwargs)


def DiT2_XL_4(**kwargs):
    return DiT2(depth=28,
                hidden_size=1152,
                patch_size=4,
                num_heads=16,
                **kwargs)


def DiT2_XL_8(**kwargs):
    return DiT2(depth=28,
                hidden_size=1152,
                patch_size=8,
                num_heads=16,
                **kwargs)


def DiT2_L_2(**kwargs):
    return DiT2(depth=24,
                hidden_size=1024,
                patch_size=2,
                num_heads=16,
                **kwargs)


def DiT2_L_2_stage1(**kwargs):
    return DiT2(depth=24-6,
                hidden_size=1024,
                patch_size=2,
                num_heads=16,
                **kwargs)

def DiT2_L_2_stage1(**kwargs):
    return DiT2(depth=24-6,
                hidden_size=1024,
                patch_size=2,
                num_heads=16,
                **kwargs)

def DiT2_L_2_half(**kwargs):
    return DiT2(depth=24//2,
                hidden_size=1024,
                patch_size=2,
                num_heads=16,
                **kwargs)

def DiT2_L_2_half_ninelayer(**kwargs):
    return DiT2(depth=9,
                hidden_size=1024,
                patch_size=2,
                num_heads=16,
                **kwargs)



def DiT2_L_4(**kwargs):
    return DiT2(depth=24,
                hidden_size=1024,
                patch_size=4,
                num_heads=16,
                **kwargs)


def DiT2_L_8(**kwargs):
    return DiT2(depth=24,
                hidden_size=1024,
                patch_size=8,
                num_heads=16,
                **kwargs)


def DiT2_B_2(**kwargs):
    return DiT2(depth=12,
                hidden_size=768,
                patch_size=2,
                num_heads=12,
                **kwargs)


def DiT2_B_2_stage1(**kwargs):
    return DiT2(depth=12, # ! just 12, stage-2 3 layers afterwards.
                hidden_size=768,
                patch_size=2,
                num_heads=12,
                **kwargs)

def DiT2_B_4(**kwargs):
    return DiT2(depth=12,
                hidden_size=768,
                patch_size=4,
                num_heads=12,
                **kwargs)


def DiT2_B_8(**kwargs):
    return DiT2(depth=12,
                hidden_size=768,
                patch_size=8,
                num_heads=12,
                **kwargs)


def DiT2_B_16(**kwargs):  # ours cfg
    return DiT2(depth=12,
                hidden_size=768,
                patch_size=16,
                num_heads=12,
                **kwargs)


def DiT2_S_2(**kwargs):
    return DiT2(depth=12, hidden_size=384, patch_size=2, num_heads=6, **kwargs)


def DiT2_S_4(**kwargs):
    return DiT2(depth=12, hidden_size=384, patch_size=4, num_heads=6, **kwargs)


def DiT2_S_8(**kwargs):
    return DiT2(depth=12, hidden_size=384, patch_size=8, num_heads=6, **kwargs)


DiT2_models = {
    'DiT2-XL/2': DiT2_XL_2,
    'DiT2-XL/2/half': DiT2_XL_2_half,
    'DiT2-XL/4': DiT2_XL_4,
    'DiT2-XL/8': DiT2_XL_8,
    'DiT2-L/2': DiT2_L_2,
    'DiT2-L/2/S1': DiT2_L_2_stage1,
    'DiT2-L/2/S1-v2': DiT2_L_2_stage1,
    'DiT2-B/2/S1': DiT2_B_2_stage1,
    'DiT2-L/4': DiT2_L_4,
    'DiT2-L/2-half': DiT2_L_2_half,
    'DiT2-L/2-ninelayer': DiT2_L_2_half_ninelayer,
    'DiT2-L/8': DiT2_L_8,
    'DiT2-B/2': DiT2_B_2,
    'DiT2-B/4': DiT2_B_4,
    'DiT2-B/8': DiT2_B_8,
    'DiT2-B/16': DiT2_B_16,
    'DiT2-S/2': DiT2_S_2,
    'DiT2-S/4': DiT2_S_4,
    'DiT2-S/8': DiT2_S_8,
}