File size: 22,914 Bytes
ddc8a59
3304f7d
 
 
 
ddc8a59
 
3304f7d
 
 
 
ddc8a59
 
 
 
 
 
 
 
 
 
3304f7d
ddc8a59
 
 
 
3304f7d
 
 
 
 
 
 
ddc8a59
 
 
 
 
 
3304f7d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
ddc8a59
 
 
3304f7d
 
 
 
 
 
ddc8a59
 
3304f7d
 
 
 
 
 
ddc8a59
 
 
 
 
 
 
 
 
3304f7d
ddc8a59
 
 
 
 
 
 
 
 
 
 
 
3304f7d
 
 
 
 
 
ddc8a59
3304f7d
ddc8a59
3304f7d
 
 
ddc8a59
 
3304f7d
 
 
 
 
 
ddc8a59
3304f7d
 
 
 
 
 
ddc8a59
3304f7d
 
 
 
 
 
 
ddc8a59
 
 
 
 
3304f7d
 
 
 
 
 
ddc8a59
 
 
3304f7d
 
 
 
 
 
 
 
 
 
 
ddc8a59
 
 
 
 
 
 
3304f7d
 
ddc8a59
 
 
 
 
 
 
3304f7d
 
 
 
 
 
 
 
ddc8a59
3304f7d
 
 
 
 
 
 
 
 
 
ddc8a59
3304f7d
 
 
 
ddc8a59
3304f7d
 
 
 
 
 
 
 
ddc8a59
 
 
3304f7d
 
 
 
 
 
 
ddc8a59
 
3304f7d
 
 
 
 
 
ddc8a59
3304f7d
 
 
 
 
 
ddc8a59
 
 
3304f7d
 
ddc8a59
3304f7d
ddc8a59
3304f7d
ddc8a59
 
3304f7d
 
 
 
 
 
 
 
ddc8a59
3304f7d
 
 
 
 
 
 
 
 
 
ddc8a59
3304f7d
 
 
 
ddc8a59
3304f7d
 
 
 
 
 
 
 
ddc8a59
 
 
3304f7d
 
 
 
 
 
ddc8a59
 
 
3304f7d
 
 
 
 
 
ddc8a59
3304f7d
 
 
 
 
 
ddc8a59
3304f7d
ddc8a59
3304f7d
 
ddc8a59
 
 
 
 
 
 
 
3304f7d
 
 
 
 
 
 
 
ddc8a59
3304f7d
 
 
 
 
 
 
 
 
 
ddc8a59
3304f7d
 
 
 
ddc8a59
3304f7d
 
 
 
 
 
 
 
ddc8a59
 
 
3304f7d
 
 
 
 
 
 
ddc8a59
 
3304f7d
 
 
 
 
 
ddc8a59
3304f7d
 
 
 
 
 
 
ddc8a59
 
 
 
 
 
 
 
 
 
 
3304f7d
ddc8a59
 
3304f7d
 
 
 
 
 
ddc8a59
3304f7d
 
 
 
 
 
ddc8a59
 
 
3304f7d
 
 
 
 
ddc8a59
 
 
3304f7d
 
 
 
 
 
ddc8a59
 
 
 
 
 
3304f7d
 
 
 
 
 
ddc8a59
3304f7d
 
 
 
 
 
ddc8a59
 
 
3304f7d
 
 
 
 
ddc8a59
 
 
3304f7d
 
 
 
 
 
ddc8a59
 
3304f7d
 
 
 
 
 
ddc8a59
 
 
 
 
3304f7d
 
 
 
 
 
ddc8a59
3304f7d
 
 
 
 
 
ddc8a59
 
 
3304f7d
 
 
 
 
ddc8a59
 
 
3304f7d
 
 
 
 
 
 
ddc8a59
 
3304f7d
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
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
from itertools import product
from typing import Dict

import tensorflow as tf
import torch
from keras_cv.models import stable_diffusion


def port_transformer_block(
    transformer_block: tf.keras.Model, up_down: int, block_id: int, attention_id: int
) -> Dict[str, torch.Tensor]:
    """Populates a Transformer block."""
    transformer_dict = dict()
    if block_id is not None:
        prefix = f"{up_down}_blocks.{block_id}"
    else:
        prefix = "mid_block"

    # Norms.
    for i in range(1, 4):
        if i == 1:
            norm = transformer_block.norm1
        elif i == 2:
            norm = transformer_block.norm2
        elif i == 3:
            norm = transformer_block.norm3
        transformer_dict[
            f"{prefix}.attentions.{attention_id}.transformer_blocks.0.norm{i}.weight"
        ] = torch.from_numpy(norm.get_weights()[0])
        transformer_dict[
            f"{prefix}.attentions.{attention_id}.transformer_blocks.0.norm{i}.bias"
        ] = torch.from_numpy(norm.get_weights()[1])

    # Attentions.
    for i in range(1, 3):
        if i == 1:
            attn = transformer_block.attn1
        else:
            attn = transformer_block.attn2
        transformer_dict[
            f"{prefix}.attentions.{attention_id}.transformer_blocks.0.attn{i}.to_q.weight"
        ] = torch.from_numpy(attn.to_q.get_weights()[0].transpose())
        transformer_dict[
            f"{prefix}.attentions.{attention_id}.transformer_blocks.0.attn{i}.to_k.weight"
        ] = torch.from_numpy(attn.to_k.get_weights()[0].transpose())
        transformer_dict[
            f"{prefix}.attentions.{attention_id}.transformer_blocks.0.attn{i}.to_v.weight"
        ] = torch.from_numpy(attn.to_v.get_weights()[0].transpose())
        transformer_dict[
            f"{prefix}.attentions.{attention_id}.transformer_blocks.0.attn{i}.to_out.0.weight"
        ] = torch.from_numpy(attn.out_proj.get_weights()[0].transpose())
        transformer_dict[
            f"{prefix}.attentions.{attention_id}.transformer_blocks.0.attn{i}.to_out.0.bias"
        ] = torch.from_numpy(attn.out_proj.get_weights()[1])

    # Dense.
    for i in range(0, 3, 2):
        if i == 0:
            layer = transformer_block.geglu.dense
            transformer_dict[
                f"{prefix}.attentions.{attention_id}.transformer_blocks.0.ff.net.{i}.proj.weight"
            ] = torch.from_numpy(layer.get_weights()[0].transpose())
            transformer_dict[
                f"{prefix}.attentions.{attention_id}.transformer_blocks.0.ff.net.{i}.proj.bias"
            ] = torch.from_numpy(layer.get_weights()[1])
        else:
            layer = transformer_block.dense
            transformer_dict[
                f"{prefix}.attentions.{attention_id}.transformer_blocks.0.ff.net.{i}.weight"
            ] = torch.from_numpy(layer.get_weights()[0].transpose())
            transformer_dict[
                f"{prefix}.attentions.{attention_id}.transformer_blocks.0.ff.net.{i}.bias"
            ] = torch.from_numpy(layer.get_weights()[1])

    return transformer_dict


def populate_unet(tf_unet: tf.keras.Model) -> Dict[str, torch.Tensor]:
    """Populates the state dict from the provided TensorFlow model
    (applicable only for the UNet)."""
    unet_state_dict = dict()

    timstep_emb = 1
    padded_conv = 1
    up_block = 0

    up_res_blocks = list(product([0, 1, 2, 3], [0, 1, 2]))
    up_res_block_flag = 0

    up_spatial_transformer_blocks = list(product([1, 2, 3], [0, 1, 2]))
    up_spatial_transformer_flag = 0

    for layer in tf_unet.layers:
        # Timstep embedding.
        if isinstance(layer, tf.keras.layers.Dense):
            unet_state_dict[
                f"time_embedding.linear_{timstep_emb}.weight"
            ] = torch.from_numpy(layer.get_weights()[0].transpose())
            unet_state_dict[
                f"time_embedding.linear_{timstep_emb}.bias"
            ] = torch.from_numpy(layer.get_weights()[1])
            timstep_emb += 1

        # Padded convs (downsamplers).
        elif isinstance(
            layer, stable_diffusion.__internal__.layers.padded_conv2d.PaddedConv2D
        ):
            if padded_conv == 1:
                # Transposition axes taken from https://github.com/huggingface/transformers/blob/main/src/transformers/modeling_tf_pytorch_utils.py#L104
                unet_state_dict["conv_in.weight"] = torch.from_numpy(
                    layer.get_weights()[0].transpose(3, 2, 0, 1)
                )
                unet_state_dict["conv_in.bias"] = torch.from_numpy(
                    layer.get_weights()[1]
                )
            elif padded_conv in [2, 3, 4]:
                unet_state_dict[
                    f"down_blocks.{padded_conv-2}.downsamplers.0.conv.weight"
                ] = torch.from_numpy(layer.get_weights()[0].transpose(3, 2, 0, 1))
                unet_state_dict[
                    f"down_blocks.{padded_conv-2}.downsamplers.0.conv.bias"
                ] = torch.from_numpy(layer.get_weights()[1])
            elif padded_conv == 5:
                unet_state_dict["conv_out.weight"] = torch.from_numpy(
                    layer.get_weights()[0].transpose(3, 2, 0, 1)
                )
                unet_state_dict["conv_out.bias"] = torch.from_numpy(
                    layer.get_weights()[1]
                )

            padded_conv += 1

        # Upsamplers.
        elif isinstance(layer, stable_diffusion.diffusion_model.Upsample):
            conv = layer.conv
            unet_state_dict[
                f"up_blocks.{up_block}.upsamplers.0.conv.weight"
            ] = torch.from_numpy(conv.get_weights()[0].transpose(3, 2, 0, 1))
            unet_state_dict[
                f"up_blocks.{up_block}.upsamplers.0.conv.bias"
            ] = torch.from_numpy(conv.get_weights()[1])
            up_block += 1

        # Output norms.
        elif isinstance(
            layer,
            stable_diffusion.__internal__.layers.group_normalization.GroupNormalization,
        ):
            unet_state_dict["conv_norm_out.weight"] = torch.from_numpy(
                layer.get_weights()[0]
            )
            unet_state_dict["conv_norm_out.bias"] = torch.from_numpy(
                layer.get_weights()[1]
            )

        # All ResBlocks.
        elif isinstance(layer, stable_diffusion.diffusion_model.ResBlock):
            layer_name = layer.name
            parts = layer_name.split("_")

            # Down.
            if len(parts) == 2 or int(parts[-1]) < 8:
                entry_flow = layer.entry_flow
                embedding_flow = layer.embedding_flow
                exit_flow = layer.exit_flow

                down_block_id = 0 if len(parts) == 2 else int(parts[-1]) // 2
                down_resnet_id = 0 if len(parts) == 2 else int(parts[-1]) % 2

                # Conv blocks.
                first_conv_layer = entry_flow[-1]
                unet_state_dict[
                    f"down_blocks.{down_block_id}.resnets.{down_resnet_id}.conv1.weight"
                ] = torch.from_numpy(
                    first_conv_layer.get_weights()[0].transpose(3, 2, 0, 1)
                )
                unet_state_dict[
                    f"down_blocks.{down_block_id}.resnets.{down_resnet_id}.conv1.bias"
                ] = torch.from_numpy(first_conv_layer.get_weights()[1])
                second_conv_layer = exit_flow[-1]
                unet_state_dict[
                    f"down_blocks.{down_block_id}.resnets.{down_resnet_id}.conv2.weight"
                ] = torch.from_numpy(
                    second_conv_layer.get_weights()[0].transpose(3, 2, 0, 1)
                )
                unet_state_dict[
                    f"down_blocks.{down_block_id}.resnets.{down_resnet_id}.conv2.bias"
                ] = torch.from_numpy(second_conv_layer.get_weights()[1])

                # Residual blocks.
                if hasattr(layer, "residual_projection"):
                    if isinstance(
                        layer.residual_projection,
                        stable_diffusion.__internal__.layers.padded_conv2d.PaddedConv2D,
                    ):
                        residual = layer.residual_projection
                        unet_state_dict[
                            f"down_blocks.{down_block_id}.resnets.{down_resnet_id}.conv_shortcut.weight"
                        ] = torch.from_numpy(
                            residual.get_weights()[0].transpose(3, 2, 0, 1)
                        )
                        unet_state_dict[
                            f"down_blocks.{down_block_id}.resnets.{down_resnet_id}.conv_shortcut.bias"
                        ] = torch.from_numpy(residual.get_weights()[1])

                # Timestep embedding.
                embedding_proj = embedding_flow[-1]
                unet_state_dict[
                    f"down_blocks.{down_block_id}.resnets.{down_resnet_id}.time_emb_proj.weight"
                ] = torch.from_numpy(embedding_proj.get_weights()[0].transpose())
                unet_state_dict[
                    f"down_blocks.{down_block_id}.resnets.{down_resnet_id}.time_emb_proj.bias"
                ] = torch.from_numpy(embedding_proj.get_weights()[1])

                # Norms.
                first_group_norm = entry_flow[0]
                unet_state_dict[
                    f"down_blocks.{down_block_id}.resnets.{down_resnet_id}.norm1.weight"
                ] = torch.from_numpy(first_group_norm.get_weights()[0])
                unet_state_dict[
                    f"down_blocks.{down_block_id}.resnets.{down_resnet_id}.norm1.bias"
                ] = torch.from_numpy(first_group_norm.get_weights()[1])
                second_group_norm = exit_flow[0]
                unet_state_dict[
                    f"down_blocks.{down_block_id}.resnets.{down_resnet_id}.norm2.weight"
                ] = torch.from_numpy(second_group_norm.get_weights()[0])
                unet_state_dict[
                    f"down_blocks.{down_block_id}.resnets.{down_resnet_id}.norm2.bias"
                ] = torch.from_numpy(second_group_norm.get_weights()[1])

            # Middle.
            elif int(parts[-1]) == 8 or int(parts[-1]) == 9:
                entry_flow = layer.entry_flow
                embedding_flow = layer.embedding_flow
                exit_flow = layer.exit_flow

                mid_resnet_id = int(parts[-1]) % 2

                # Conv blocks.
                first_conv_layer = entry_flow[-1]
                unet_state_dict[
                    f"mid_block.resnets.{mid_resnet_id}.conv1.weight"
                ] = torch.from_numpy(
                    first_conv_layer.get_weights()[0].transpose(3, 2, 0, 1)
                )
                unet_state_dict[
                    f"mid_block.resnets.{mid_resnet_id}.conv1.bias"
                ] = torch.from_numpy(first_conv_layer.get_weights()[1])
                second_conv_layer = exit_flow[-1]
                unet_state_dict[
                    f"mid_block.resnets.{mid_resnet_id}.conv2.weight"
                ] = torch.from_numpy(
                    second_conv_layer.get_weights()[0].transpose(3, 2, 0, 1)
                )
                unet_state_dict[
                    f"mid_block.resnets.{mid_resnet_id}.conv2.bias"
                ] = torch.from_numpy(second_conv_layer.get_weights()[1])

                # Residual blocks.
                if hasattr(layer, "residual_projection"):
                    if isinstance(
                        layer.residual_projection,
                        stable_diffusion.__internal__.layers.padded_conv2d.PaddedConv2D,
                    ):
                        residual = layer.residual_projection
                        unet_state_dict[
                            f"mid_block.resnets.{mid_resnet_id}.conv_shortcut.weight"
                        ] = torch.from_numpy(
                            residual.get_weights()[0].transpose(3, 2, 0, 1)
                        )
                        unet_state_dict[
                            f"mid_block.resnets.{mid_resnet_id}.conv_shortcut.bias"
                        ] = torch.from_numpy(residual.get_weights()[1])

                # Timestep embedding.
                embedding_proj = embedding_flow[-1]
                unet_state_dict[
                    f"mid_block.resnets.{mid_resnet_id}.time_emb_proj.weight"
                ] = torch.from_numpy(embedding_proj.get_weights()[0].transpose())
                unet_state_dict[
                    f"mid_block.resnets.{mid_resnet_id}.time_emb_proj.bias"
                ] = torch.from_numpy(embedding_proj.get_weights()[1])

                # Norms.
                first_group_norm = entry_flow[0]
                unet_state_dict[
                    f"mid_block.resnets.{mid_resnet_id}.norm1.weight"
                ] = torch.from_numpy(first_group_norm.get_weights()[0])
                unet_state_dict[
                    f"mid_block.resnets.{mid_resnet_id}.norm1.bias"
                ] = torch.from_numpy(first_group_norm.get_weights()[1])
                second_group_norm = exit_flow[0]
                unet_state_dict[
                    f"mid_block.resnets.{mid_resnet_id}.norm2.weight"
                ] = torch.from_numpy(second_group_norm.get_weights()[0])
                unet_state_dict[
                    f"mid_block.resnets.{mid_resnet_id}.norm2.bias"
                ] = torch.from_numpy(second_group_norm.get_weights()[1])

            # Up.
            elif int(parts[-1]) > 9 and up_res_block_flag < len(up_res_blocks):
                entry_flow = layer.entry_flow
                embedding_flow = layer.embedding_flow
                exit_flow = layer.exit_flow

                up_res_block = up_res_blocks[up_res_block_flag]
                up_block_id = up_res_block[0]
                up_resnet_id = up_res_block[1]

                # Conv blocks.
                first_conv_layer = entry_flow[-1]
                unet_state_dict[
                    f"up_blocks.{up_block_id}.resnets.{up_resnet_id}.conv1.weight"
                ] = torch.from_numpy(
                    first_conv_layer.get_weights()[0].transpose(3, 2, 0, 1)
                )
                unet_state_dict[
                    f"up_blocks.{up_block_id}.resnets.{up_resnet_id}.conv1.bias"
                ] = torch.from_numpy(first_conv_layer.get_weights()[1])
                second_conv_layer = exit_flow[-1]
                unet_state_dict[
                    f"up_blocks.{up_block_id}.resnets.{up_resnet_id}.conv2.weight"
                ] = torch.from_numpy(
                    second_conv_layer.get_weights()[0].transpose(3, 2, 0, 1)
                )
                unet_state_dict[
                    f"up_blocks.{up_block_id}.resnets.{up_resnet_id}.conv2.bias"
                ] = torch.from_numpy(second_conv_layer.get_weights()[1])

                # Residual blocks.
                if hasattr(layer, "residual_projection"):
                    if isinstance(
                        layer.residual_projection,
                        stable_diffusion.__internal__.layers.padded_conv2d.PaddedConv2D,
                    ):
                        residual = layer.residual_projection
                        unet_state_dict[
                            f"up_blocks.{up_block_id}.resnets.{up_resnet_id}.conv_shortcut.weight"
                        ] = torch.from_numpy(
                            residual.get_weights()[0].transpose(3, 2, 0, 1)
                        )
                        unet_state_dict[
                            f"up_blocks.{up_block_id}.resnets.{up_resnet_id}.conv_shortcut.bias"
                        ] = torch.from_numpy(residual.get_weights()[1])

                # Timestep embedding.
                embedding_proj = embedding_flow[-1]
                unet_state_dict[
                    f"up_blocks.{up_block_id}.resnets.{up_resnet_id}.time_emb_proj.weight"
                ] = torch.from_numpy(embedding_proj.get_weights()[0].transpose())
                unet_state_dict[
                    f"up_blocks.{up_block_id}.resnets.{up_resnet_id}.time_emb_proj.bias"
                ] = torch.from_numpy(embedding_proj.get_weights()[1])

                # Norms.
                first_group_norm = entry_flow[0]
                unet_state_dict[
                    f"up_blocks.{up_block_id}.resnets.{up_resnet_id}.norm1.weight"
                ] = torch.from_numpy(first_group_norm.get_weights()[0])
                unet_state_dict[
                    f"up_blocks.{up_block_id}.resnets.{up_resnet_id}.norm1.bias"
                ] = torch.from_numpy(first_group_norm.get_weights()[1])
                second_group_norm = exit_flow[0]
                unet_state_dict[
                    f"up_blocks.{up_block_id}.resnets.{up_resnet_id}.norm2.weight"
                ] = torch.from_numpy(second_group_norm.get_weights()[0])
                unet_state_dict[
                    f"up_blocks.{up_block_id}.resnets.{up_resnet_id}.norm2.bias"
                ] = torch.from_numpy(second_group_norm.get_weights()[1])

                up_res_block_flag += 1

        # All SpatialTransformer blocks.
        elif isinstance(layer, stable_diffusion.diffusion_model.SpatialTransformer):
            layer_name = layer.name
            parts = layer_name.split("_")

            # Down.
            if len(parts) == 2 or int(parts[-1]) < 6:
                down_block_id = 0 if len(parts) == 2 else int(parts[-1]) // 2
                down_attention_id = 0 if len(parts) == 2 else int(parts[-1]) % 2

                # Convs.
                proj1 = layer.proj1
                unet_state_dict[
                    f"down_blocks.{down_block_id}.attentions.{down_attention_id}.proj_in.weight"
                ] = torch.from_numpy(proj1.get_weights()[0].transpose(3, 2, 0, 1))
                unet_state_dict[
                    f"down_blocks.{down_block_id}.attentions.{down_attention_id}.proj_in.bias"
                ] = torch.from_numpy(proj1.get_weights()[1])
                proj2 = layer.proj2
                unet_state_dict[
                    f"down_blocks.{down_block_id}.attentions.{down_attention_id}.proj_out.weight"
                ] = torch.from_numpy(proj2.get_weights()[0].transpose(3, 2, 0, 1))
                unet_state_dict[
                    f"down_blocks.{down_block_id}.attentions.{down_attention_id}.proj_out.bias"
                ] = torch.from_numpy(proj2.get_weights()[1])

                # Transformer blocks.
                transformer_block = layer.transformer_block
                unet_state_dict.update(
                    port_transformer_block(
                        transformer_block, "down", down_block_id, down_attention_id
                    )
                )

                # Norms.
                norm = layer.norm
                unet_state_dict[
                    f"down_blocks.{down_block_id}.attentions.{down_attention_id}.norm.weight"
                ] = torch.from_numpy(norm.get_weights()[0])
                unet_state_dict[
                    f"down_blocks.{down_block_id}.attentions.{down_attention_id}.norm.bias"
                ] = torch.from_numpy(norm.get_weights()[1])

            # Middle.
            elif int(parts[-1]) == 6:
                mid_attention_id = int(parts[-1]) % 2
                # Convs.
                proj1 = layer.proj1
                unet_state_dict[
                    f"mid_block.attentions.{mid_attention_id}.proj_in.weight"
                ] = torch.from_numpy(proj1.get_weights()[0].transpose(3, 2, 0, 1))
                unet_state_dict[
                    f"mid_block.attentions.{mid_attention_id}.proj_in.bias"
                ] = torch.from_numpy(proj1.get_weights()[1])
                proj2 = layer.proj2
                unet_state_dict[
                    f"mid_block.attentions.{mid_resnet_id}.proj_out.weight"
                ] = torch.from_numpy(proj2.get_weights()[0].transpose(3, 2, 0, 1))
                unet_state_dict[
                    f"mid_block.attentions.{mid_attention_id}.proj_out.bias"
                ] = torch.from_numpy(proj2.get_weights()[1])

                # Transformer blocks.
                transformer_block = layer.transformer_block
                unet_state_dict.update(
                    port_transformer_block(
                        transformer_block, "mid", None, mid_attention_id
                    )
                )

                # Norms.
                norm = layer.norm
                unet_state_dict[
                    f"mid_block.attentions.{mid_attention_id}.norm.weight"
                ] = torch.from_numpy(norm.get_weights()[0])
                unet_state_dict[
                    f"mid_block.attentions.{mid_attention_id}.norm.bias"
                ] = torch.from_numpy(norm.get_weights()[1])

            # Up.
            elif int(parts[-1]) > 6 and up_spatial_transformer_flag < len(
                up_spatial_transformer_blocks
            ):
                up_spatial_transformer_block = up_spatial_transformer_blocks[
                    up_spatial_transformer_flag
                ]
                up_block_id = up_spatial_transformer_block[0]
                up_attention_id = up_spatial_transformer_block[1]

                # Convs.
                proj1 = layer.proj1
                unet_state_dict[
                    f"up_blocks.{up_block_id}.attentions.{up_attention_id}.proj_in.weight"
                ] = torch.from_numpy(proj1.get_weights()[0].transpose(3, 2, 0, 1))
                unet_state_dict[
                    f"up_blocks.{up_block_id}.attentions.{up_attention_id}.proj_in.bias"
                ] = torch.from_numpy(proj1.get_weights()[1])
                proj2 = layer.proj2
                unet_state_dict[
                    f"up_blocks.{up_block_id}.attentions.{up_attention_id}.proj_out.weight"
                ] = torch.from_numpy(proj2.get_weights()[0].transpose(3, 2, 0, 1))
                unet_state_dict[
                    f"up_blocks.{up_block_id}.attentions.{up_attention_id}.proj_out.bias"
                ] = torch.from_numpy(proj2.get_weights()[1])

                # Transformer blocks.
                transformer_block = layer.transformer_block
                unet_state_dict.update(
                    port_transformer_block(
                        transformer_block, "up", up_block_id, up_attention_id
                    )
                )

                # Norms.
                norm = layer.norm
                unet_state_dict[
                    f"up_blocks.{up_block_id}.attentions.{up_attention_id}.norm.weight"
                ] = torch.from_numpy(norm.get_weights()[0])
                unet_state_dict[
                    f"up_blocks.{up_block_id}.attentions.{up_attention_id}.norm.bias"
                ] = torch.from_numpy(norm.get_weights()[1])

                up_spatial_transformer_flag += 1

    return unet_state_dict