{{ message }}
Repository navigation
Expand file tree
/
Copy pathconvert_dcae_to_diffusers.py
More file actions
323 lines (289 loc) · 11.2 KB
/
Copy pathconvert_dcae_to_diffusers.py
File metadata and controls
323 lines (289 loc) · 11.2 KB
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 argparse
from typing import Any, Dict
import torch
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
from diffusers import AutoencoderDC
def remap_qkv_(key: str, state_dict: Dict[str, Any]):
qkv = state_dict.pop(key)
q, k, v = torch.chunk(qkv, 3, dim=0)
parent_module, _, _ = key.rpartition(".qkv.conv.weight")
state_dict[f"{parent_module}.to_q.weight"] = q.squeeze()
state_dict[f"{parent_module}.to_k.weight"] = k.squeeze()
state_dict[f"{parent_module}.to_v.weight"] = v.squeeze()
def remap_proj_conv_(key: str, state_dict: Dict[str, Any]):
parent_module, _, _ = key.rpartition(".proj.conv.weight")
state_dict[f"{parent_module}.to_out.weight"] = state_dict.pop(key).squeeze()
AE_KEYS_RENAME_DICT = {
# common
"main.": "",
"op_list.": "",
"context_module": "attn",
"local_module": "conv_out",
# NOTE: The below two lines work because scales in the available configs only have a tuple length of 1
# If there were more scales, there would be more layers, so a loop would be better to handle this
"aggreg.0.0": "to_qkv_multiscale.0.proj_in",
"aggreg.0.1": "to_qkv_multiscale.0.proj_out",
"depth_conv.conv": "conv_depth",
"inverted_conv.conv": "conv_inverted",
"point_conv.conv": "conv_point",
"point_conv.norm": "norm",
"conv.conv.": "conv.",
"conv1.conv": "conv1",
"conv2.conv": "conv2",
"conv2.norm": "norm",
"proj.norm": "norm_out",
# encoder
"encoder.project_in.conv": "encoder.conv_in",
"encoder.project_out.0.conv": "encoder.conv_out",
"encoder.stages": "encoder.down_blocks",
# decoder
"decoder.project_in.conv": "decoder.conv_in",
"decoder.project_out.0": "decoder.norm_out",
"decoder.project_out.2.conv": "decoder.conv_out",
"decoder.stages": "decoder.up_blocks",
}
AE_F32C32_KEYS = {
# encoder
"encoder.project_in.conv": "encoder.conv_in.conv",
# decoder
"decoder.project_out.2.conv": "decoder.conv_out.conv",
}
AE_F64C128_KEYS = {
# encoder
"encoder.project_in.conv": "encoder.conv_in.conv",
# decoder
"decoder.project_out.2.conv": "decoder.conv_out.conv",
}
AE_F128C512_KEYS = {
# encoder
"encoder.project_in.conv": "encoder.conv_in.conv",
# decoder
"decoder.project_out.2.conv": "decoder.conv_out.conv",
}
AE_SPECIAL_KEYS_REMAP = {
"qkv.conv.weight": remap_qkv_,
"proj.conv.weight": remap_proj_conv_,
}
def get_state_dict(saved_dict: Dict[str, Any]) -> Dict[str, Any]:
state_dict = saved_dict
if "model" in saved_dict.keys():
state_dict = state_dict["model"]
if "module" in saved_dict.keys():
state_dict = state_dict["module"]
if "state_dict" in saved_dict.keys():
state_dict = state_dict["state_dict"]
return state_dict
def update_state_dict_(state_dict: Dict[str, Any], old_key: str, new_key: str) -> Dict[str, Any]:
state_dict[new_key] = state_dict.pop(old_key)
def convert_ae(config_name: str, dtype: torch.dtype):
config = get_ae_config(config_name)
hub_id = f"mit-han-lab/{config_name}"
ckpt_path = hf_hub_download(hub_id, "model.safetensors")
original_state_dict = get_state_dict(load_file(ckpt_path))
ae = AutoencoderDC(**config).to(dtype=dtype)
for key in list(original_state_dict.keys()):
new_key = key[:]
for replace_key, rename_key in AE_KEYS_RENAME_DICT.items():
new_key = new_key.replace(replace_key, rename_key)
update_state_dict_(original_state_dict, key, new_key)
for key in list(original_state_dict.keys()):
for special_key, handler_fn_inplace in AE_SPECIAL_KEYS_REMAP.items():
if special_key not in key:
continue
handler_fn_inplace(key, original_state_dict)
ae.load_state_dict(original_state_dict, strict=True)
return ae
def get_ae_config(name: str):
if name in ["dc-ae-f32c32-sana-1.0"]:
config = {
"latent_channels": 32,
"encoder_block_types": (
"ResBlock",
"ResBlock",
"ResBlock",
"EfficientViTBlock",
"EfficientViTBlock",
"EfficientViTBlock",
),
"decoder_block_types": (
"ResBlock",
"ResBlock",
"ResBlock",
"EfficientViTBlock",
"EfficientViTBlock",
"EfficientViTBlock",
),
"encoder_block_out_channels": (128, 256, 512, 512, 1024, 1024),
"decoder_block_out_channels": (128, 256, 512, 512, 1024, 1024),
"encoder_qkv_multiscales": ((), (), (), (5,), (5,), (5,)),
"decoder_qkv_multiscales": ((), (), (), (5,), (5,), (5,)),
"encoder_layers_per_block": (2, 2, 2, 3, 3, 3),
"decoder_layers_per_block": [3, 3, 3, 3, 3, 3],
"downsample_block_type": "conv",
"upsample_block_type": "interpolate",
"decoder_norm_types": "rms_norm",
"decoder_act_fns": "silu",
"scaling_factor": 0.41407,
}
elif name in ["dc-ae-f32c32-in-1.0", "dc-ae-f32c32-mix-1.0"]:
AE_KEYS_RENAME_DICT.update(AE_F32C32_KEYS)
config = {
"latent_channels": 32,
"encoder_block_types": [
"ResBlock",
"ResBlock",
"ResBlock",
"EfficientViTBlock",
"EfficientViTBlock",
"EfficientViTBlock",
],
"decoder_block_types": [
"ResBlock",
"ResBlock",
"ResBlock",
"EfficientViTBlock",
"EfficientViTBlock",
"EfficientViTBlock",
],
"encoder_block_out_channels": [128, 256, 512, 512, 1024, 1024],
"decoder_block_out_channels": [128, 256, 512, 512, 1024, 1024],
"encoder_layers_per_block": [0, 4, 8, 2, 2, 2],
"decoder_layers_per_block": [0, 5, 10, 2, 2, 2],
"encoder_qkv_multiscales": ((), (), (), (), (), ()),
"decoder_qkv_multiscales": ((), (), (), (), (), ()),
"decoder_norm_types": ["batch_norm", "batch_norm", "batch_norm", "rms_norm", "rms_norm", "rms_norm"],
"decoder_act_fns": ["relu", "relu", "relu", "silu", "silu", "silu"],
}
if name == "dc-ae-f32c32-in-1.0":
config["scaling_factor"] = 0.3189
elif name == "dc-ae-f32c32-mix-1.0":
config["scaling_factor"] = 0.4552
elif name in ["dc-ae-f64c128-in-1.0", "dc-ae-f64c128-mix-1.0"]:
AE_KEYS_RENAME_DICT.update(AE_F64C128_KEYS)
config = {
"latent_channels": 128,
"encoder_block_types": [
"ResBlock",
"ResBlock",
"ResBlock",
"EfficientViTBlock",
"EfficientViTBlock",
"EfficientViTBlock",
"EfficientViTBlock",
],
"decoder_block_types": [
"ResBlock",
"ResBlock",
"ResBlock",
"EfficientViTBlock",
"EfficientViTBlock",
"EfficientViTBlock",
"EfficientViTBlock",
],
"encoder_block_out_channels": [128, 256, 512, 512, 1024, 1024, 2048],
"decoder_block_out_channels": [128, 256, 512, 512, 1024, 1024, 2048],
"encoder_layers_per_block": [0, 4, 8, 2, 2, 2, 2],
"decoder_layers_per_block": [0, 5, 10, 2, 2, 2, 2],
"encoder_qkv_multiscales": ((), (), (), (), (), (), ()),
"decoder_qkv_multiscales": ((), (), (), (), (), (), ()),
"decoder_norm_types": [
"batch_norm",
"batch_norm",
"batch_norm",
"rms_norm",
"rms_norm",
"rms_norm",
"rms_norm",
],
"decoder_act_fns": ["relu", "relu", "relu", "silu", "silu", "silu", "silu"],
}
if name == "dc-ae-f64c128-in-1.0":
config["scaling_factor"] = 0.2889
elif name == "dc-ae-f64c128-mix-1.0":
config["scaling_factor"] = 0.4538
elif name in ["dc-ae-f128c512-in-1.0", "dc-ae-f128c512-mix-1.0"]:
AE_KEYS_RENAME_DICT.update(AE_F128C512_KEYS)
config = {
"latent_channels": 512,
"encoder_block_types": [
"ResBlock",
"ResBlock",
"ResBlock",
"EfficientViTBlock",
"EfficientViTBlock",
"EfficientViTBlock",
"EfficientViTBlock",
"EfficientViTBlock",
],
"decoder_block_types": [
"ResBlock",
"ResBlock",
"ResBlock",
"EfficientViTBlock",
"EfficientViTBlock",
"EfficientViTBlock",
"EfficientViTBlock",
"EfficientViTBlock",
],
"encoder_block_out_channels": [128, 256, 512, 512, 1024, 1024, 2048, 2048],
"decoder_block_out_channels": [128, 256, 512, 512, 1024, 1024, 2048, 2048],
"encoder_layers_per_block": [0, 4, 8, 2, 2, 2, 2, 2],
"decoder_layers_per_block": [0, 5, 10, 2, 2, 2, 2, 2],
"encoder_qkv_multiscales": ((), (), (), (), (), (), (), ()),
"decoder_qkv_multiscales": ((), (), (), (), (), (), (), ()),
"decoder_norm_types": [
"batch_norm",
"batch_norm",
"batch_norm",
"rms_norm",
"rms_norm",
"rms_norm",
"rms_norm",
"rms_norm",
],
"decoder_act_fns": ["relu", "relu", "relu", "silu", "silu", "silu", "silu", "silu"],
}
if name == "dc-ae-f128c512-in-1.0":
config["scaling_factor"] = 0.4883
elif name == "dc-ae-f128c512-mix-1.0":
config["scaling_factor"] = 0.3620
else:
raise ValueError("Invalid config name provided.")
return config
def get_args():
parser = argparse.ArgumentParser()
parser.add_argument(
"--config_name",
type=str,
default="dc-ae-f32c32-sana-1.0",
choices=[
"dc-ae-f32c32-sana-1.0",
"dc-ae-f32c32-in-1.0",
"dc-ae-f32c32-mix-1.0",
"dc-ae-f64c128-in-1.0",
"dc-ae-f64c128-mix-1.0",
"dc-ae-f128c512-in-1.0",
"dc-ae-f128c512-mix-1.0",
],
help="The DCAE checkpoint to convert",
)
parser.add_argument("--output_path", type=str, required=True, help="Path where converted model should be saved")
parser.add_argument("--dtype", default="fp32", help="Torch dtype to save the model in.")
return parser.parse_args()
DTYPE_MAPPING = {
"fp32": torch.float32,
"fp16": torch.float16,
"bf16": torch.bfloat16,
}
VARIANT_MAPPING = {
"fp32": None,
"fp16": "fp16",
"bf16": "bf16",
}
if __name__ == "__main__":
args = get_args()
dtype = DTYPE_MAPPING[args.dtype]
variant = VARIANT_MAPPING[args.dtype]
ae = convert_ae(args.config_name, dtype)
ae.save_pretrained(args.output_path, safe_serialization=True, max_shard_size="5GB", variant=variant)
You can’t perform that action at this time.