-
Notifications
You must be signed in to change notification settings - Fork 45
/
format_convert.py
177 lines (132 loc) 路 6.95 KB
/
format_convert.py
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
"""
Copyright 2023, Haofan Wang.
This script is for LoRA weights converting. It is NOT elegant, just working for temporary usage.
(1) Train in diffusers, the LoRA is saved in .bin format. You can convert it to .safetensors for stable-diffusion-webui. Note that diffusers only support adding LoRA to UNet.
(2) Download from other platforms such as civitai, the LoRA is saved in .safetensors. You can convert it to .bin then and load in unet directly using diffusers API.
All is about weight mapping. Below are weight namings of .bin and .safetensors for a specific layer.
# model layer
# Linear(in_features=320, out_features=320, bias=False)
pipeline.unet.down_blocks[0].attentions[0].transformer_blocks[0].attn1.to_q
- .bin
'down_blocks.0.attentions.0.transformer_blocks.0.attn1.processor.to_q_lora.up.weight' # torch.Size([320, rank])
'down_blocks.0.attentions.0.transformer_blocks.0.attn1.processor.to_q_lora.down.weight' # torch.Size([rank, 320])
- .safetensors
'lora_unet_down_blocks_0_attentions_0_transformer_blocks_0_attn1_to_q.lora_up.weight' # torch.Size([320, rank])
'lora_unet_down_blocks_0_attentions_0_transformer_blocks_0_attn1_to_q.lora_down.weight' # torch.Size([rank, 320])
"""
import torch
from safetensors.torch import load_file, save_file
from diffusers import StableDiffusionPipeline
from diffusers import DPMSolverMultistepScheduler
LORA_PREFIX_UNET = 'lora_unet'
def convert_name_to_bin(name):
# down_blocks_0_attentions_0_transformer_blocks_0_attn1_to_q.lora_up
new_name = name.replace(LORA_PREFIX_UNET+'_', '')
new_name = new_name.replace('.weight', '')
# ['down_blocks_0_attentions_0_transformer_blocks_0_attn1_to_q', 'lora.up']
parts = new_name.split('.')
#parts[0] = parts[0].replace('_0', '')
if 'out' in parts[0]:
parts[0] = "_".join(parts[0].split('_')[:-1])
parts[1] = parts[1].replace('_', '.')
# ['down', 'blocks', '0', 'attentions', '0', 'transformer', 'blocks', '0', 'attn1', 'to', 'q']
# ['mid', 'block', 'attentions', '0', 'transformer', 'blocks', '0', 'attn2', 'to', 'out']
sub_parts = parts[0].split('_')
# down_blocks.0.attentions.0.transformer_blocks.0.attn1.to_q_
new_sub_parts = ""
for i in range(len(sub_parts)):
if sub_parts[i] in ['block', 'blocks', 'attentions'] or sub_parts[i].isnumeric() or 'attn' in sub_parts[i]:
if 'attn' in sub_parts[i]:
new_sub_parts += sub_parts[i] + ".processor."
else:
new_sub_parts += sub_parts[i] + "."
else:
new_sub_parts += sub_parts[i] + "_"
# down_blocks.0.attentions.0.transformer_blocks.0.attn1.processor.to_q_lora.up
new_sub_parts += parts[1]
new_name = new_sub_parts + '.weight'
return new_name
def safetensors_to_bin(safetensor_path, bin_path):
bin_state_dict = {}
safetensors_state_dict = load_file(safetensor_path)
for key_safetensors in safetensors_state_dict:
# these if are required by current diffusers' API
# remove these may have negative effect as not all LoRAs are used
if 'text' in key_safetensors:
continue
if 'unet' not in key_safetensors:
continue
if 'transformer_blocks' not in key_safetensors:
continue
if 'ff_net' in key_safetensors or 'alpha' in key_safetensors:
continue
key_bin = convert_name_to_bin(key_safetensors)
bin_state_dict[key_bin] = safetensors_state_dict[key_safetensors]
torch.save(bin_state_dict, bin_path)
def convert_name_to_safetensors(name):
# ['down_blocks', '0', 'attentions', '0', 'transformer_blocks', '0', 'attn1', 'processor', 'to_q_lora', 'up', 'weight']
parts = name.split('.')
# ['down_blocks', '_0', 'attentions', '_0', 'transformer_blocks', '_0', 'attn1', 'processor', 'to_q_lora', 'up', 'weight']
for i in range(len(parts)):
if parts[i].isdigit():
parts[i] = '_' + parts[i]
if "to" in parts[i] and "lora" in parts[i]:
parts[i] = parts[i].replace('_lora', '.lora')
new_parts = []
for i in range(len(parts)):
if i == 0:
new_parts.append(LORA_PREFIX_UNET + '_' + parts[i])
elif i == len(parts) - 2:
new_parts.append(parts[i] + '_to_' + parts[i+1])
new_parts[-1] = new_parts[-1].replace('_to_weight', '')
elif i == len(parts) - 1:
new_parts[-1] += '.' + parts[i]
elif parts[i] != 'processor':
new_parts.append(parts[i])
new_name = '_'.join(new_parts)
new_name = new_name.replace('__', '_')
new_name = new_name.replace('_to_out.', '_to_out_0.')
return new_name
def bin_to_safetensors(bin_path, safetensor_path):
bin_state_dict = torch.load(bin_path)
safetensors_state_dict = {}
for key_bin in bin_state_dict:
key_safetensors = convert_name_to_safetensors(key_bin)
safetensors_state_dict[key_safetensors] = bin_state_dict[key_bin]
save_file(safetensors_state_dict, safetensor_path)
if __name__ == "__main__":
"""
sanity check
"""
#desire_output = 'lora_unet_down_blocks_0_attentions_0_transformer_blocks_0_attn1_to_q.lora_up.weight'
#output = convert_name_to_safetensors('down_blocks.0.attentions.0.transformer_blocks.0.attn1.processor.to_q_lora.up.weight')
#desire_output = 'down_blocks.0.attentions.0.transformer_blocks.0.attn1.processor.to_q_lora.up.weight'
#output = convert_name_to_bin('lora_unet_down_blocks_0_attentions_0_transformer_blocks_0_attn1_to_q.lora_up.weight')
# down_blocks.0.attentions.0.transformer_blocks.0.attn1.processor.to_out_lora.up.weight
# print(convert_name_to_bin('lora_unet_down_blocks_0_attentions_0_transformer_blocks_0_attn1_to_out_0.lora_up.weight'))
"""
from safetensor to bin
"""
# load diffusers model
model_id = "runwayml/stable-diffusion-v1-5"
pipeline = StableDiffusionPipeline.from_pretrained(model_id,torch_dtype=torch.float32)
# convert
# you have to download a suitable safetensors, not all is supported!
# download example from https://huggingface.co/SenY/LoRA/tree/main
# wget https://huggingface.co/SenY/LoRA/resolve/main/CheapCotton.safetensors
safetensor_path = "CheapCotton.safetensors"
bin_path = "CheapCotton.bin"
safetensors_to_bin(safetensor_path, bin_path)
# load it into UNet
# please note that diffusers' load_attn_procs only support add LoRA into attention
# if you have LoRA with other insertion, it does not support now
pipeline.unet.load_attn_procs(bin_path)
"""
from bin to safetensor
"""
# convert
# download example from https://huggingface.co/sayakpaul/sd-model-finetuned-lora-t4/tree/main
# wget https://huggingface.co/sayakpaul/sd-model-finetuned-lora-t4/resolve/main/pytorch_lora_weights.bin
bin_path = "pytorch_lora_weights.bin"
safetensor_path = "pytorch_lora_weights.safetensors"
bin_to_safetensors(bin_path, safetensor_path)