-
Notifications
You must be signed in to change notification settings - Fork 2
/
merge_to_original.py
133 lines (110 loc) · 5.08 KB
/
merge_to_original.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
import argparse
import torch
argparser = argparse.ArgumentParser()
argparser.add_argument('--processor_path', type=str)
argparser.add_argument('--output_path', type=str)
argparser.add_argument('--model_path', type=str)
args = argparser.parse_args()
processor_path = args.processor_path
output_path = args.output_path
model_path = args.model_path
processor = torch.load(processor_path, map_location='cpu')
state_dict = torch.load(model_path, map_location='cpu')
# from https://gist.github.com/jachiam/8a5c0b607e38fcc585168b90c686eb05
layer_types_to_processor = {
'to_q': 'W_Q',
'to_k': 'W_K',
'to_v': 'W_V',
}
unet_conversion_map = [
# (stable-diffusion, HF Diffusers)
("time_embed.0.weight", "time_embedding.linear_1.weight"),
("time_embed.0.bias", "time_embedding.linear_1.bias"),
("time_embed.2.weight", "time_embedding.linear_2.weight"),
("time_embed.2.bias", "time_embedding.linear_2.bias"),
("input_blocks.0.0.weight", "conv_in.weight"),
("input_blocks.0.0.bias", "conv_in.bias"),
("out.0.weight", "conv_norm_out.weight"),
("out.0.bias", "conv_norm_out.bias"),
("out.2.weight", "conv_out.weight"),
("out.2.bias", "conv_out.bias"),
]
unet_conversion_map_resnet = [
# (stable-diffusion, HF Diffusers)
("in_layers.0", "norm1"),
("in_layers.2", "conv1"),
("out_layers.0", "norm2"),
("out_layers.3", "conv2"),
("emb_layers.1", "time_emb_proj"),
("skip_connection", "conv_shortcut"),
]
unet_conversion_map_layer = []
# hardcoded number of downblocks and resnets/attentions...
# would need smarter logic for other networks.
for i in range(4):
# loop over downblocks/upblocks
for j in range(2):
# loop over resnets/attentions for downblocks
hf_down_res_prefix = f"down_blocks.{i}.resnets.{j}."
sd_down_res_prefix = f"input_blocks.{3*i + j + 1}.0."
unet_conversion_map_layer.append((sd_down_res_prefix, hf_down_res_prefix))
if i < 3:
# no attention layers in down_blocks.3
hf_down_atn_prefix = f"down_blocks.{i}.attentions.{j}."
sd_down_atn_prefix = f"input_blocks.{3*i + j + 1}.1."
unet_conversion_map_layer.append((sd_down_atn_prefix, hf_down_atn_prefix))
for j in range(3):
# loop over resnets/attentions for upblocks
hf_up_res_prefix = f"up_blocks.{i}.resnets.{j}."
sd_up_res_prefix = f"output_blocks.{3*i + j}.0."
unet_conversion_map_layer.append((sd_up_res_prefix, hf_up_res_prefix))
if i > 0:
# no attention layers in up_blocks.0
hf_up_atn_prefix = f"up_blocks.{i}.attentions.{j}."
sd_up_atn_prefix = f"output_blocks.{3*i + j}.1."
unet_conversion_map_layer.append((sd_up_atn_prefix, hf_up_atn_prefix))
if i < 3:
# no downsample in down_blocks.3
hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0.conv."
sd_downsample_prefix = f"input_blocks.{3*(i+1)}.0.op."
unet_conversion_map_layer.append((sd_downsample_prefix, hf_downsample_prefix))
# no upsample in up_blocks.3
hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0."
sd_upsample_prefix = f"output_blocks.{3*i + 2}.{1 if i == 0 else 2}."
unet_conversion_map_layer.append((sd_upsample_prefix, hf_upsample_prefix))
hf_mid_atn_prefix = "mid_block.attentions.0."
sd_mid_atn_prefix = "middle_block.1."
unet_conversion_map_layer.append((sd_mid_atn_prefix, hf_mid_atn_prefix))
for j in range(2):
hf_mid_res_prefix = f"mid_block.resnets.{j}."
sd_mid_res_prefix = f"middle_block.{2*j}."
unet_conversion_map_layer.append((sd_mid_res_prefix, hf_mid_res_prefix))
def get_name_for_processor(name):
for sd_name, hf_name in unet_conversion_map:
name = name.replace(sd_name, hf_name)
for sd_name, hf_name in unet_conversion_map_layer:
name = name.replace(sd_name, hf_name)
return name
for k in state_dict.keys():
if ".attn2." in k and 'to_out' not in k and '.weight' in k and 'model.diffusion_model.' in k:
# print(k, v)
layer = '.'.join(k.split(".")[:-2]).replace('model.diffusion_model.', '')
processor_key = get_name_for_processor(layer) + '.processor'
weights = processor['weights'][processor_key]
r = processor["parameters"][processor_key]["r"]
hidden_size = processor["parameters"][processor_key]["hidden_size"]
block_size = hidden_size // r
layer_type = k.split(".")[-2]
processor_name = layer_types_to_processor[layer_type]
W = weights[processor_name]
block_Q = W - W.transpose(1, 2)
norm_Q = torch.norm(block_Q.flatten())
new_norm_Q = torch.clamp(norm_Q, max=processor["parameters"][processor_key]["constraint"])
block_Q = block_Q * ((new_norm_Q + 1e-8) / (norm_Q + 1e-8))
I = torch.eye(block_size, device=W.device).unsqueeze(0).repeat(r, 1, 1)
block_R = torch.matmul(I + block_Q, (I - block_Q).inverse())
R = torch.block_diag(*block_R).to(torch.float32)
state_dict[k] = torch.matmul(R.T, state_dict[k].to(torch.float32)).to(state_dict[k].dtype)
torch.save({
'state_dict': state_dict
}, output_path)