-
Notifications
You must be signed in to change notification settings - Fork 0
/
LoadPretrained.py
192 lines (172 loc) · 8.63 KB
/
LoadPretrained.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
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
import torch
# res->vae
def load_res_vae(model, param):
model.s[0].load_state_dict(param.norm1.state_dict())
model.s[2].load_state_dict(param.conv1.state_dict())
model.s[3].load_state_dict(param.norm2.state_dict())
model.s[5].load_state_dict(param.conv2.state_dict())
if isinstance(model.res, torch.nn.Module):
model.res.load_state_dict(param.conv_shortcut.state_dict())
# attention->vae
def load_atten(model, param):
model.norm.load_state_dict(param.group_norm.state_dict())
model.q.load_state_dict(param.query.state_dict())
model.k.load_state_dict(param.key.state_dict())
model.v.load_state_dict(param.value.state_dict())
model.out.load_state_dict(param.proj_attn.state_dict())
# vae:encoder+decoder
def load_vae(model, params):
# encoder.in
model.encoder[0].load_state_dict(params.encoder.conv_in.state_dict())
# encoder.down
for i in range(4):
load_res_vae(model.encoder[i + 1][0], params.encoder.down_blocks[i].resnets[0])
load_res_vae(model.encoder[i + 1][1], params.encoder.down_blocks[i].resnets[1])
if i != 3:
model.encoder[i + 1][2][1].load_state_dict(
params.encoder.down_blocks[i].downsamplers[0].conv.state_dict())
# encoder.mid
load_res_vae(model.encoder[5][0], params.encoder.mid_block.resnets[0])
load_res_vae(model.encoder[5][2], params.encoder.mid_block.resnets[1])
load_atten(model.encoder[5][1], params.encoder.mid_block.attentions[0])
# encoder.out
model.encoder[6][0].load_state_dict(params.encoder.conv_norm_out.state_dict())
model.encoder[6][2].load_state_dict(params.encoder.conv_out.state_dict())
# encoder.正态分布层
model.encoder[7].load_state_dict(params.quant_conv.state_dict())
# decoder.正态分布层
model.decoder[0].load_state_dict(params.post_quant_conv.state_dict())
# decoder.in
model.decoder[1].load_state_dict(params.decoder.conv_in.state_dict())
# decoder.mid
load_res_vae(model.decoder[2][0], params.decoder.mid_block.resnets[0])
load_res_vae(model.decoder[2][2], params.decoder.mid_block.resnets[1])
load_atten(model.decoder[2][1], params.decoder.mid_block.attentions[0])
# decoder.up
for i in range(4):
load_res_vae(model.decoder[i + 3][0], params.decoder.up_blocks[i].resnets[0])
load_res_vae(model.decoder[i + 3][1], params.decoder.up_blocks[i].resnets[1])
load_res_vae(model.decoder[i + 3][2], params.decoder.up_blocks[i].resnets[2])
if i != 3:
model.decoder[i + 3][4].load_state_dict(
params.decoder.up_blocks[i].upsamplers[0].conv.state_dict())
# decoder.out
model.decoder[7][0].load_state_dict(params.decoder.conv_norm_out.state_dict())
model.decoder[7][2].load_state_dict(params.decoder.conv_out.state_dict())
# text encoder:词嵌入+12×clip encoder
def load_text(model, params):
model.embed.embed.load_state_dict(
params.text_model.embeddings.token_embedding.state_dict())
model.embed.pos_embed.load_state_dict(
params.text_model.embeddings.position_embedding.state_dict())
for i in range(12):
# 第一层norm
model.encoder[i].s1[0].load_state_dict(
params.text_model.encoder.layers[i].layer_norm1.state_dict())
# 注意力q矩阵
model.encoder[i].s1[1].q.load_state_dict(
params.text_model.encoder.layers[i].self_attn.q_proj.state_dict())
# 注意力k矩阵
model.encoder[i].s1[1].k.load_state_dict(
params.text_model.encoder.layers[i].self_attn.k_proj.state_dict())
# 注意力v矩阵
model.encoder[i].s1[1].v.load_state_dict(
params.text_model.encoder.layers[i].self_attn.v_proj.state_dict())
# 注意力out
model.encoder[i].s1[1].out.load_state_dict(
params.text_model.encoder.layers[i].self_attn.out_proj.state_dict())
# 第二层norm
model.encoder[i].s2[0].load_state_dict(
params.text_model.encoder.layers[i].layer_norm2.state_dict())
# mlp第一层fc
model.encoder[i].s2[1].load_state_dict(
params.text_model.encoder.layers[i].mlp.fc1.state_dict())
# mlp第二层fc
model.encoder[i].s3.load_state_dict(
params.text_model.encoder.layers[i].mlp.fc2.state_dict())
model.encoder[12].load_state_dict(params.text_model.final_layer_norm.state_dict())
# transformer->down block / up block->unet
def load_tf(model, param):
model.norm_in.load_state_dict(param.norm.state_dict())
model.cnn_in.load_state_dict(param.proj_in.state_dict())
model.atten1.q.load_state_dict(
param.transformer_blocks[0].attn1.to_q.state_dict())
model.atten1.k.load_state_dict(
param.transformer_blocks[0].attn1.to_k.state_dict())
model.atten1.v.load_state_dict(
param.transformer_blocks[0].attn1.to_v.state_dict())
model.atten1.out.load_state_dict(
param.transformer_blocks[0].attn1.to_out[0].state_dict())
model.atten2.q.load_state_dict(
param.transformer_blocks[0].attn2.to_q.state_dict())
model.atten2.k.load_state_dict(
param.transformer_blocks[0].attn2.to_k.state_dict())
model.atten2.v.load_state_dict(
param.transformer_blocks[0].attn2.to_v.state_dict())
model.atten2.out.load_state_dict(
param.transformer_blocks[0].attn2.to_out[0].state_dict())
model.fc0.load_state_dict(
param.transformer_blocks[0].ff.net[0].proj.state_dict())
model.fc1.load_state_dict(
param.transformer_blocks[0].ff.net[2].state_dict())
model.norm_atten0.load_state_dict(
param.transformer_blocks[0].norm1.state_dict())
model.norm_atten1.load_state_dict(
param.transformer_blocks[0].norm2.state_dict())
model.norm_act.load_state_dict(
param.transformer_blocks[0].norm3.state_dict())
model.cnn_out.load_state_dict(param.proj_out.state_dict())
# resnet->down block / up block->unet
def load_res_unet(model, param):
model.time[1].load_state_dict(param.time_emb_proj.state_dict())
model.s0[0].load_state_dict(param.norm1.state_dict())
model.s0[2].load_state_dict(param.conv1.state_dict())
model.s1[0].load_state_dict(param.norm2.state_dict())
model.s1[2].load_state_dict(param.conv2.state_dict())
if isinstance(model.res, torch.nn.Module):
model.res.load_state_dict(param.conv_shortcut.state_dict())
# down block->unet
def load_down_block(model, param):
load_tf(model.tf0, param.attentions[0])
load_tf(model.tf1, param.attentions[1])
load_res_unet(model.res0, param.resnets[0])
load_res_unet(model.res1, param.resnets[1])
model.out.load_state_dict(param.downsamplers[0].conv.state_dict())
# up block->unet
def load_up_block(model, param):
load_tf(model.tf0, param.attentions[0])
load_tf(model.tf1, param.attentions[1])
load_tf(model.tf2, param.attentions[2])
load_res_unet(model.res0, param.resnets[0])
load_res_unet(model.res1, param.resnets[1])
load_res_unet(model.res2, param.resnets[2])
if isinstance(model.out, torch.nn.Module):
model.out[1].load_state_dict(param.upsamplers[0].conv.state_dict())
# unet模型：1层in, 4层down, 1层middle, 4层up, 1层out
def load_unet(model, param):
# in
model.in_vae.load_state_dict(param.conv_in.state_dict())
model.in_time[0].load_state_dict(param.time_embedding.linear_1.state_dict())
model.in_time[2].load_state_dict(param.time_embedding.linear_2.state_dict())
# down
load_down_block(model.down_block0, param.down_blocks[0])
load_down_block(model.down_block1, param.down_blocks[1])
load_down_block(model.down_block2, param.down_blocks[2])
load_res_unet(model.down_res0, param.down_blocks[3].resnets[0])
load_res_unet(model.down_res1, param.down_blocks[3].resnets[1])
# mid
load_tf(model.mid_tf, param.mid_block.attentions[0])
load_res_unet(model.mid_res0, param.mid_block.resnets[0])
load_res_unet(model.mid_res1, param.mid_block.resnets[1])
# up
load_res_unet(model.up_res0, param.up_blocks[0].resnets[0])
load_res_unet(model.up_res1, param.up_blocks[0].resnets[1])
load_res_unet(model.up_res2, param.up_blocks[0].resnets[2])
model.up_in[1].load_state_dict(
param.up_blocks[0].upsamplers[0].conv.state_dict())
load_up_block(model.up_block0, param.up_blocks[1])
load_up_block(model.up_block1, param.up_blocks[2])
load_up_block(model.up_block2, param.up_blocks[3])
# out
model.out[0].load_state_dict(param.conv_norm_out.state_dict())
model.out[2].load_state_dict(param.conv_out.state_dict())