-
Notifications
You must be signed in to change notification settings - Fork 1
/
savemodel.py
150 lines (115 loc) · 4.44 KB
/
savemodel.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
import sys
import struct
import json
import torch
import numpy as np
#from transformers import AutoModel, AutoTokenizer
from sentence_transformers import SentenceTransformer
import re
if len(sys.argv) > 1:
dir_model = sys.argv[1]
else:
dir_model = "msmarco-distilbert-base-dot-prod-v3"
with open(dir_model + "/tokenizer.json", "r", encoding="utf-8") as f:
encoder = json.load(f)
with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
hparams = json.load(f)
with open(dir_model + "/modules.json", "r", encoding="utf-8") as f:
modules = json.load(f)
st_model = SentenceTransformer(dir_model)
list_vars = st_model[0].state_dict() # transformer
def strip(x: str):
x = "auto_model." + x
print(x)
y = list_vars[x]
assert y.view(-1)[0].dtype == torch.float32
return y.numpy()
if len(sys.argv) > 2:
outfile = sys.argv[2]
else:
outfile = "msmarco-distilbert-base-dot-prod-v3_converted_full.bin"
with open(outfile,mode='wb') as of:
#write up front stuff
header = struct.pack(
'iiiiiii',
hparams['dim'], hparams['hidden_dim'], hparams['n_layers'],
hparams['n_heads'], 0, len(encoder['model']['vocab']),
hparams['max_position_embeddings'],
)
of.write(header)
w = strip('embeddings.word_embeddings.weight')
of.write(memoryview(w))
w = strip('embeddings.position_embeddings.weight')
of.write(memoryview(w))
w = strip('embeddings.LayerNorm.weight')
of.write(memoryview(w))
w = strip('embeddings.LayerNorm.bias')
of.write(memoryview(w))
layers = hparams['n_layers']
for l in range(layers):
w = strip(f'transformer.layer.{l}.attention.q_lin.weight')
of.write(memoryview(w))
for l in range(layers):
w = strip(f'transformer.layer.{l}.attention.q_lin.bias')
of.write(memoryview(w))
for l in range(layers):
w = strip(f'transformer.layer.{l}.attention.k_lin.weight')
of.write(memoryview(w))
for l in range(layers):
w = strip(f'transformer.layer.{l}.attention.k_lin.bias')
of.write(memoryview(w))
for l in range(layers):
w = strip(f'transformer.layer.{l}.attention.v_lin.weight')
of.write(memoryview(w))
for l in range(layers):
w = strip(f'transformer.layer.{l}.attention.v_lin.bias')
of.write(memoryview(w))
for l in range(layers):
w = strip(f'transformer.layer.{l}.attention.out_lin.weight')
of.write(memoryview(w))
for l in range(layers):
w = strip(f'transformer.layer.{l}.attention.out_lin.bias')
of.write(memoryview(w))
for l in range(layers):
w = strip(f'transformer.layer.{l}.sa_layer_norm.weight')
of.write(memoryview(w))
for l in range(layers):
w = strip(f'transformer.layer.{l}.sa_layer_norm.bias')
of.write(memoryview(w))
for l in range(layers):
w = strip(f'transformer.layer.{l}.ffn.lin1.weight')
of.write(memoryview(w))
for l in range(layers):
w = strip(f'transformer.layer.{l}.ffn.lin1.bias')
of.write(memoryview(w))
for l in range(layers):
w = strip(f'transformer.layer.{l}.ffn.lin2.weight')
of.write(memoryview(w))
for l in range(layers):
w = strip(f'transformer.layer.{l}.ffn.lin2.bias')
of.write(memoryview(w))
for l in range(layers):
w = strip(f'transformer.layer.{l}.output_layer_norm.weight')
of.write(memoryview(w))
for l in range(layers):
w = strip(f'transformer.layer.{l}.output_layer_norm.bias')
of.write(memoryview(w))
# just stick the linear weights at the end
print("linear.weight")
y = st_model[2].state_dict()['linear.weight']
assert y.view(-1)[0].dtype == torch.float32
of.write(memoryview(y.numpy()))
if len(sys.argv) > 3:
vname = sys.argv[3]
else:
vname = "tokenizer.bin"
vocab = encoder["model"]["vocab"]
# write out vocab
max_len = max([len(bytes(v,"utf-8")) for v in vocab])
print("Maximum word size: ", max_len)
with open(vname, "wb") as f:
f.write(struct.pack("i", max_len))
for v in vocab:
vb = bytes(v,"utf-8")
f.write(struct.pack("ii", 0, len(vb)))
f.write(struct.pack(f"{len(vb)}s",vb))