ApfelSchorle
commited on
Commit
•
9e92d30
1
Parent(s):
844e982
upload All
Browse files- .gitattributes +1 -0
- 3BSanokaKai2/AI-Large.py +708 -0
- 3BSanokaKai2/LLM1.pth +3 -0
- 3BSanokaKai2/LLM2.pth +3 -0
- 3BSanokaKai2/LLM3.pth +3 -0
- 3BSanokaKai2/LLM4.pth +3 -0
- 3BSanokaKai2/LLM5.pth +3 -0
- 3BSanokaKai2/LLM6.pth +3 -0
- 3BSanokaKai2/licence.txt +7 -0
- 3BSanokaKai2/output.pth +3 -0
- 3BSanokaKai2/readme.txt +38 -0
- 3BSanokaKai2/table.txt +3 -0
- 3BSanokaKai2/tokenizer.model +3 -0
- 3BSanokaKai2/tokenizer.vocab +0 -0
- 3BSanokaKai2/word2vec.model +3 -0
.gitattributes
CHANGED
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
36 |
+
3BSanokaKai2/table.txt filter=lfs diff=lfs merge=lfs -text
|
3BSanokaKai2/AI-Large.py
ADDED
@@ -0,0 +1,708 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# -*- coding: utf-8 -*-
|
2 |
+
"""
|
3 |
+
Created on Thu Mar 21 10:34:46 2024
|
4 |
+
|
5 |
+
@author: takan
|
6 |
+
"""
|
7 |
+
|
8 |
+
import MeCab
|
9 |
+
import torch
|
10 |
+
import copy
|
11 |
+
import time
|
12 |
+
import matplotlib.pyplot as plt
|
13 |
+
import re
|
14 |
+
import math
|
15 |
+
import numpy as np
|
16 |
+
from gensim.models import Word2Vec
|
17 |
+
import pickle
|
18 |
+
import threading
|
19 |
+
import sentencepiece as spm
|
20 |
+
|
21 |
+
class DenseBlock(torch.nn.Module):
|
22 |
+
def __init__(self, dim, mul=1):
|
23 |
+
super().__init__()
|
24 |
+
self.I = torch.nn.Linear(dim, dim*mul)
|
25 |
+
self.O = torch.nn.Linear(dim*mul, dim)
|
26 |
+
def forward(self, x):
|
27 |
+
x = self.I(x)
|
28 |
+
x = torch.nn.functional.elu(x)
|
29 |
+
x = self.O(x)
|
30 |
+
return x
|
31 |
+
|
32 |
+
class AttentionBlock(torch.nn.Module):
|
33 |
+
def __init__(self, dim, mul=1):
|
34 |
+
super().__init__()
|
35 |
+
self.Q = torch.nn.Linear(dim, dim*mul)
|
36 |
+
self.K = torch.nn.Linear(dim, dim*mul)
|
37 |
+
self.V = torch.nn.Linear(dim, dim*mul)
|
38 |
+
self.O = torch.nn.Linear(dim*mul, dim)
|
39 |
+
def forward(self, q,k,v):
|
40 |
+
q = self.Q(q)
|
41 |
+
k = self.K(k)
|
42 |
+
v = self.V(v)
|
43 |
+
x = torch.nn.functional.softmax(q * k, dim=-1) * v
|
44 |
+
x = self.O(x)
|
45 |
+
return x
|
46 |
+
"""
|
47 |
+
class AttentionBlock(torch.nn.Module):
|
48 |
+
def __init__(self, dim, mul=1):
|
49 |
+
super().__init__()
|
50 |
+
self.attn = torch.nn.MultiheadAttention(dim, 16, batch_first=True)
|
51 |
+
def forward(self, q,k,v):
|
52 |
+
x = self.attn(q, k, v)[0]
|
53 |
+
return x
|
54 |
+
"""
|
55 |
+
class SanokaLayer(torch.nn.Module):
|
56 |
+
def __init__(self, dim, mul=1):
|
57 |
+
super().__init__()
|
58 |
+
self.x = None
|
59 |
+
self.A = AttentionBlock(dim, mul)
|
60 |
+
self.B = DenseBlock(dim, mul)
|
61 |
+
def reset(self, x=None):
|
62 |
+
self.x = x
|
63 |
+
def forward(self, u):
|
64 |
+
if (self.x != None):
|
65 |
+
uu = torch.nn.functional.normalize(u)
|
66 |
+
xx = torch.nn.functional.normalize(self.x)
|
67 |
+
x = self.A(uu, xx, xx)
|
68 |
+
y = self.B(torch.nn.functional.normalize(x)) + u
|
69 |
+
self.x = x + self.x
|
70 |
+
return y
|
71 |
+
else:
|
72 |
+
uu = torch.nn.functional.normalize(u)
|
73 |
+
x = self.A(uu, uu, uu)
|
74 |
+
y = self.B(torch.nn.functional.normalize(x)) + u
|
75 |
+
self.x = x
|
76 |
+
return y
|
77 |
+
|
78 |
+
class SanokaModel(torch.nn.Module):
|
79 |
+
def __init__(self, dim, mul=1, Top=True):
|
80 |
+
super().__init__()
|
81 |
+
self.Top = Top
|
82 |
+
if (Top):
|
83 |
+
self.I = torch.nn.Linear(128, dim)
|
84 |
+
self.A = SanokaLayer(dim, mul)
|
85 |
+
self.B = SanokaLayer(dim, mul)
|
86 |
+
self.C = SanokaLayer(dim, mul)
|
87 |
+
self.D = SanokaLayer(dim, mul)
|
88 |
+
self.E = SanokaLayer(dim, mul)
|
89 |
+
self.F = SanokaLayer(dim, mul)
|
90 |
+
def reset(self):
|
91 |
+
self.A.reset()
|
92 |
+
self.B.reset()
|
93 |
+
self.C.reset()
|
94 |
+
self.D.reset()
|
95 |
+
self.E.reset()
|
96 |
+
self.F.reset()
|
97 |
+
|
98 |
+
def forward(self, x):
|
99 |
+
if (self.Top):
|
100 |
+
x = self.I(x)
|
101 |
+
x = self.A(x)
|
102 |
+
x = self.B(x)
|
103 |
+
x = self.C(x)
|
104 |
+
x = self.D(x)
|
105 |
+
x = self.E(x)
|
106 |
+
x = self.F(x)
|
107 |
+
|
108 |
+
return x
|
109 |
+
|
110 |
+
class OutputLayer (torch.nn.Module):
|
111 |
+
def __init__(self, hiddendim, worddim=59000, heads=4):
|
112 |
+
super().__init__()
|
113 |
+
self.H = torch.nn.Linear(hiddendim, worddim)
|
114 |
+
def forward(self, inpute):
|
115 |
+
x = inpute
|
116 |
+
x = self.H(x)
|
117 |
+
return x
|
118 |
+
|
119 |
+
def GOILOAD():
|
120 |
+
fuf = open("table.txt", "r", encoding="UTF-8")
|
121 |
+
goi = fuf.read().split("\n")
|
122 |
+
fuf.close()
|
123 |
+
chardim = len(goi[1:])
|
124 |
+
charid = {goi[i+1].split()[0]:i for i in range(chardim-1)}
|
125 |
+
return charid, [goi[ia+1].split()[0] for ia in range(chardim-1)]
|
126 |
+
|
127 |
+
datas = []
|
128 |
+
trues = []
|
129 |
+
lens = []
|
130 |
+
dones = 0
|
131 |
+
def Convert(buns, table, maxlen=256):
|
132 |
+
buns = buns.split("\n")
|
133 |
+
sp = spm.SentencePieceProcessor()
|
134 |
+
sp.Load("tokenizer.model")
|
135 |
+
w2v = Word2Vec.load("word2vec.model")
|
136 |
+
data = []
|
137 |
+
true = []
|
138 |
+
lena = []
|
139 |
+
for datac in range(len(buns)):
|
140 |
+
#print(datac)
|
141 |
+
#print(buns[datac])
|
142 |
+
error = False
|
143 |
+
try:
|
144 |
+
buna = sp.EncodeAsPieces(buns[datac])[:maxlen]
|
145 |
+
a = torch.from_numpy(w2v.wv[buna])
|
146 |
+
b = torch.tensor([table[buna[ii]] for ii in range(len(buna))])
|
147 |
+
ll = len(buna)
|
148 |
+
c = ll
|
149 |
+
except:
|
150 |
+
print("ERROR")
|
151 |
+
else:
|
152 |
+
data.append(a)
|
153 |
+
true.append(b)
|
154 |
+
lena.append(c)
|
155 |
+
print(datac)
|
156 |
+
f = open("Train_Data.bin", "wb")
|
157 |
+
pickle.dump((data, true, lena), f)
|
158 |
+
f.close()
|
159 |
+
return
|
160 |
+
|
161 |
+
def SPMake():
|
162 |
+
|
163 |
+
spm.SentencePieceTrainer.Train(f"--input=train_data.txt --model_prefix=tokenizer --vocab_size=20000 --train_extremely_large_corpus=True")
|
164 |
+
def W2VMake(filepath="train_data.txt", mincount=50, worker=60):
|
165 |
+
sp = spm.SentencePieceProcessor()
|
166 |
+
sp.Load("tokenizer.model")
|
167 |
+
f = open(filepath, mode="r", encoding="UTF-8")
|
168 |
+
texts = f.read().split("\n")
|
169 |
+
f.close()
|
170 |
+
dat = []
|
171 |
+
print(len(texts))
|
172 |
+
for a in range(len(texts)):
|
173 |
+
dat.append(sp.EncodeAsPieces(texts[a]))
|
174 |
+
print(a)
|
175 |
+
|
176 |
+
model = Word2Vec(sentences=dat, vector_size=128, window=100, min_count=mincount, workers=worker)
|
177 |
+
model.save("word2vec.model")
|
178 |
+
model.wv.save_word2vec_format('table.txt')
|
179 |
+
|
180 |
+
def DataMake(filepath="train_data.txt", maxlen=129):
|
181 |
+
table, i2w = GOILOAD()
|
182 |
+
print(len(table))
|
183 |
+
time.sleep(1)
|
184 |
+
f = open(filepath, mode="r", encoding="UTF-8")
|
185 |
+
txt = f.read()
|
186 |
+
f.close()
|
187 |
+
Convert(txt, table)
|
188 |
+
return None
|
189 |
+
|
190 |
+
def PreTrain(Load=False, dim=512, outputdim=40000, lr=1e-04, epoch=10, epochload=1000,usedata=480000, onestep=100, uselen=64):
|
191 |
+
global datas
|
192 |
+
global trues
|
193 |
+
global lens
|
194 |
+
torch.manual_seed(1293431)
|
195 |
+
#torch.manual_seed(576765)
|
196 |
+
device1 = torch.device("cuda:0")
|
197 |
+
device2 = torch.device("cuda:1")
|
198 |
+
device3 = torch.device("cuda:2")
|
199 |
+
device4 = torch.device("cuda:3")
|
200 |
+
device5 = torch.device("cuda:4")
|
201 |
+
device6 = torch.device("cuda:5")
|
202 |
+
device7 = torch.device("cuda:6")
|
203 |
+
lossf = torch.nn.CrossEntropyLoss()
|
204 |
+
model1 = SanokaModel(dim, 2, True).to(torch.bfloat16).to(device1)
|
205 |
+
model2 = SanokaModel(dim, 2, False).to(torch.bfloat16).to(device2)
|
206 |
+
model3 = SanokaModel(dim, 2, False).to(torch.bfloat16).to(device3)
|
207 |
+
model4 = SanokaModel(dim, 2, False).to(torch.bfloat16).to(device4)
|
208 |
+
model5 = SanokaModel(dim, 2, False).to(torch.bfloat16).to(device5)
|
209 |
+
model6 = SanokaModel(dim, 2, False).to(torch.bfloat16).to(device6)
|
210 |
+
output = OutputLayer(dim, outputdim).to(torch.bfloat16).to(device7)
|
211 |
+
|
212 |
+
if (Load):
|
213 |
+
model1.load_state_dict(torch.load("LLM1.pth", map_location=device1))
|
214 |
+
model2.load_state_dict(torch.load("LLM2.pth", map_location=device2))
|
215 |
+
model3.load_state_dict(torch.load("LLM3.pth", map_location=device3))
|
216 |
+
model4.load_state_dict(torch.load("LLM4.pth", map_location=device4))
|
217 |
+
model5.load_state_dict(torch.load("LLM5.pth", map_location=device5))
|
218 |
+
model6.load_state_dict(torch.load("LLM6.pth", map_location=device6))
|
219 |
+
output.load_state_dict(torch.load("output.pth", map_location=device7))
|
220 |
+
model1Optim = torch.optim.Adam(model1.parameters(), lr=lr)
|
221 |
+
model2Optim = torch.optim.Adam(model2.parameters(), lr=lr)
|
222 |
+
model3Optim = torch.optim.Adam(model3.parameters(), lr=lr)
|
223 |
+
model4Optim = torch.optim.Adam(model4.parameters(), lr=lr)
|
224 |
+
model5Optim = torch.optim.Adam(model5.parameters(), lr=lr)
|
225 |
+
model6Optim = torch.optim.Adam(model6.parameters(), lr=lr)
|
226 |
+
outputO = torch.optim.Adam(output.parameters(), lr=lr)
|
227 |
+
f = open("Train_Data.bin", "rb")
|
228 |
+
datas, trues, lens = pickle.load(f)
|
229 |
+
f.close()
|
230 |
+
train_x = torch.zeros((epochload, uselen, 128)).to(torch.bfloat16).to(device1)
|
231 |
+
train_y = torch.full((epochload, uselen), outputdim - 1, dtype=torch.long).to(device7)
|
232 |
+
table, i2w = GOILOAD()
|
233 |
+
base = 0
|
234 |
+
epoch = int(np.floor((len(datas) / epochload) * epoch))
|
235 |
+
print("データ量", len(datas))
|
236 |
+
for epochs in range(epoch):
|
237 |
+
train_x = train_x.detach()
|
238 |
+
train_y = train_y.detach()
|
239 |
+
if (base < len(datas) - epochload*2):
|
240 |
+
base += epochload
|
241 |
+
else:
|
242 |
+
base = 0
|
243 |
+
if (base > usedata):
|
244 |
+
base = 0
|
245 |
+
for b in range(epochload):
|
246 |
+
a = b + base
|
247 |
+
leng = lens[a]
|
248 |
+
if (leng > uselen):
|
249 |
+
leng = uselen
|
250 |
+
|
251 |
+
train_x[b, :datas[a].shape[0]] = datas[a].to(torch.bfloat16).to(device1)[:uselen]
|
252 |
+
train_y[b, :trues[a].shape[0]] = trues[a].to(device7).to(torch.long)[:uselen]
|
253 |
+
epls = 0.00
|
254 |
+
timem = time.time()
|
255 |
+
for steps in range(epochload//onestep):
|
256 |
+
model1.reset()
|
257 |
+
model2.reset()
|
258 |
+
model3.reset()
|
259 |
+
model4.reset()
|
260 |
+
model5.reset()
|
261 |
+
model6.reset()
|
262 |
+
oa = ""
|
263 |
+
model1Optim.zero_grad()
|
264 |
+
model2Optim.zero_grad()
|
265 |
+
model3Optim.zero_grad()
|
266 |
+
model4Optim.zero_grad()
|
267 |
+
model5Optim.zero_grad()
|
268 |
+
model6Optim.zero_grad()
|
269 |
+
outputO.zero_grad()
|
270 |
+
loss = 0.00
|
271 |
+
for b in range(uselen-1):
|
272 |
+
out = model1(train_x[steps*onestep:steps*onestep+onestep, b])
|
273 |
+
out = model2(out.to(device2))
|
274 |
+
out = model3(out.to(device3))
|
275 |
+
out = model4(out.to(device4))
|
276 |
+
out = model5(out.to(device5))
|
277 |
+
out = model6(out.to(device6))
|
278 |
+
out = output(out.to(device7))
|
279 |
+
loss += lossf(out, train_y[steps*onestep:steps*onestep+onestep, b+1])
|
280 |
+
epls += loss
|
281 |
+
|
282 |
+
sfo = torch.nn.functional.softmax(out[0], dim=-1)
|
283 |
+
wid = torch.argmax(sfo, dim=-1).item()
|
284 |
+
try:
|
285 |
+
wd = i2w[wid]
|
286 |
+
except:
|
287 |
+
oa = oa + "ERROR"
|
288 |
+
else:
|
289 |
+
oa = oa + wd
|
290 |
+
|
291 |
+
loss.backward()
|
292 |
+
#print(b)
|
293 |
+
model1Optim.step()
|
294 |
+
model2Optim.step()
|
295 |
+
model3Optim.step()
|
296 |
+
model4Optim.step()
|
297 |
+
model5Optim.step()
|
298 |
+
model6Optim.step()
|
299 |
+
outputO.step()
|
300 |
+
print("出力サンプル> ", oa[:32].replace("?", ""))
|
301 |
+
print("epoch", epochs,"Train_epoch_sum_loss", epls.item(), "time", time.time() - timem)
|
302 |
+
if (epochs % 10 == 9):
|
303 |
+
torch.save(model1.state_dict(), "LLM1.pth")
|
304 |
+
torch.save(model2.state_dict(), "LLM2.pth")
|
305 |
+
torch.save(model3.state_dict(), "LLM3.pth")
|
306 |
+
torch.save(model4.state_dict(), "LLM4.pth")
|
307 |
+
torch.save(model5.state_dict(), "LLM5.pth")
|
308 |
+
torch.save(model6.state_dict(), "LLM6.pth")
|
309 |
+
torch.save(output.state_dict(), "output.pth")
|
310 |
+
def Fineturning(Load=False, dim=512, outputdim=40000, lr=1e-04, epoch=10000, epochload=1000, onestep=200, uselen=32):
|
311 |
+
global datas
|
312 |
+
global trues
|
313 |
+
global lens
|
314 |
+
torch.manual_seed(1293431)
|
315 |
+
#torch.manual_seed(576765)
|
316 |
+
device1 = torch.device("cuda:0")
|
317 |
+
device2 = torch.device("cuda:1")
|
318 |
+
device3 = torch.device("cuda:2")
|
319 |
+
device4 = torch.device("cuda:3")
|
320 |
+
device5 = torch.device("cuda:4")
|
321 |
+
device6 = torch.device("cuda:5")
|
322 |
+
device7 = torch.device("cuda:6")
|
323 |
+
lossf = torch.nn.CrossEntropyLoss()
|
324 |
+
model1 = SanokaModel(dim, 2, True).to(torch.bfloat16).to(device1)
|
325 |
+
model2 = SanokaModel(dim, 2, False).to(torch.bfloat16).to(device2)
|
326 |
+
model3 = SanokaModel(dim, 2, False).to(torch.bfloat16).to(device3)
|
327 |
+
model4 = SanokaModel(dim, 2, False).to(torch.bfloat16).to(device4)
|
328 |
+
model5 = SanokaModel(dim, 2, False).to(torch.bfloat16).to(device5)
|
329 |
+
model6 = SanokaModel(dim, 2, False).to(torch.bfloat16).to(device6)
|
330 |
+
output = OutputLayer(dim, outputdim).to(torch.bfloat16).to(device7)
|
331 |
+
|
332 |
+
model1.load_state_dict(torch.load("LLM1.pth", map_location=device1))
|
333 |
+
model2.load_state_dict(torch.load("LLM2.pth", map_location=device2))
|
334 |
+
model3.load_state_dict(torch.load("LLM3.pth", map_location=device3))
|
335 |
+
model4.load_state_dict(torch.load("LLM4.pth", map_location=device4))
|
336 |
+
model5.load_state_dict(torch.load("LLM5.pth", map_location=device5))
|
337 |
+
model6.load_state_dict(torch.load("LLM6.pth", map_location=device6))
|
338 |
+
output.load_state_dict(torch.load("output.pth", map_location=device7))
|
339 |
+
model1Optim = torch.optim.Adam(model1.parameters(), lr=lr)
|
340 |
+
model2Optim = torch.optim.Adam(model2.parameters(), lr=lr)
|
341 |
+
model3Optim = torch.optim.Adam(model3.parameters(), lr=lr)
|
342 |
+
model4Optim = torch.optim.Adam(model4.parameters(), lr=lr)
|
343 |
+
model5Optim = torch.optim.Adam(model5.parameters(), lr=lr)
|
344 |
+
model6Optim = torch.optim.Adam(model6.parameters(), lr=lr/500)
|
345 |
+
outputO = torch.optim.Adam(output.parameters(), lr=lr)
|
346 |
+
f = open("Train_Data.bin", "rb")
|
347 |
+
datas, trues, lens = pickle.load(f)
|
348 |
+
f.close()
|
349 |
+
train_x = torch.zeros((epochload, uselen, 128)).to(torch.bfloat16).to(device1)
|
350 |
+
train_y = torch.full((epochload, uselen), outputdim - 1, dtype=torch.long).to(device7)
|
351 |
+
table, i2w = GOILOAD()
|
352 |
+
base = 0
|
353 |
+
epoch = int(np.floor((len(datas) / epochload) * epoch))
|
354 |
+
#print(epoch)
|
355 |
+
for epochs in range(epoch):
|
356 |
+
train_x = train_x.detach()
|
357 |
+
train_y = train_y.detach()
|
358 |
+
if (base < len(datas) - epochload*2):
|
359 |
+
base += epochload
|
360 |
+
else:
|
361 |
+
base = 0
|
362 |
+
for b in range(epochload):
|
363 |
+
a = b + base
|
364 |
+
#print(a)
|
365 |
+
leng = lens[a]
|
366 |
+
if (leng > uselen):
|
367 |
+
leng = uselen
|
368 |
+
|
369 |
+
train_x[b, :datas[a].shape[0]] = datas[a].to(torch.bfloat16).to(device1)[:uselen]
|
370 |
+
train_y[b, :trues[a].shape[0]] = trues[a].to(device7).to(torch.long)[:uselen]
|
371 |
+
epls = 0.00
|
372 |
+
timem = time.time()
|
373 |
+
for steps in range(epochload//onestep):
|
374 |
+
model1.reset()
|
375 |
+
model2.reset()
|
376 |
+
model3.reset()
|
377 |
+
model4.reset()
|
378 |
+
model5.reset()
|
379 |
+
model6.reset()
|
380 |
+
oa = ""
|
381 |
+
loss = 0.00
|
382 |
+
model1Optim.zero_grad()
|
383 |
+
model2Optim.zero_grad()
|
384 |
+
model3Optim.zero_grad()
|
385 |
+
model4Optim.zero_grad()
|
386 |
+
model5Optim.zero_grad()
|
387 |
+
model6Optim.zero_grad()
|
388 |
+
outputO.zero_grad()
|
389 |
+
for b in range(uselen-1):
|
390 |
+
with torch.no_grad():
|
391 |
+
out = model1(train_x[steps*onestep:steps*onestep+onestep, b])
|
392 |
+
out = model2(out.to(device2))
|
393 |
+
out = model3(out.to(device3))
|
394 |
+
out = model4(out.to(device4))
|
395 |
+
out = model5(out.to(device5))
|
396 |
+
out = model6(out.to(device6))
|
397 |
+
out = output(out.to(device7))
|
398 |
+
loss += lossf(out, train_y[steps*onestep:steps*onestep+onestep, b+1])
|
399 |
+
epls += loss.item()
|
400 |
+
|
401 |
+
sfo = torch.nn.functional.softmax(out[0], dim=-1)
|
402 |
+
wid = torch.argmax(sfo, dim=-1).item()
|
403 |
+
try:
|
404 |
+
wd = i2w[wid]
|
405 |
+
except:
|
406 |
+
oa = oa + "ERROR"
|
407 |
+
else:
|
408 |
+
oa = oa + wd
|
409 |
+
loss.backward()
|
410 |
+
#model6Optim.step()
|
411 |
+
outputO.step()
|
412 |
+
print("出力サンプル> ", oa[:32].replace("?", ""))
|
413 |
+
print("epoch", epochs,"Train_epoch_sum_loss", epls, "time", time.time() - timem)
|
414 |
+
if (epochs % 10 == 9):
|
415 |
+
#torch.save(model6.state_dict(), "LLM6F.pth")
|
416 |
+
torch.save(output.state_dict(), "fineturning.pth")
|
417 |
+
def Predict(dim=512, outputdim=40000, maxlen=32):
|
418 |
+
|
419 |
+
torch.manual_seed(1293431)
|
420 |
+
|
421 |
+
table, i2w = GOILOAD()
|
422 |
+
sp = spm.SentencePieceProcessor()
|
423 |
+
sp.Load("tokenizer.model")
|
424 |
+
|
425 |
+
w2v = Word2Vec.load("word2vec.model")
|
426 |
+
|
427 |
+
device1 = torch.device("cuda:0")
|
428 |
+
|
429 |
+
device2 = torch.device("cuda:1")
|
430 |
+
|
431 |
+
device3 = torch.device("cuda:2")
|
432 |
+
|
433 |
+
device4 = torch.device("cuda:3")
|
434 |
+
|
435 |
+
device5 = torch.device("cuda:4")
|
436 |
+
|
437 |
+
device6 = torch.device("cuda:5")
|
438 |
+
|
439 |
+
device7 = torch.device("cuda:6")
|
440 |
+
|
441 |
+
lossf = torch.nn.CrossEntropyLoss()
|
442 |
+
|
443 |
+
model1 = SanokaModel(dim, 2, True).to(torch.bfloat16).to(device1)
|
444 |
+
|
445 |
+
model2 = SanokaModel(dim, 2, False).to(torch.bfloat16).to(device2)
|
446 |
+
|
447 |
+
model3 = SanokaModel(dim, 2, False).to(torch.bfloat16).to(device3)
|
448 |
+
|
449 |
+
model4 = SanokaModel(dim, 2, False).to(torch.bfloat16).to(device4)
|
450 |
+
|
451 |
+
model5 = SanokaModel(dim, 2, False).to(torch.bfloat16).to(device5)
|
452 |
+
|
453 |
+
model6 = SanokaModel(dim, 2, False).to(torch.bfloat16).to(device6)
|
454 |
+
|
455 |
+
output = OutputLayer(dim, outputdim).to(torch.bfloat16).to(device7)
|
456 |
+
|
457 |
+
|
458 |
+
|
459 |
+
model1.load_state_dict(torch.load("LLM1.pth", map_location=device1))
|
460 |
+
|
461 |
+
model2.load_state_dict(torch.load("LLM2.pth", map_location=device2))
|
462 |
+
|
463 |
+
model3.load_state_dict(torch.load("LLM3.pth", map_location=device3))
|
464 |
+
|
465 |
+
model4.load_state_dict(torch.load("LLM4.pth", map_location=device4))
|
466 |
+
|
467 |
+
model5.load_state_dict(torch.load("LLM5.pth", map_location=device5))
|
468 |
+
|
469 |
+
model6.load_state_dict(torch.load("LLM6.pth", map_location=device6))
|
470 |
+
|
471 |
+
output.load_state_dict(torch.load("fineturning.pth", map_location=device7))
|
472 |
+
|
473 |
+
while(1):
|
474 |
+
|
475 |
+
dd = input("Q> ")# + ","
|
476 |
+
|
477 |
+
|
478 |
+
|
479 |
+
data = []
|
480 |
+
|
481 |
+
buna = sp.EncodeAsPieces(dd)
|
482 |
+
|
483 |
+
print(buna)
|
484 |
+
|
485 |
+
for a in range(len(buna)):
|
486 |
+
|
487 |
+
try:
|
488 |
+
|
489 |
+
data.append(torch.from_numpy(w2v.wv[buna[a]]).view(1, 1, 128).to(device1))
|
490 |
+
|
491 |
+
except KeyError:
|
492 |
+
|
493 |
+
print("Not Found")
|
494 |
+
|
495 |
+
dat = torch.cat(data, dim=1).to(device1)
|
496 |
+
|
497 |
+
oa = ""
|
498 |
+
|
499 |
+
with torch.no_grad():
|
500 |
+
|
501 |
+
model1.reset()
|
502 |
+
|
503 |
+
model2.reset()
|
504 |
+
|
505 |
+
model3.reset()
|
506 |
+
|
507 |
+
model4.reset()
|
508 |
+
|
509 |
+
model5.reset()
|
510 |
+
|
511 |
+
model6.reset()
|
512 |
+
|
513 |
+
oa = ""
|
514 |
+
|
515 |
+
for a in range(dat.shape[1] - 1):
|
516 |
+
|
517 |
+
out = model1(dat[:, a].to(torch.bfloat16))
|
518 |
+
|
519 |
+
out = model2(out.to(device2))
|
520 |
+
|
521 |
+
out = model3(out.to(device3))
|
522 |
+
|
523 |
+
out = model4(out.to(device4))
|
524 |
+
|
525 |
+
out = model5(out.to(device5))
|
526 |
+
|
527 |
+
out = model6(out.to(device6))
|
528 |
+
|
529 |
+
out = output(out.to(device7))
|
530 |
+
|
531 |
+
for b in range(maxlen - dat.shape[1]):
|
532 |
+
|
533 |
+
out = model1(dat[:, -1].to(torch.bfloat16))
|
534 |
+
|
535 |
+
out = model2(out.to(device2))
|
536 |
+
|
537 |
+
out = model3(out.to(device3))
|
538 |
+
|
539 |
+
out = model4(out.to(device4))
|
540 |
+
|
541 |
+
out = model5(out.to(device5))
|
542 |
+
|
543 |
+
out = model6(out.to(device6))
|
544 |
+
|
545 |
+
out = output(out.to(device7))
|
546 |
+
|
547 |
+
sfo = torch.nn.functional.softmax(out, dim=-1)
|
548 |
+
|
549 |
+
wid = torch.argmax(sfo, dim=-1).item()
|
550 |
+
|
551 |
+
if (wid != outputdim - 1):
|
552 |
+
|
553 |
+
try:
|
554 |
+
|
555 |
+
wd = i2w[wid]
|
556 |
+
|
557 |
+
except:
|
558 |
+
|
559 |
+
oa = oa + "ERROR"
|
560 |
+
|
561 |
+
else:
|
562 |
+
|
563 |
+
oa = oa + wd
|
564 |
+
|
565 |
+
dat = torch.cat([dat, torch.from_numpy(w2v.wv[wd]).to(device1).view(1, 1, 128)], dim=1)
|
566 |
+
|
567 |
+
print("A> ", oa.replace("?", ""))
|
568 |
+
|
569 |
+
def ValidationLoss(dim=512, outputdim=40000, maxlen=32):
|
570 |
+
|
571 |
+
torch.manual_seed(1293431)
|
572 |
+
|
573 |
+
table, i2w = GOILOAD()
|
574 |
+
|
575 |
+
tagger = MeCab.Tagger("-Owakati")
|
576 |
+
|
577 |
+
w2v = Word2Vec.load("word2vec.model")
|
578 |
+
|
579 |
+
device1 = torch.device("cuda:0")
|
580 |
+
|
581 |
+
device2 = torch.device("cuda:1")
|
582 |
+
|
583 |
+
device3 = torch.device("cuda:2")
|
584 |
+
|
585 |
+
device4 = torch.device("cuda:3")
|
586 |
+
|
587 |
+
device5 = torch.device("cuda:4")
|
588 |
+
|
589 |
+
device6 = torch.device("cuda:5")
|
590 |
+
|
591 |
+
device7 = torch.device("cuda:6")
|
592 |
+
|
593 |
+
lossf = torch.nn.CrossEntropyLoss()
|
594 |
+
|
595 |
+
model1 = SanokaModel(dim, 2, True).to(torch.bfloat16).to(device1)
|
596 |
+
|
597 |
+
model2 = SanokaModel(dim, 2, False).to(torch.bfloat16).to(device2)
|
598 |
+
|
599 |
+
model3 = SanokaModel(dim, 2, False).to(torch.bfloat16).to(device3)
|
600 |
+
|
601 |
+
model4 = SanokaModel(dim, 2, False).to(torch.bfloat16).to(device4)
|
602 |
+
|
603 |
+
model5 = SanokaModel(dim, 2, False).to(torch.bfloat16).to(device5)
|
604 |
+
|
605 |
+
model6 = SanokaModel(dim, 2, False).to(torch.bfloat16).to(device6)
|
606 |
+
|
607 |
+
output = OutputLayer(dim, outputdim).to(torch.bfloat16).to(device7)
|
608 |
+
|
609 |
+
|
610 |
+
|
611 |
+
model1.load_state_dict(torch.load("LLM1.pth", map_location=device1))
|
612 |
+
|
613 |
+
model2.load_state_dict(torch.load("LLM2.pth", map_location=device2))
|
614 |
+
|
615 |
+
model3.load_state_dict(torch.load("LLM3.pth", map_location=device3))
|
616 |
+
|
617 |
+
model4.load_state_dict(torch.load("LLM4.pth", map_location=device4))
|
618 |
+
|
619 |
+
model5.load_state_dict(torch.load("LLM5.pth", map_location=device5))
|
620 |
+
|
621 |
+
model6.load_state_dict(torch.load("LLM6.pth", map_location=device6))
|
622 |
+
|
623 |
+
output.load_state_dict(torch.load("output.pth", map_location=device7))
|
624 |
+
|
625 |
+
dd = input("TestData> ")
|
626 |
+
|
627 |
+
lossf = torch.nn.CrossEntropyLoss()
|
628 |
+
|
629 |
+
data = []
|
630 |
+
|
631 |
+
buna = tagger.parse(dd).split()
|
632 |
+
|
633 |
+
trued = torch.tensor([table[dfg] for dfg in buna]).to(torch.long).unsqueeze(dim=0)
|
634 |
+
|
635 |
+
print(buna)
|
636 |
+
|
637 |
+
print(trued)
|
638 |
+
|
639 |
+
for a in range(len(buna)):
|
640 |
+
|
641 |
+
try:
|
642 |
+
|
643 |
+
data.append(torch.from_numpy(w2v.wv[buna[a]]).view(1, 1, 128).to(device1))
|
644 |
+
|
645 |
+
except KeyError:
|
646 |
+
|
647 |
+
print("Not Found")
|
648 |
+
|
649 |
+
dat = torch.cat(data, dim=1).to(device1)
|
650 |
+
|
651 |
+
oa = ""
|
652 |
+
|
653 |
+
loss = 0.00
|
654 |
+
|
655 |
+
with torch.no_grad():
|
656 |
+
|
657 |
+
model1.reset()
|
658 |
+
|
659 |
+
model2.reset()
|
660 |
+
|
661 |
+
model3.reset()
|
662 |
+
|
663 |
+
model4.reset()
|
664 |
+
|
665 |
+
model5.reset()
|
666 |
+
|
667 |
+
model6.reset()
|
668 |
+
|
669 |
+
oa = ""
|
670 |
+
|
671 |
+
for a in range(dat.shape[1] - 1):
|
672 |
+
|
673 |
+
out = model1(dat[:, a])
|
674 |
+
|
675 |
+
out = model2(out.to(device2))
|
676 |
+
|
677 |
+
out = model3(out.to(device3))
|
678 |
+
|
679 |
+
out = model4(out.to(device4))
|
680 |
+
|
681 |
+
out = model5(out.to(device5))
|
682 |
+
|
683 |
+
out = model6(out.to(device6))
|
684 |
+
|
685 |
+
out = output(out.to(device7))
|
686 |
+
|
687 |
+
sfo = torch.nn.functional.softmax(out, dim=-1)
|
688 |
+
|
689 |
+
wid = torch.argmax(sfo, dim=-1).item()
|
690 |
+
|
691 |
+
try:
|
692 |
+
|
693 |
+
wd = i2w[wid]
|
694 |
+
|
695 |
+
except:
|
696 |
+
|
697 |
+
oa = oa + "ERROR"
|
698 |
+
|
699 |
+
else:
|
700 |
+
oa = oa + wd
|
701 |
+
|
702 |
+
loss += lossf(out, trued[:, a+1].to(device2))
|
703 |
+
|
704 |
+
print("validationloss", loss.item() / dat.shape[1], "preview", oa)
|
705 |
+
if __name__ == "__main__":
|
706 |
+
#DataMake()
|
707 |
+
#Fineturning(Load=False,dim=2048, outputdim=21000,lr=1e-03, onestep=300, uselen=128)
|
708 |
+
#Predict(dim=2048, outputdim=21000, maxlen=128)
|
3BSanokaKai2/LLM1.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:012d20b8fc0d8af6f4db67bafb2a28ca59a2ee56423eac2e601a1697beffe298
|
3 |
+
size 604773006
|
3BSanokaKai2/LLM2.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:4ee5932033baa47419e20a437951b9afe371d083837e4f687ad8222116b41936
|
3 |
+
size 604244122
|
3BSanokaKai2/LLM3.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:57b7e2c4dd0e3fa6a05bdf75bdb1d9668586e9bc9cc542e76475649f699cb480
|
3 |
+
size 604244122
|
3BSanokaKai2/LLM4.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:b064b7d5fdd79e6e78a754c3382806b5c52eea990ad6df864a056b79c512fb2e
|
3 |
+
size 604244122
|
3BSanokaKai2/LLM5.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:f448ad6bd54de6dbc6210a1846358ff7ed77a3741d51f4d2d277b0b9c55879a1
|
3 |
+
size 604244122
|
3BSanokaKai2/LLM6.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:9bb731e1a1df5a2e205df89e1ceb5d856a81e871d88faf1ebb3ec05bb3880be7
|
3 |
+
size 604244122
|
3BSanokaKai2/licence.txt
ADDED
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Copyright (c) <2024> <Apfel X:@KyoumeiProject>
|
2 |
+
|
3 |
+
以下に定める条件に従い、本ソフトウェアおよび関連文書のファイル(以下「ソフトウェア」)の複製を取得するすべての人に対し、ソフトウェアを無制限に扱うことを無償で許可します。これには、ソフトウェアの複製を使用、複写、変更、結合、掲載、頒布、サブライセンス、および/または販売する権利、およびソフトウェアを提供する相手に同じことを許可する権利も無制限に含まれます。
|
4 |
+
|
5 |
+
上記の著作権表示および本許諾表示を、ソフトウェアのすべての複製または重要な部分に記載するものとします。
|
6 |
+
|
7 |
+
ソフトウェアは「現状のまま」で、明示であるか暗黙であるかを問わず、何らの保証もなく提供されます。ここでいう保証とは、商品性、特定の目的への適合性、および権利非侵害についての保証も含みますが、それに限定されるものではありません。 作者または著作権者は、契約行為、不法行為、またはそれ以外であろうと、ソフトウェアに起因または関連し、あるいはソフトウェアの使用またはその他の扱いによって生じる一切の請求、損害、その他の義務について何らの責任も負わないものとします。
|
3BSanokaKai2/output.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:4d62f04138b99cc0735f102e2179672be006c45a37f4b69342f3389b939fba28
|
3 |
+
size 86059474
|
3BSanokaKai2/readme.txt
ADDED
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
AI-Large.pyがトレーニングコードです。
|
2 |
+
ファインチューニング済みデータがないので、
|
3 |
+
ファインチューニング関数を用意しています。
|
4 |
+
|
5 |
+
警告:FT含め学習はメインメモリを128GB積んでいないマシンを推奨。ブルスク出すかもしれません。
|
6 |
+
注意:GPUを7台使用する設定になっています。もし変更したい場合は"cuda:n"となっている所を探し、希望のGPU番号、またはcpuを選択してください。
|
7 |
+
|
8 |
+
使用ライブラリ
|
9 |
+
|
10 |
+
|
11 |
+
import MeCab
|
12 |
+
import unidic
|
13 |
+
import torch
|
14 |
+
import copy
|
15 |
+
import time
|
16 |
+
import matplotlib.pyplot as plt
|
17 |
+
import re
|
18 |
+
import math
|
19 |
+
import numpy as np
|
20 |
+
from gensim.models import Word2Vec
|
21 |
+
import pickle
|
22 |
+
import threading
|
23 |
+
import sentencepiece
|
24 |
+
|
25 |
+
# ファインチューニングの方法
|
26 |
+
まず、「train_data.txt」と言うファイルを用意します。
|
27 |
+
その中に、ファインチューニング用のデータを用意してください。
|
28 |
+
train_data.txtは、改行ごとに別の時系列として扱われます。
|
29 |
+
train_data.txtを用意したら、AI-Large.pyを実行してください。
|
30 |
+
実行すると、DataMake()関数により、学習データがベクトル化されます。
|
31 |
+
次ににFineturning()を実行されます。
|
32 |
+
これで学習が行われます。
|
33 |
+
学習が始まると出力サンプルが表示されるので、ある程度の日本語になったらctrl+cを使い止めましょう。
|
34 |
+
最初は、50epochと表示される位でctrl+cを実行することをお勧めします。
|
35 |
+
これでfineturning.pthが生成されます。
|
36 |
+
最後に、Fineturning()とDataMake()をコメントアウトし、Predict()を実行すると、使用できます。
|
37 |
+
「Q>」と表示されるので、そこに入力を入れましょう。
|
38 |
+
そうすると「A>」の横に出力が出るはずです。(FT不足だと、何も出力されない場合があります。)
|
3BSanokaKai2/table.txt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:90a6e7244d7c9f6d7baaac0fde820a1eb41724e6a0f9fdd793f00e5c02b62059
|
3 |
+
size 27069230
|
3BSanokaKai2/tokenizer.model
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:8e0594d183dc437f0b24fd52db43c8ef068d39c0f5bdec0cc1fd5b867214675f
|
3 |
+
size 577009
|
3BSanokaKai2/tokenizer.vocab
ADDED
The diff for this file is too large to render.
See raw diff
|
|
3BSanokaKai2/word2vec.model
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:b70326897d6913da9aa1fc2e837e7531458359740bae17580c8c9d82a7782efe
|
3 |
+
size 21157728
|