|
|
|
src="KoichiYasuoka/modernbert-base-thai-wikipedia-upos" |
|
tgt="KoichiYasuoka/modernbert-base-thai-wikipedia-ud-triangular" |
|
import os |
|
os.system("""D=spaCy-Thai/UD_Thai-Corpora |
|
test -d $D || git clone --depth=1 https://github.com/KoichiYasuoka/spaCy-Thai |
|
nawk 'BEGIN{FS=OFS="\\t"} |
|
{if(NF==10&&$1~/^[1-9][0-9]*$/||$0~/^# text =/)u=u$0"\\n"; |
|
else if($0==""){f=(FILENAME~/test/)?"test":(FILENAME~/dev/)?"dev":"train"; |
|
if(u~/\\t0\\troot\\t/)print u>f".conllu"; |
|
u=""} |
|
}' $D/*-ud-*.conllu""") |
|
class UDTriangularDataset(object): |
|
def __init__(self,conllu,tokenizer): |
|
self.conllu=open(conllu,"r",encoding="utf-8") |
|
self.tokenizer=tokenizer |
|
self.seeks=[0] |
|
label=set(["SYM|x","X|x"]) |
|
dep=set(["X|x|r-goeswith"]) |
|
s=self.conllu.readline() |
|
while s!="": |
|
if s=="\n": |
|
self.seeks.append(self.conllu.tell()) |
|
else: |
|
w=s.split("\t") |
|
if len(w)==10: |
|
if w[0].isdecimal(): |
|
p=w[3] |
|
q="" if w[5]=="_" else "|"+w[5] |
|
d=("|" if w[6]=="0" else "|l-" if int(w[0])<int(w[6]) else "|r-")+w[7] |
|
label.add(p+"|o"+q) |
|
label.add(p+"|x"+q) |
|
dep.add(p+"|o"+q+d) |
|
dep.add(p+"|x"+q+d) |
|
s=self.conllu.readline() |
|
lid={l:i for i,l in enumerate(sorted(label))} |
|
for i,d in enumerate(sorted(dep),len(lid)): |
|
lid[d]=i |
|
self.label2id=lid |
|
def __call__(*args): |
|
lid={l:i for i,l in enumerate(sorted(set(sum([list(t.label2id) for t in args],[]))))} |
|
for t in args: |
|
t.label2id=lid |
|
return lid |
|
def __del__(self): |
|
self.conllu.close() |
|
__len__=lambda self:len(self.seeks)-1 |
|
def __getitem__(self,i): |
|
self.conllu.seek(self.seeks[i]) |
|
c,t,s=[],[""],False |
|
while t[0]!="\n": |
|
t=self.conllu.readline().split("\t") |
|
if len(t)==10 and t[0].isdecimal(): |
|
if s: |
|
t[1]=" "+t[1] |
|
c.append(t) |
|
s=t[9].find("SpaceAfter=No")<0 |
|
v=self.tokenizer([t[1] for t in c],add_special_tokens=False)["input_ids"] |
|
for i in range(len(v)-1,-1,-1): |
|
for j in range(1,len(v[i])): |
|
c.insert(i+1,[c[i][0],"_","_","X","_","_",c[i][0],"goeswith","_","_"]) |
|
y=["0"]+[t[0] for t in c] |
|
h=[i if t[6]=="0" else y.index(t[6]) for i,t in enumerate(c,1)] |
|
x=["o" if k>i or sum([1 if j==i+1 else 0 for j in h[i+1:]])>0 else "x" for i,k in enumerate(h)] |
|
p=[t[3]+"|"+x[i] if t[5]=="_" else t[3]+"|"+x[i]+"|"+t[5] for i,t in enumerate(c)] |
|
d=[t[7] if t[6]=="0" else "l-"+t[7] if int(t[0])<int(t[6]) else "r-"+t[7] for t in c] |
|
v=sum(v,[]) |
|
ids=[self.tokenizer.cls_token_id] |
|
upos=["SYM|x"] |
|
for i,k in enumerate(v): |
|
if len(v)<128 or x[i]=="o": |
|
ids.append(k) |
|
upos.append(p[i]+"|"+d[i] if h[i]==i+1 else p[i]) |
|
for j in range(i+1,len(v)): |
|
ids.append(v[j]) |
|
upos.append(p[j]+"|"+d[j] if h[j]==i+1 else p[i]+"|"+d[i] if h[i]==j+1 else p[j]) |
|
ids.append(self.tokenizer.sep_token_id) |
|
upos.append("SYM|x") |
|
return {"input_ids":ids[:8192],"labels":[self.label2id[p] for p in upos[:8192]]} |
|
from transformers import AutoTokenizer,AutoConfig,AutoModelForTokenClassification,DataCollatorForTokenClassification,TrainingArguments,Trainer |
|
tkz=AutoTokenizer.from_pretrained(src) |
|
trainDS=UDTriangularDataset("train.conllu",tkz) |
|
devDS=UDTriangularDataset("dev.conllu",tkz) |
|
testDS=UDTriangularDataset("test.conllu",tkz) |
|
lid=trainDS(devDS,testDS) |
|
cfg=AutoConfig.from_pretrained(src,num_labels=len(lid),label2id=lid,id2label={i:l for l,i in lid.items()},ignore_mismatched_sizes=True,trust_remote_code=True) |
|
mdl=AutoModelForTokenClassification.from_pretrained(src,config=cfg,ignore_mismatched_sizes=True,trust_remote_code=True) |
|
arg=TrainingArguments(num_train_epochs=3,per_device_train_batch_size=1,dataloader_pin_memory=False,output_dir=tgt,overwrite_output_dir=True,save_total_limit=2,learning_rate=5e-05,warmup_ratio=0.1,save_safetensors=False) |
|
trn=Trainer(args=arg,data_collator=DataCollatorForTokenClassification(tkz),model=mdl,train_dataset=trainDS) |
|
trn.train() |
|
trn.save_model(tgt) |
|
tkz.save_pretrained(tgt) |
|
|
