KoichiYasuoka's picture
release after the tokenizer refined
03272ba
raw
history blame
4.8 kB
#! /usr/bin/python3
src="rinna/japanese-gpt2-xsmall"
tgt="KoichiYasuoka/rinna-gpt2-xsmall-japanese-ud-causal"
url="https://github.com/UniversalDependencies/UD_Japanese-GSDLUW"
import os,json
from transformers import AutoTokenizer,PreTrainedTokenizerFast,AutoConfig,GPT2ForTokenClassification,DefaultDataCollator,TrainingArguments,Trainer
d=os.path.basename(url)
os.system("test -d "+d+" || git clone --depth=1 "+url)
os.system("for F in train dev test ; do cp "+d+"/*-$F.conllu $F.conllu ; done")
tkz=AutoTokenizer.from_pretrained(src,add_prefix_space=False,legacy=False,model_max_length=1024)
tkz.save_pretrained("tmpdir")
d=json.loads(tkz.backend_tokenizer.to_str())
d["normalizer"]["normalizers"].append({"type":"Lowercase"})
form=set()
with open("train.conllu","r",encoding="utf-8") as r:
for s in r:
w=s.split("\t")
if len(w)==10 and w[0].isdecimal():
form.add(w[1])
for t in d["model"]["vocab"]:
if t[0] not in form:
t[1]*=len(t[0])
tkz.backend_tokenizer.from_str(json.dumps(d)).save("tmpdir/tokenizer.json")
tkz=PreTrainedTokenizerFast.from_pretrained("tmpdir")
class UDCausalDataset(object):
def __init__(self,conllu,tokenizer,embeddings=None):
self.conllu=open(conllu,"r",encoding="utf-8")
self.tokenizer=tokenizer
self.embeddings=embeddings
self.max_tokens=3
self.seeks=[(0,0)]
label=set(["SYM"])
dep=set()
s=self.conllu.readline()
while s!="":
if s=="\n":
self.seeks.append((self.conllu.tell(),0))
else:
w=s.split("\t")
if len(w)==10:
if w[0].isdecimal():
p=w[3] if w[5]=="_" else w[3]+"|"+w[5]
label.add(p)
dep.add(p+("|" if w[6]=="0" else "|l-" if int(w[0])<int(w[6]) else "|r-")+w[7])
self.seeks.append((self.seeks[-1][0],int(w[0])))
self.max_tokens=max(self.max_tokens,int(w[0])*2+1)
s=self.conllu.readline()
lid={}
for i,l in enumerate(sorted(label)):
lid[l],lid["B-"+l],lid["I-"+l]=i*3,i*3+1,i*3+2
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):
s,t=self.seeks[i]
self.conllu.seek(s)
form,upos,deps,w=[],[],[],[""]
while w[0]!="\n":
w=self.conllu.readline().split("\t")
if len(w)==10:
form.append(w[1])
if w[0].isdecimal():
upos.append(w[3] if w[5]=="_" else w[3]+"|"+w[5])
deps.append((int(w[6]),w[7]))
v=self.tokenizer(form,add_special_tokens=False)
if t==0:
i,u=[],[]
for j,(x,y) in enumerate(zip(v["input_ids"],upos)):
if x!=[]:
i+=x
u+=[y] if len(x)==1 else ["B-"+y]+["I-"+y]*(len(x)-1)
emb=self.embeddings
pad=self.tokenizer.pad_token_id
else:
import torch
m=[]
for x in v["input_ids"]:
if x==[]:
m.append(self.embeddings[self.tokenizer.unk_token_id,:])
else:
m.append(self.embeddings[x,:].sum(axis=0))
m.append(self.embeddings[self.tokenizer.sep_token_id,:])
m.append(self.embeddings[self.tokenizer.pad_token_id,:])
emb=torch.stack(m)
i,u=list(range(len(upos)+1)),upos+["SYM"]
i.append(t-1)
k,d=deps[t-1]
u.append(upos[t-1]+"|"+d if k==0 else upos[t-1])
for j in range(t,len(upos)):
i.append(j)
a,b=deps[j]
u.append(upos[j]+"|r-"+b if a==t else upos[t-1]+"|l-"+d if j+1==k else upos[j])
pad=-1
j=self.max_tokens-len(i)
if j>0:
ids=i+[pad]*j
upos=u+["SYM"]*j
else:
ids=i[0:self.max_tokens]
upos=u[0:self.max_tokens]
return {"inputs_embeds":emb[ids,:],"labels":[self.label2id[p] for p in upos]}
trainDS=UDCausalDataset("train.conllu",tkz)
devDS=UDCausalDataset("dev.conllu",tkz)
testDS=UDCausalDataset("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)
mdl=GPT2ForTokenClassification.from_pretrained(src,config=cfg,ignore_mismatched_sizes=True)
trainDS.embeddings=mdl.get_input_embeddings().weight
trainDS.max_tokens=min(trainDS.max_tokens,cfg.max_position_embeddings)
arg=TrainingArguments(num_train_epochs=3,per_device_train_batch_size=32,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=DefaultDataCollator(),model=mdl,train_dataset=trainDS)
trn.train()
trn.save_model(tgt)
tkz.save_pretrained(tgt)