maker.py

#! /usr/bin/python3
src="Kendamarron/Tokara-0.5B-v0.1"
tgt="KoichiYasuoka/Tokara-0.5B-ud-embeds"
url="https://github.com/UniversalDependencies/UD_Japanese-GSDLUW"
import os
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")
class UDEmbedsDataset(object):
  def __init__(self,conllu,tokenizer,oldtokenizer=None,embeddings=None):
    self.conllu=open(conllu,"r",encoding="utf-8")
    self.tokenizer=tokenizer
    self.oldtokenizer=oldtokenizer if oldtokenizer else tokenizer
    self.embeddings=embeddings
    self.seeks=[0]
    label=set(["SYM","SYM.","SYM|_"])
    dep=set()
    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]
            for k in [p,p+".","B-"+p,"B-"+p+".","I-"+p,"I-"+p+".",p+q+"|_",p+q+d]:
              label.add(k)
      s=self.conllu.readline()
    self.label2id={l:i for i,l in enumerate(sorted(label))}
  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):
    import torch
    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
    x=[True if t[6]=="0" or int(t[6])>j or sum([1 if int(c[i][6])==j+1 else 0 for i in range(j+1,len(c))])>0 else False for j,t in enumerate(c)]
    v=self.tokenizer([t[1] for t in c],add_special_tokens=False)["input_ids"]
    ids,upos=[],[]
    for i,(j,k) in enumerate(zip(v,c)):
      if j==[]:
        j=self.tokenizer("〓")["input_ids"]
      p=k[3] if x[i] else k[3]+"."
      ids+=j
      upos+=[p] if len(j)==1 else ["B-"+p]+["I-"+p]*(len(j)-1)
    v=self.oldtokenizer([t[1] for t in c],add_special_tokens=False)["input_ids"]
    p=[t[3] if t[5]=="_" else t[3]+"|"+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]
    idx=[-1]
    upos.append("SYM|_")
    for i in range(len(x)):
      idx.append(i)
      upos.append(p[i]+"|"+d[i] if c[i][6]=="0" else p[i]+"|_")
      for j in range(i+1,len(x)):
        idx.append(j)
        upos.append(p[j]+"|"+d[j] if int(c[j][6])==i+1 else p[i]+"|"+d[i] if int(c[i][6])==j+1 else p[j]+"|_")
      idx.append(-1)
      upos.append("SYM|_")
    with torch.no_grad():
      m=[]
      for j in v:
        if j==[]:
          j=self.tokenizer("〓")["input_ids"]
        m.append(self.embeddings[j,:].sum(axis=0))
      m.append(self.embeddings[self.tokenizer("<|im_end|>")["input_ids"][0],:])
      emb=torch.stack(m)
    return{"inputs_embeds":torch.vstack((self.embeddings[ids,:],emb[idx,:])),"labels":[self.label2id[p] for p in upos]}
from transformers import AutoTokenizer,AutoConfig,AutoModelForTokenClassification,DefaultDataCollator,TrainingArguments,Trainer
from tokenizers.pre_tokenizers import Sequence,Split
from tokenizers import Regex
from copy import deepcopy
otk=AutoTokenizer.from_pretrained(src)
ntk=deepcopy(otk)
ntk.backend_tokenizer.pre_tokenizer=Sequence([Split(Regex("[ぁ-ん]"),"isolated"),otk.backend_tokenizer.pre_tokenizer])
trainDS=UDEmbedsDataset("train.conllu",ntk,otk)
devDS=UDEmbedsDataset("dev.conllu",ntk,otk)
testDS=UDEmbedsDataset("test.conllu",ntk,otk)
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)
trainDS.embeddings=mdl.get_input_embeddings().weight
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=DefaultDataCollator(),model=mdl,train_dataset=trainDS)
trn.train()
trn.save_model(tgt)
otk.save_pretrained(tgt)