maker.py

#! /usr/bin/python3
src="KoichiYasuoka/modernbert-large-thai-wikipedia-upos"
tgt="KoichiYasuoka/modernbert-large-thai-wikipedia-ud-embeds"
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 UDEmbedsDataset(object):
  def __init__(self,conllu,tokenizer,embeddings=None):
    self.conllu=open(conllu,"r",encoding="utf-8")
    self.tokenizer=tokenizer
    self.embeddings=embeddings
    self.seeks=[0]
    label=set(["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)*2
  def __getitem__(self,i):
    self.conllu.seek(self.seeks[int(i/2)])
    z,c,t,s=i%2,[],[""],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"]
    if z==0:
      ids,upos=[self.tokenizer.cls_token_id],["SYM."]
      for i,(j,k) in enumerate(zip(v,c)):
        if j==[]:
          j=[self.tokenizer.unk_token_id]
        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)
      ids.append(self.tokenizer.sep_token_id)
      upos.append("SYM.")
      emb=self.embeddings
    else:
      import torch
      if len(x)<128:
        x=[True]*len(x)
      else:
        w=sum([len(x)-i+1 if b else 0 for i,b in enumerate(x)])+1
        for i in range(len(x)):
          if x[i]==False and w+len(x)-i<8192:
            x[i]=True
            w+=len(x)-i+1
      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]
      ids,upos=[-1],["SYM|_"]
      for i in range(len(x)):
        if x[i]:
          ids.append(i)
          upos.append(p[i]+"|"+d[i] if c[i][6]=="0" else p[i]+"|_")
          for j in range(i+1,len(x)):
            ids.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]+"|_")
          ids.append(-1)
          upos.append("SYM|_")
      with torch.no_grad():
        m=[]
        for j in v:
          if j==[]:
            j=[self.tokenizer.unk_token_id]
          m.append(self.embeddings[j,:].sum(axis=0))
        m.append(self.embeddings[self.tokenizer.sep_token_id,:])
        emb=torch.stack(m)
    return{"inputs_embeds":emb[ids[:8192],:],"labels":[self.label2id[p] for p in upos[:8192]]}
from transformers import AutoTokenizer,AutoConfig,AutoModelForTokenClassification,DefaultDataCollator,TrainingArguments,Trainer
tkz=AutoTokenizer.from_pretrained(src)
trainDS=UDEmbedsDataset("train.conllu",tkz)
devDS=UDEmbedsDataset("dev.conllu",tkz)
testDS=UDEmbedsDataset("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)
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)
tkz.save_pretrained(tgt)