maker.sh

#! /bin/sh
test -f ja_gsd_modern.conllu || curl -LO https://github.com/KoichiYasuoka/SuPar-UniDic/raw/main/suparunidic/suparmodels/ja_gsd_modern.conllu
curl -L https://huggingface.co/KoichiYasuoka/Swallow-MS-7b-upos/resolve/main/tokenizer.json | env LANG=ja_JP.utf8 egrep -v '"[ぁ-ん] [ぁ-ん]",$' > newtokenizer.json

TMP=./maker$$.py
cat << 'EOF' > $TMP
#! /usr/bin/env deepspeed
src="KoichiYasuoka/Swallow-MS-7b-upos"
tgt="KoichiYasuoka/Swallow-MS-7b-char-upos"
from transformers import LlamaTokenizerFast,MistralForTokenClassification,AutoConfig,DataCollatorForTokenClassification,TrainingArguments,Trainer

class UPOSFileDataset(object):
  def __init__(self,conllu,tokenizer):
    self.conllu=open(conllu,"r",encoding="utf-8")
    self.tokenizer=tokenizer
    self.seeks=[0]
    self.multiword={}
    label=set(["SYM"])
    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():
            label.add(w[3] if w[5]=="_" else w[3]+"|"+w[5])
          elif w[0].find("-")>0:
            t=w[0].split("-")
            f,j,k=w[1],[],[]
            for i in range(int(t[0]),int(t[1])+1):
              w=self.conllu.readline().split("\t")
              j.append(w[3] if w[5]=="_" else w[3]+"|"+w[5])
              k.append(w[1])
            p="+".join(j)
            label.add(p)
            if p in self.multiword:
              self.multiword[p][f]=list(k)
            else:
              self.multiword[p]={f:list(k)}
      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
    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])
    form,upos=[],[]
    while self.conllu.tell()<self.seeks[i+1]:
      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])
        elif w[0].find("-")>0:
          t=w[0].split("-")
          u=[]
          for j in range(int(t[0]),int(t[1])+1):
            k=self.conllu.readline().split("\t")
            u.append(k[3] if k[5]=="_" else k[3]+"|"+k[5])
          upos.append("+".join(u))
    v=self.tokenizer(form,add_special_tokens=False)
    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)
    if len(i)<self.tokenizer.model_max_length-3:
      ids=[self.tokenizer.cls_token_id]+i+[self.tokenizer.sep_token_id]
      upos=["SYM"]+u+["SYM"]
    else:
      ids=i[0:self.tokenizer.model_max_length-2]
      upos=u[0:self.tokenizer.model_max_length-2]
    return {"input_ids":ids,"labels":[self.label2id[t] for t in upos]}

tkz=LlamaTokenizerFast.from_pretrained(src,tokenizer_file="newtokenizer.json")
trainDS=UPOSFileDataset("ja_gsd_modern.conllu",tkz)
lid=trainDS.label2id
cfg=AutoConfig.from_pretrained(src,num_labels=len(lid),label2id=lid,id2label={i:l for l,i in lid.items()},ignore_mismatched_sizes=True)
dsp={"fp16":{"enabled":"auto"},"optimizer":{"type":"AdamW"},"scheduler":{"type":"WarmupLR","params":{}},"train_batch_size":"auto","train_micro_batch_size_per_gpu":"auto","zero_optimization":{"stage":3,"offload_optimizer":{"device":"cpu","pin_memory":True},"offload_param":{"device":"cpu","pin_memory":True},"overlap_comm":True,"contiguous_gradients":True,"reduce_bucket_size":"auto","stage3_prefetch_bucket_size":"auto","stage3_param_persistence_threshold":"auto","stage3_gather_16bit_weights_on_model_save":True}}
arg=TrainingArguments(num_train_epochs=3,per_device_train_batch_size=8,deepspeed=dsp,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=MistralForTokenClassification.from_pretrained(src,config=cfg,ignore_mismatched_sizes=True),train_dataset=trainDS)
trn.train()
trn.save_model(tgt)
tkz.save_pretrained(tgt)
EOF
chmod 755 $TMP
$TMP
exit