model improved

config.json

@@ -367,7 +367,7 @@
   "tie_word_embeddings": false,
   "tokenizer_class": "PreTrainedTokenizerFast",
   "torch_dtype": "float32",
-  "transformers_version": "4.42.4",
+  "transformers_version": "4.44.2",
   "use_cache": true,
   "use_parallel_residual": false,
   "vocab_size": 44416

maker.py

@@ -3,7 +3,7 @@ src="rinna/japanese-gpt-neox-small"
 tgt="KoichiYasuoka/rinna-gpt-neox-small-japanese-ud-causal"
 url="https://github.com/UniversalDependencies/UD_Japanese-GSDLUW"
 
-import os,json
+import os,json,unicodedata
 from transformers import AutoTokenizer,PreTrainedTokenizerFast,AutoConfig,GPTNeoXForTokenClassification,DefaultDataCollator,TrainingArguments,Trainer
 d=os.path.basename(url)
 os.system("test -d "+d+" || git clone --depth=1 "+url)
@@ -11,6 +11,7 @@ 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=2048)
 tkz.save_pretrained("tmpdir")
 d=json.loads(tkz.backend_tokenizer.to_str())
+tkz.backend_tokenizer.from_str(json.dumps(d)).save("tmpdir/oldtokenizer.json")
 form=set()
 with open("train.conllu","r",encoding="utf-8") as r:
   for s in r:
@@ -20,13 +21,17 @@ with open("train.conllu","r",encoding="utf-8") as r:
 for t in d["model"]["vocab"]:
   if t[0] not in form:
     t[1]*=len(t[0])
+  elif len(t[0])>1 and unicodedata.name(t[0][0]).startswith("HIRAGANA"):
+    t[1]*=len(t[0])
 tkz.backend_tokenizer.from_str(json.dumps(d)).save("tmpdir/tokenizer.json")
-tkz=PreTrainedTokenizerFast.from_pretrained("tmpdir")
+ntk=PreTrainedTokenizerFast.from_pretrained("tmpdir")
+otk=PreTrainedTokenizerFast.from_pretrained("tmpdir",tokenizer_file="tmpdir/oldtokenizer.json")
 
 class UDCausalDataset(object):
-  def __init__(self,conllu,tokenizer,embeddings=None):
+  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.max_tokens=3
     self.seeks=[(0,0)]
@@ -71,8 +76,8 @@ class UDCausalDataset(object):
         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:
+      v=self.tokenizer(form,add_special_tokens=False)
       i,u=[],[]
       for j,(x,y) in enumerate(zip(v["input_ids"],upos)):
         if x!=[]:
@@ -82,6 +87,7 @@ class UDCausalDataset(object):
       pad=self.tokenizer.pad_token_id
     else:
       import torch
+      v=self.oldtokenizer(form,add_special_tokens=False)
       m=[]
       for x in v["input_ids"]:
         if x==[]:
@@ -109,9 +115,9 @@ class UDCausalDataset(object):
       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)
+trainDS=UDCausalDataset("train.conllu",ntk,otk)
+devDS=UDCausalDataset("dev.conllu",ntk,otk)
+testDS=UDCausalDataset("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)
 mdl=GPTNeoXForTokenClassification.from_pretrained(src,config=cfg,ignore_mismatched_sizes=True)
@@ -121,4 +127,4 @@ arg=TrainingArguments(num_train_epochs=3,per_device_train_batch_size=32,dataload
 trn=Trainer(args=arg,data_collator=DefaultDataCollator(),model=mdl,train_dataset=trainDS)
 trn.train()
 trn.save_model(tgt)
-tkz.save_pretrained(tgt)
+ntk.save_pretrained(tgt)

pytorch_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:bb8c94b610e3cb65a776ba947f8ad8edd6534aa050ebff4fbc5283b99e2980c7
+oid sha256:8fffdfbf75da7de6f3a3c543821983faf9a3efc331f88476a253d2f527f3094a
 size 477227998

ud.py

@@ -1,5 +1,10 @@
 import numpy
-from transformers import TokenClassificationPipeline
+from transformers import TokenClassificationPipeline,AutoTokenizer
+try:
+  from transformers.utils import cached_file
+except:
+  from transformers.file_utils import cached_path,hf_bucket_url
+  cached_file=lambda x,y:os.path.join(x,y) if os.path.isdir(x) else cached_path(hf_bucket_url(x,y))
 
 class BellmanFordTokenClassificationPipeline(TokenClassificationPipeline):
   def __init__(self,**kwargs):
@@ -42,6 +47,7 @@ class UniversalDependenciesCausalPipeline(BellmanFordTokenClassificationPipeline
   def __init__(self,**kwargs):
     kwargs["aggregation_strategy"]="simple"
     super().__init__(**kwargs)
+    self.oldtokenizer=AutoTokenizer.from_pretrained(self.tokenizer.name_or_path,tokenizer_file=cached_file(self.tokenizer.name_or_path,"oldtokenizer.json"))
     x=self.model.config.label2id
     self.root=numpy.full((len(x)),numpy.nan)
     self.left_arc=numpy.full((len(x)),numpy.nan)
@@ -87,7 +93,7 @@ class UniversalDependenciesCausalPipeline(BellmanFordTokenClassificationPipeline
       if d[i].strip()=="":
         d.pop(i)
         w.pop(i)
-    v=self.tokenizer(d,add_special_tokens=False)
+    v=self.oldtokenizer(d,add_special_tokens=False)
     e=self.model.get_input_embeddings().weight
     m=[]
     for x in v["input_ids"]: