initial release

README.md

@@ -0,0 +1,29 @@
+---
+language:
+- "pt"
+tags:
+- "portuguese"
+- "token-classification"
+- "pos"
+- "dependency-parsing"
+base_model: eliasjacob/ModernBERT-large-portuguese
+datasets:
+- "universal_dependencies"
+license: "apache-2.0"
+pipeline_tag: "token-classification"
+---
+
+# modernbert-large-portuguese-ud-embeds
+
+## Model Description
+
+This is a ModernBERT model for POS-tagging and dependency-parsing, derived from [ModernBERT-large-portuguese](https://huggingface.co/eliasjacob/ModernBERT-large-portuguese).
+
+## How to Use
+
+```py
+from transformers import pipeline
+nlp=pipeline("universal-dependencies","KoichiYasuoka/modernbert-large-portuguese-ud-embeds",trust_remote_code=True)
+print(nlp("Foi quebrado pelo pelo do gato"))
+```
+

maker.py

@@ -0,0 +1,133 @@
+#! /usr/bin/python3
+import os
+src="eliasjacob/ModernBERT-large-portuguese"
+tgt="KoichiYasuoka/modernbert-large-portuguese-ud-embeds"
+url="https://github.com/UniversalDependencies/UD_Portuguese-"
+for e in ["Bosque","GSD","PetroGold"]:
+  u=url+e
+  d=os.path.basename(u)
+  os.system("test -d "+d+" || git clone --depth=1 "+u)
+s='BEGIN{FS=OFS="\\t";a[1]=a[2]=""};{if(NF==10){if($1~/-/&&$2~/-/)split($1,a,"-");else{if($1==a[1])$10="SpaceAfter=No";else if($1==a[2])$2="-"$2;print}}else{print;a[1]=a[2]=""}}'
+os.system("for F in train dev test ; do nawk '"+s+"' UD_Portuguese-*/*-$F.conllu >$F.conllu ; done")
+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.","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,e,w,m=i%2,[],[""],"_","_",None,False
+    while t[0]!="\n":
+      t=self.conllu.readline().split("\t") 
+      if len(t)==10:
+        if t[0].isdecimal():
+          if m:
+            t[1]=" "+t[1] 
+          if w==None:
+            c.append(t)
+            m=t[9].find("SpaceAfter=No")
+          elif s==t[0]:
+            t[1]=w
+            t[6]="0"
+            c.append(t)
+          elif e==t[0]:
+            s=e="_"
+            w=None
+        elif z==0:
+          k=t[0].split("-")
+          if len(k)==2:
+            s,e=k
+            w=" "+t[1] if m else t[1]
+            m=t[9].find("SpaceAfter=No")
+    x=[True if t[6]=="0" or int(t[6])>int(t[0]) or sum([1 if int(c[i][6])==int(t[0]) 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)<127:
+        x=[True]*len(x)
+        w=(len(x)+2)*(len(x)+1)/2
+      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]+"|_")
+          if w>8192 and i>0:
+            while w>8192 and upos[-1].endswith("|_"):
+              upos.pop(-1)
+              ids.pop(-1)
+              w-=1
+          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
+from tokenizers.pre_tokenizers import Sequence,Punctuation
+tkz=AutoTokenizer.from_pretrained(src)
+tkz.backend_tokenizer.pre_tokenizer=Sequence([Punctuation(),tkz.backend_tokenizer.pre_tokenizer])
+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)

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cc56e811d16418f8000018ef672be70765f5e2ed897412183467c091d7290d92
+size 1598942274

special_tokens_map.json

@@ -0,0 +1,37 @@
+{
+  "cls_token": {
+    "content": "[CLS]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "mask_token": {
+    "content": "[MASK]",
+    "lstrip": true,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "[PAD]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "sep_token": {
+    "content": "[SEP]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "[UNK]",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer_config.json

@@ -0,0 +1,73 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<|padding|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "[UNK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "3": {
+      "content": "[CLS]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "4": {
+      "content": "[SEP]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "5": {
+      "content": "[PAD]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "6": {
+      "content": "[MASK]",
+      "lstrip": true,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "clean_up_tokenization_spaces": true,
+  "cls_token": "[CLS]",
+  "extra_special_tokens": {},
+  "mask_token": "[MASK]",
+  "model_input_names": [
+    "input_ids",
+    "attention_mask"
+  ],
+  "model_max_length": 8192,
+  "pad_token": "[PAD]",
+  "sep_token": "[SEP]",
+  "tokenizer_class": "PreTrainedTokenizerFast",
+  "unk_token": "[UNK]"
+}

ud.py

@@ -0,0 +1,197 @@
+import numpy
+from transformers import TokenClassificationPipeline
+
+class BellmanFordTokenClassificationPipeline(TokenClassificationPipeline):
+  def __init__(self,**kwargs):
+    super().__init__(**kwargs)
+    x=self.model.config.label2id
+    y=[k for k in x if k.find("|")<0 and not k.startswith("I-")]
+    self.transition=numpy.full((len(x),len(x)),-numpy.inf)
+    self.ilabel=numpy.full(len(x),-numpy.inf)
+    self.slabel=numpy.full(len(x),-numpy.inf)
+    for k,v in x.items():
+      if k.find("|")<0:
+        for j in ["I-"+k[2:]] if k.startswith("B-") else [k]+y if k.startswith("I-") else y:
+          self.transition[v,x[j]]=0
+        if k.startswith("I-"):
+          self.ilabel[v]=0
+        elif k.startswith("SYM"):
+          self.slabel[v]=0
+  def check_model_type(self,supported_models):
+    pass
+  def postprocess(self,model_outputs,**kwargs):
+    if "logits" not in model_outputs:
+      return self.postprocess(model_outputs[0],**kwargs)
+    return self.bellman_ford_token_classification(model_outputs,**kwargs)
+  def bellman_ford_token_classification(self,model_outputs,**kwargs):
+    m=model_outputs["logits"][0].numpy()
+    x=model_outputs["offset_mapping"][0].tolist()
+    for i,(s,e) in enumerate(x):
+      if s==0 and e==0:
+        m[i]+=self.slabel
+      elif i>0 and s<e and x[i-1][1]>s:
+        m[i]+=self.ilabel
+    e=numpy.exp(m-numpy.max(m,axis=-1,keepdims=True))
+    z=e/e.sum(axis=-1,keepdims=True)
+    for i in range(m.shape[0]-1,0,-1):
+      m[i-1]+=numpy.max(m[i]+self.transition,axis=1)
+    k=[numpy.argmax(m[0]+self.transition[0])]
+    for i in range(1,m.shape[0]):
+      k.append(numpy.argmax(m[i]+self.transition[k[-1]]))
+    w=[{"entity":self.model.config.id2label[j],"start":s,"end":e,"score":z[i,j]} for i,((s,e),j) in enumerate(zip(x,k)) if s<e]
+    if "aggregation_strategy" in kwargs and kwargs["aggregation_strategy"]!="none":
+      for i,t in reversed(list(enumerate(w))):
+        p=t.pop("entity")
+        if i>0 and p.startswith("I-"):
+          w[i-1]["score"]=min(w[i-1]["score"],t["score"])
+          w[i-1]["end"]=w.pop(i)["end"]
+        elif i>0 and w[i-1]["end"]>t["start"]:
+          w[i-1]["score"]=min(w[i-1]["score"],t["score"])
+          w[i-1]["end"]=w.pop(i)["end"]
+        elif p.startswith("B-"):
+          t["entity_group"]=p[2:]
+        else:
+          t["entity_group"]=p
+    for t in w:
+      t["text"]=model_outputs["sentence"][t["start"]:t["end"]]
+    return w
+
+class UniversalDependenciesPipeline(BellmanFordTokenClassificationPipeline):
+  def __init__(self,**kwargs):
+    kwargs["aggregation_strategy"]="simple"
+    super().__init__(**kwargs)
+    x=self.model.config.label2id
+    self.root=numpy.full((len(x)),-numpy.inf)
+    self.left_arc=numpy.full((len(x)),-numpy.inf)
+    self.right_arc=numpy.full((len(x)),-numpy.inf)
+    for k,v in x.items():
+      if k.endswith("|root"):
+        self.root[v]=0
+      elif k.find("|l-")>0:
+        self.left_arc[v]=0
+      elif k.find("|r-")>0:
+        self.right_arc[v]=0
+    self.multiword={}
+    if self.model.config.task_specific_params:
+      if "upos_multiword" in self.model.config.task_specific_params:
+        self.multiword=self.model.config.task_specific_params["upos_multiword"]
+  def postprocess(self,model_outputs,**kwargs):
+    import torch
+    kwargs["aggregation_strategy"]="simple"
+    if "logits" not in model_outputs:
+      return self.postprocess(model_outputs[0],**kwargs)
+    w=self.bellman_ford_token_classification(model_outputs,**kwargs)
+    off=[(t["start"],t["end"]) for t in w]
+    for i,(s,e) in reversed(list(enumerate(off))):
+      if s<e:
+        d=w[i]["text"]
+        j=len(d)-len(d.lstrip())
+        if j>0:
+          d=d.lstrip()
+          off[i]=(off[i][0]+j,off[i][1])
+        j=len(d)-len(d.rstrip())
+        if j>0:
+          d=d.rstrip()
+          off[i]=(off[i][0],off[i][1]-j)
+        if d.strip()=="":
+          off.pop(i)
+          w.pop(i)
+        else:
+          p=w[i]["entity_group"]
+          if p in self.multiword:
+            d=d.lower()
+            if d in self.multiword[p]:
+              j=self.multiword[p][d]
+              if "".join(j)==d:
+                for k in reversed(j[1:]):
+                  e=off[i][1]
+                  w.insert(i+1,{"start":e-len(k),"end":e,"text":k,"entity_group":"","score":w[i]["score"]})
+                  off.insert(i+1,(e-len(k),e))
+                  w[i]["end"]=e-len(k)
+                  off[i]=(off[i][0],e-len(k))
+                w[i]["text"]=" "+j[0]
+                w[i]["entity_group"]=""
+              else:
+                s,e=off[i]
+                for k in reversed(j[1:]):
+                  w.insert(i+1,{"start":s,"end":e,"text":" "+k,"entity_group":"+","score":w[i]["score"]})
+                  off.insert(i+1,(s,e))
+                w[i]["text"]=" "+j[0]
+                w[i]["entity_group"]=f"+{len(j)}"
+    v=self.tokenizer([t["text"] for t in w],add_special_tokens=False)
+    x=[not t["entity_group"].endswith(".") for t in w]
+    if len(x)<127:
+      x=[True]*len(x)
+    else:
+      k=sum([len(x)-i+1 if b else 0 for i,b in enumerate(x)])+1
+      for i in numpy.argsort(numpy.array([t["score"] for t in w])):
+        if x[i]==False and k+len(x)-i<8192:
+          x[i]=True
+          k+=len(x)-i+1
+    ids=[-1]
+    for i in range(len(x)):
+      if x[i]:
+        ids.append(i)
+        for j in range(i+1,len(x)):
+          ids.append(j)
+        ids.append(-1)
+    with torch.no_grad():
+      e=self.model.get_input_embeddings().weight
+      m=[]
+      for j in v["input_ids"]:
+        if j==[]:
+          j=[self.tokenizer.unk_token_id]
+        m.append(e[j,:].sum(axis=0))
+      m.append(e[self.tokenizer.sep_token_id,:])
+      m=torch.stack(m).to(self.device)
+      e=self.model(inputs_embeds=torch.unsqueeze(m[ids,:],0))
+    m=e.logits[0].cpu().numpy()
+    e=numpy.full((len(x),len(x),m.shape[-1]),m.min())
+    k=1
+    for i in range(len(x)):
+      if x[i]:
+        e[i,i]=m[k]+self.root
+        k+=1
+        for j in range(1,len(x)-i):
+          e[i+j,i]=m[k]+self.left_arc
+          e[i,i+j]=m[k]+self.right_arc
+          k+=1
+        k+=1
+    m,p=numpy.max(e,axis=2),numpy.argmax(e,axis=2)
+    h=self.chu_liu_edmonds(m)
+    z=[i for i,j in enumerate(h) if i==j]
+    if len(z)>1:
+      k,h=z[numpy.argmax(m[z,z])],numpy.min(m)-numpy.max(m)
+      m[:,z]+=[[0 if j in z and (i!=j or i==k) else h for i in z] for j in range(m.shape[0])]
+      h=self.chu_liu_edmonds(m)
+    q=[self.model.config.id2label[p[j,i]].split("|") for i,j in enumerate(h)]
+    t=model_outputs["sentence"].replace("\n"," ")
+    u="# text = "+t+"\n"
+    for i,(s,e) in enumerate(off):
+      m=w[i]["entity_group"]
+      if m.startswith("+"):
+        if m!="+":
+          u+="\t".join([f"{i+1}-{i+int(m)}",t[s:e],"_","_","_","_","_","_","_","_" if i+int(m)<len(off) and e<off[i+int(m)][0] else "SpaceAfter=No"])+"\n"
+        u+="\t".join([str(i+1),w[i]["text"].strip(),"_",q[i][0],"_","_" if len(q[i])<3 else "|".join(q[i][1:-1]),str(0 if h[i]==i else h[i]+1),"root" if q[i][-1]=="root" else q[i][-1][2:],"_","_"])+"\n"
+      else:
+        u+="\t".join([str(i+1),t[s:e],"_",q[i][0],"_","_" if len(q[i])<3 else "|".join(q[i][1:-1]),str(0 if h[i]==i else h[i]+1),"root" if q[i][-1]=="root" else q[i][-1][2:],"_","_" if i+1<len(off) and e<off[i+1][0] else "SpaceAfter=No"])+"\n"
+    return u+"\n"
+  def chu_liu_edmonds(self,matrix):
+    h=numpy.argmax(matrix,axis=0)
+    x=[-1 if i==j else j for i,j in enumerate(h)]
+    for b in [lambda x,i,j:-1 if i not in x else x[i],lambda x,i,j:-1 if j<0 else x[j]]:
+      y=[]
+      while x!=y:
+        y=list(x)
+        for i,j in enumerate(x):
+          x[i]=b(x,i,j)
+      if max(x)<0:
+        return h
+    y,x=[i for i,j in enumerate(x) if j==max(x)],[i for i,j in enumerate(x) if j<max(x)]
+    z=matrix-numpy.max(matrix,axis=0)
+    m=numpy.block([[z[x,:][:,x],numpy.max(z[x,:][:,y],axis=1).reshape(len(x),1)],[numpy.max(z[y,:][:,x],axis=0),numpy.max(z[y,y])]])
+    k=[j if i==len(x) else x[j] if j<len(x) else y[numpy.argmax(z[y,x[i]])] for i,j in enumerate(self.chu_liu_edmonds(m))]
+    h=[j if i in y else k[x.index(i)] for i,j in enumerate(h)]
+    i=y[numpy.argmax(z[x[k[-1]],y] if k[-1]<len(x) else z[y,y])]
+    h[i]=x[k[-1]] if k[-1]<len(x) else i
+    return h