from transformers import TokenClassificationPipeline

class UniversalDependenciesPipeline(TokenClassificationPipeline):
  def _forward(self,model_input):
    import torch
    v=model_input["input_ids"][0].tolist()
    with torch.no_grad():
      e=self.model(input_ids=torch.tensor([v[0:i]+[self.tokenizer.mask_token_id]+v[i+1:]+[j] for i,j in enumerate(v[1:-1],1)]))
    return {"logits":e.logits[:,1:-2,:],**model_input}
  def postprocess(self,model_output,**kwargs):
    import numpy
    import ufal.chu_liu_edmonds
    e=model_output["logits"].numpy()
    r=[1 if i==0 else -1 if j.endswith("|root") else 0 for i,j in sorted(self.model.config.id2label.items())]
    e+=numpy.where(numpy.add.outer(numpy.identity(e.shape[0]),r)==0,0,numpy.nan)
    m=numpy.full((e.shape[0]+1,e.shape[1]+1),numpy.nan)
    m[1:,1:]=numpy.nanmax(e,axis=2).transpose()
    p=numpy.zeros(m.shape)
    p[1:,1:]=numpy.nanargmax(e,axis=2).transpose()
    for i in range(1,m.shape[0]):
      m[i,0],m[i,i],p[i,0]=m[i,i],numpy.nan,p[i,i]
    h=ufal.chu_liu_edmonds.chu_liu_edmonds(m)[0]
    if [0 for i in h if i==0]!=[0]:
      m[:,0]+=numpy.where(m[:,0]<numpy.nanmax(m[:,0]),numpy.nan,0)
      h=ufal.chu_liu_edmonds.chu_liu_edmonds(m)[0]
    v=[(s,e) for s,e in model_output["offset_mapping"][0].tolist() if s<e]
    q=[self.model.config.id2label[p[i,j]].split("|") for i,j in enumerate(h)]
    if "aggregation_strategy" in kwargs and kwargs["aggregation_strategy"]!="none":
      for i,j in reversed(list(enumerate(q[2:],2))):
        if j[-1]=="goeswith":
          h=[b if i>b else b-1 for a,b in enumerate(h) if i!=a]
          v[i-2]=(v[i-2][0],v.pop(i-1)[1])
          q.pop(i)
    t=model_output["sentence"]
    u="# text = "+t+"\n"
    for i,(s,e) in enumerate(v,1):
      u+="\t".join([str(i),t[s:e],"_",q[i][0],"_","|".join(q[i][1:-1]),str(h[i]),q[i][-1],"_","_" if i<len(v) and e<v[i][0] else "SpaceAfter=No"])+"\n"
    return u+"\n"