ud.py

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.startswith("B-") or not (k.startswith("I-") or k.endswith("|root") or k.find("|l-")>0 or k.find("|r-")>0)]
    self.transition=numpy.full((len(x),len(x)),-numpy.inf)
    for k,v in x.items():
      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
  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()
    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(model_outputs["offset_mapping"][0].tolist(),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 p.startswith("I-"):
          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 UniversalDependenciesCausalPipeline(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
  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)
    d=[t["text"] for t in w]
    v=self.tokenizer(d,add_special_tokens=False)
    e=self.model.get_input_embeddings().weight
    m=[]
    for x in v["input_ids"]:
      if x==[]:
        x=[self.tokenizer.unk_token_id]
      m.append(e[x,:].sum(axis=0))
    m.append(e[self.tokenizer.sep_token_id,:])
    m.append(e[self.tokenizer.pad_token_id,:])
    m=torch.stack(m).to(self.device)
    k=list(range(len(d)+1))
    e=[]
    with torch.no_grad():
      for i in range(len(d)):
        e.append(self.model(inputs_embeds=torch.unsqueeze(m[k+list(range(i,len(d)))+[-1]*i,:],0)).logits[0,-len(d):,:])
    e=torch.stack(e).cpu().numpy()
    for i in range(len(d)):
      for j in range(i):
        e[-j-1,-i-1],e[-i-1,-j-1]=e[-i-1,i-j]+self.left_arc,e[-i-1,i-j]+self.right_arc
      e[-i-1,-i-1]=e[-i-1,0]+self.root
    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,j in enumerate(d):
      u+="\t".join([str(i+1),j,"_",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(d) and w[i]["end"]<w[i+1]["start"] 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