[DL]基于pytorch的Elman RNN语言模型

留个坑，先放代码，后面再解释

#qa_pairs_parallel.txtimport torchimport torch.nn as nnimport jiebaclass Sentences(object):    def __init__(self,dirname):        self.dirname=dirname    def __iter__(self):        for line in open(self.dirname,'r').readlines():            yield line.replace('\n','')import jiebaimport copysentences=Sentences('./corpus_data/qa_pairs_parallel.txt')x=[]y=[]for line in sentences:    arrs=line.split('$$')    for arr in arrs[:2]:        temp=list(arr)        tempx=copy.copy(temp);tempy=copy.copy(temp)        tempx.insert(0,'<start>');x.append(tempx)        tempy.append('<end>');y.append(tempy)print(len(x),x[:5])print(len(y),y[:5])dictionary=dict()for lists in [x,y]:    for words in lists:        for word in words:            if word not in dictionary:                dictionary[word]=len(dictionary)print(len(dictionary))id2word=dict()for word in dictionary:    id2word[dictionary[word]]=wordprint(dictionary['<end>'])print(dictionary['<start>'])print(id2word[1])print(id2word[222])class RNN(nn.Module):    def __init__(self,vocab_size,input_size,hidden_size):        super(RNN,self).__init__()        self.input_size=input_size        self.hidden_size=hidden_size        self.vocab_size=vocab_size        self.sigmoid=nn.Sigmoid()        self.softmax=nn.Softmax()        self.linear1=nn.Linear(self.vocab_size+self.hidden_size,self.hidden_size)        self.linear2=nn.Linear(self.hidden_size,self.vocab_size)    def forward(self,input_x,input_hidden):        x_t=torch.cat((input_x.view(1,-1),input_hidden.view(1,-1)),dim=1)        hidden=self.sigmoid(self.linear1(x_t))        output=self.softmax(self.linear2(hidden))        return [output,hidden]hidden_size=256vocab_size=len(dictionary)input_size=1000model=RNN(vocab_size,input_size,256)if torch.cuda.is_available():    model.cuda()loss_fun = nn.CrossEntropyLoss()optimizer=torch.optim.SGD(model.parameters(),lr=0.1)embedding = torch.nn.Embedding(vocab_size, vocab_size, padding_idx=0)one2n=torch.eye(vocab_size)for _ in range(5):    for i,xi in enumerate(x[:]):        input_hidden=torch.autograd.Variable(torch.randn(1,hidden_size))        loss=0.0        model.zero_grad()        for j,word in enumerate(xi):            input=torch.autograd.Variable(one2n[dictionary[word]],requires_grad=True).view(1,-1)            target=torch.autograd.Variable(torch.cuda.LongTensor([dictionary[y[i][j]]]))            out,input_hidden=model(input.cuda(),input_hidden.cuda())            loss+=loss_fun(out,target)        print('-'*89)        loss/=len(xi)        print('loss=',loss.data)        loss.backward(retain_graph=True)        optimizer.step()torch.save(model,'rnn_language_model.m')