doc2veckeras.py 文件源码

def build_keras_model_dm(index_size,vector_size,vocab_size,
                         context_size,
                         maxwords,
                         cbow_mean=False,
                         learn_doctags=True, learn_words=True, learn_hidden=True,
                         model=None ,
                         word_vectors=None,doctag_vectors=None,hidden_vectors=None,
                         sub_batch_size=1
                         ):
    """
    >>> word_vectors=np.array([[1,2],[3,4],[5,6]])
    >>> doctag_vectors=np.array([[10,20],[30,40]])
    >>> hidden_vectors=np.array([[1,0],[0,1]])
    >>> sub_batch_size=2
    >>> kerasmodel=build_keras_model_dm(index_size=2,vector_size=2,vocab_size=3,context_size=2,maxwords=2,sub_batch_size=sub_batch_size,word_vectors=word_vectors,doctag_vectors=doctag_vectors,hidden_vectors=hidden_vectors, learn_words=True )
    >>> ind=[[0],[1]]
    >>> iwd=[[1,0],[1,1]]
    >>> ipt=[[1,0],[0,1]]
    >>> tmp1=kerasmodel.predict({'index':np.array(ind),'iword':np.array(iwd),'point':np.array(ipt)})['code']
    >>> tmp2=np.array([ [(word_vectors[iwd[i]].sum(0)+doctag_vectors[i]).dot(hidden_vectors[j]) for j in ipt[i] ] for i in range(2)])
    >>> np.linalg.norm(1/(1+np.exp(-tmp2))-tmp1) < 0.001
    True
    """ 
    kerasmodel = Graph()

    kerasmodel.add_input(name='index',input_shape=(1,)     , dtype=int)
    if doctag_vectors is None :        
        kerasmodel.add_node(Embedding(index_size,   vector_size,trainable=learn_doctags,input_length=1                              ),name='embedindex', input='index')
    else:
        kerasmodel.add_node(Embedding(index_size,   vector_size,trainable=learn_doctags,input_length=1     ,weights=[doctag_vectors]),name='embedindex', input='index')
    kerasmodel.add_input(name='iword',input_shape=(maxwords,), dtype=int)

    if word_vectors is None :
        kerasmodel.add_node(Embedding(vocab_size,   vector_size,trainable=learn_words  ,input_length=maxwords                         ),name='embedword', input='iword')
    else:
        kerasmodel.add_node(Embedding(vocab_size,   vector_size,trainable=learn_words  ,input_length=maxwords,weights=[word_vectors  ]),name='embedword', input='iword')

    kerasmodel.add_input(name='point',input_shape=(sub_batch_size,)     , dtype=int)
    if hidden_vectors is None :
        kerasmodel.add_node(Embedding(context_size, vector_size,trainable=learn_hidden ,input_length=sub_batch_size                          ),name='embedpoint', input='point')
    else:
        kerasmodel.add_node(Embedding(context_size, vector_size,trainable=learn_hidden ,input_length=sub_batch_size  ,weights=[hidden_vectors]),name='embedpoint', input='point')

    if cbow_mean:
        kerasmodel.add_node(Lambda(lambda x:x.mean(1),output_shape=(vector_size,)), name='merge',inputs=['embedindex','embedword'], merge_mode='concat', concat_axis=1)
    else:
        kerasmodel.add_node(Lambda(lambda x:x.sum(1),output_shape=(vector_size,)), name='merge',inputs=['embedindex','embedword'], merge_mode='concat', concat_axis=1)

    kerasmodel.add_node(Activation('sigmoid'), name='sigmoid',inputs=['merge','embedpoint'], merge_mode='dot',dot_axes=-1)
    kerasmodel.add_output(name='code',input='sigmoid')
    kerasmodel.compile('rmsprop', {'code':'mse'})

    return kerasmodel