decoder.py 文件源码

def getTrainedCLM():
    ''' Read CLM from file '''
    #Some parameters for the CLM
    INPUT_SIZE = 29

    #Hidden layer hyper-parameters
    N_HIDDEN = 100
    HIDDEN_NONLINEARITY = 'rectify'

    #Gradient clipping
    GRAD_CLIP = 100
    l_in = lasagne.layers.InputLayer(shape = (None, None, INPUT_SIZE)) #One-hot represenntation of character indices
    l_mask = lasagne.layers.InputLayer(shape = (None, None))

    l_recurrent = lasagne.layers.RecurrentLayer(incoming = l_in, num_units=N_HIDDEN, mask_input = l_mask, learn_init=True, grad_clipping=GRAD_CLIP)
    Recurrent_output=lasagne.layers.get_output(l_recurrent)

    n_batch, n_time_steps, n_features = l_in.input_var.shape

    l_reshape = lasagne.layers.ReshapeLayer(l_recurrent, (-1, N_HIDDEN))
    Reshape_output = lasagne.layers.get_output(l_reshape)

    l_h1 = lasagne.layers.DenseLayer(l_reshape, num_units=N_HIDDEN)
    l_h2 = lasagne.layers.DenseLayer(l_h1, num_units=N_HIDDEN)
    l_dense = lasagne.layers.DenseLayer(l_h2, num_units=INPUT_SIZE, nonlinearity = lasagne.nonlinearities.softmax)
    with np.load('CLM_model.npz') as f:
        param_values = [f['arr_%d' % i] for i in range(len(f.files))]
    lasagne.layers.set_all_param_values(l_dense, param_values,trainable = True)
    output = lasagne.layers.get_output( l_dense )
    return l_in,l_mask,output


#def getCLMOneHot( sequence ):