mlp_layer.py 文件源码

def mlp_layer_softmax(tparams, layer1_input, prefix='mlp_layer'):

    """ layer1_input:  n_sample * n_feature    64*20
        input_shape: (num of hiddens, number of input features)   200*20
        pred_shape: (num of labels, number of hiddens) 2*200
        y_recon : n_label *n_sample 2*64
    """
    hidden_2_out = tensor.nnet.sigmoid(tensor.dot(layer1_input, tparams[_p(prefix,'W1')].T) + tparams[_p(prefix,'b1')] )  # 64*200  
    y_recons = tensor.dot(hidden_2_out, tparams[_p(prefix,'V1')].T) + tparams[_p(prefix,'c1')]  
    #y_recons = tensor.tanh(y_recons) * 10   # avoid numerical issues/label smoothing
    #y_recons = tensor.nnet.softmax(y_recons) # 64*2



    max_w = tensor.max(y_recons, axis = 1, keepdims=True)
    e0 = tensor.exp(y_recons - max_w)
    y_recons = e0 / tensor.sum(e0, axis = 1, keepdims=True)

    return y_recons