nn.py 文件源码

def l2normalize(layer, train_scale=True):
    W_param = layer.W
    s = W_param.get_value().shape
    if len(s)==4:
        axes_to_sum = (1,2,3)
        dimshuffle_args = [0,'x','x','x']
        k = s[0]
    else:
        axes_to_sum = 0
        dimshuffle_args = ['x',0]
        k = s[1]
    layer.W_scale = layer.add_param(lasagne.init.Constant(1.),
                          (k,), name="W_scale", trainable=train_scale, regularizable=False)
    layer.W = W_param * (layer.W_scale/T.sqrt(1e-6 + T.sum(T.square(W_param),axis=axes_to_sum))).dimshuffle(*dimshuffle_args)
    return layer

# fully connected layer with weight normalization