layers.py 文件源码

def localResponseNormalizationCrossChannel(incoming, alpha=1e-4,
                                           k=2, beta=0.75, n=5):
    """
    Implement the local response normalization cross the channels described
    in <ImageNet Classification with Deep Convolutional Neural Networks>,
    A.Krizhevsky et al. sec.3.3.
    Reference of the code:
    https://github.com/Lasagne/Lasagne/blob/master/lasagne/layers/
    normalization.py
    https://github.com/lisa-lab/pylearn2/blob/master/pylearn2/expr/normalize.py
    Parameters:
    incomping: The feature maps. (output of the convolution layer).
    alpha: float scalar
    k: float scalr
    beta: float scalar
    n: integer: number of adjacent channels. Must be odd.
    """
    if n % 2 == 0:
        raise NotImplementedError("Works only with odd n")

    input_shape = incoming.shape
    half_n = n // 2
    input_sqr = T.sqr(incoming)
    b, ch, r, c = input_shape
    extra_channels = T.alloc(0., b, ch + 2*half_n, r, c)
    input_sqr = T.set_subtensor(extra_channels[:, half_n:half_n+ch, :, :],
                                input_sqr)
    scale = k
    for i in range(n):
        scale += alpha * input_sqr[:, i:i+ch, :, :]
    scale = scale ** beta

    return incoming / scale