helper_dataset.py 文件源码

def lecun_lcn(input, kernel_size=9, threshold=1e-4, use_divisor=False):
    """
    Yann LeCun's local contrast normalization
    Orginal code in Theano by: Guillaume Desjardins

    :param input:
    :param kernel_size:
    :param threshold:
    :param use_divisor:
    :return:
    """
    input_shape = (input.shape[0], 1, input.shape[1], input.shape[2])
    input = input.reshape(input_shape).astype(floatX)

    X = T.tensor4(dtype=floatX)
    filter_shape = (1, 1, kernel_size, kernel_size)
    filters = gaussian_filter(kernel_size).reshape(filter_shape)
    filters = shared(_asarray(filters, dtype=floatX), borrow=True)

    convout = conv2d(input=X,
                     filters=filters,
                     input_shape=input.shape,
                     filter_shape=filter_shape,
                     border_mode='half')
    new_X = X - convout

    if use_divisor:
        # Scale down norm of kernel_size x kernel_size patch
        sum_sqr_XX = conv2d(input=T.sqr(T.abs_(new_X)),
                            filters=filters,
                            input_shape=input.shape,
                            filter_shape=filter_shape,
                            border_mode='half')

        denom = T.sqrt(sum_sqr_XX)
        per_img_mean = denom.mean(axis=[2, 3])
        divisor = T.largest(per_img_mean.dimshuffle(0, 1, 'x', 'x'), denom)
        divisor = T.maximum(divisor, threshold)
        new_X = new_X / divisor

    new_X = new_X.dimshuffle(0, 2, 3, 1)
    new_X = new_X.flatten(ndim=3)
    f = function([X], new_X)
    return f(input)