dataprocessing.py 文件源码

def gaussian_blur(input, filter_size, filter_sampling_range=3.5, strides=[1, 1, 1, 1], padding='SAME'):
    """
    Blur input with a 2D Gaussian filter of size filter_size x filter_size. The filter's values are 
    sampled from an evenly spaced grid on the 2D standard normal distribution in the range 
    [-filter_sampling_range, filter_sampling_range] in both dimensions. 

    :param input: A rank-4 tensor with shape=(samples, x, y, n_channels). The same Gaussian filter 
        will be applied to all n_channels feature maps of input. 
    :param filter_size: The size of one edge of the square-shaped Gaussian filter. 
    :param filter_sampling_range: The range in which to sample from the standard normal distribution in 
        both dimensions, i.e. a sampling range of 1 corresponds to sampling in a square grid that bounds 
        the standard deviation circle.
    :param strides: Param strides as passed to tf.nn.depthwise_conv2d.
    :param padding: Param padding as passed to tf.nn.depthwise_conv2d.
    :return: The result of the Gaussian blur as a rank-4 tensor with the same shape as input.
    """

    # make 2D distribution
    mu = np.repeat(np.float32(0.), 2)
    sig = np.repeat(np.float32(1.), 2)
    dist = tf.contrib.distributions.MultivariateNormalDiag(mu, sig)

    # sample from distribution on a grid
    sampling_range = tf.cast(filter_sampling_range, tf.float32)
    x_1D = tf.linspace(-sampling_range, sampling_range, filter_size)
    x = tf.stack(tf.meshgrid(x_1D, x_1D), 2)
    kern = dist.pdf(x)
    kern /= tf.reduce_sum(kern)
    kern = tf.reshape(kern, kern.shape.as_list() + [1, 1])
    kern = tf.tile(kern, [1, 1, input.shape.as_list()[-1], 1])

    return tf.nn.depthwise_conv2d(input, kern, strides, padding)