def gaussian_filter_2d_variable_sigma(input, sigmas,
window_radius=None,
border_mode='zero'
):
warnings.warn("gaussian_filter_2d_variable_sigma takes forever to compile.")
def filter_sigma(idx, kernel):
dimpattern_w = ('x', 'x', 'x', 0)
dimpattern_h = ('x', 'x', 0, 'x')
filter_w = kernel.dimshuffle(dimpattern_w)
blur_w = T.nnet.conv2d(
padded_input[idx:idx+1], filter_w,
border_mode=_get_chained_w_h_conv_border(conv_border, 'w'),
filter_shape=[1, 1, 1, None])
filter_h = kernel.dimshuffle(dimpattern_h)
return T.nnet.conv2d(
blur_w, filter_h,
border_mode=_get_chained_w_h_conv_border(conv_border, 'h'),
filter_shape=[1, 1, None, 1])
ndim = 4
assert input.ndim == ndim, \
"there must be {} dimensions, got {}".format(ndim, input.ndim)
window_radius = gaussian_kernel_default_radius(sigmas, window_radius)
padded_input, conv_border = add_border(input, window_radius, border_mode)
kernel = gaussian_kernel_1d(sigmas, window_radius)
blur, _ = theano.map(
filter_sigma,
sequences=[T.arange(sigmas.shape[0]), kernel])
return blur.reshape(input.shape)
def gaussian_filter_2d(input, sigma, window_radius=None,
border_mode='zero'):
bs, ch, h, w = input.shape
ndim = 4
assert input.ndim == ndim, \
"there must be {} dimensions, got {}".format(ndim, input.ndim)
window_radius = gaussian_kernel_default_radius(sigma, window_radius)
input = input.reshape((bs*ch, 1, h, w))
filter_1d = gaussian_kernel_1d(sigma, window_radius)
dimpattern_w = ('x', 'x', 0, 1)
dimpattern_h = ('x', 'x', 1, 0)
filter_w = filter_1d.dimshuffle(dimpattern_w)
padded_input, conv_border = add_border(input, window_radius, border_mode)
blur_w = T.nnet.conv2d(
padded_input, filter_w,
border_mode=_get_chained_w_h_conv_border(conv_border, 'w'),
filter_shape=[1, 1, 1, None])
filter_h = filter_1d.dimshuffle(dimpattern_h)
blured = T.nnet.conv2d(
blur_w, filter_h,
border_mode=_get_chained_w_h_conv_border(conv_border, 'h'),
filter_shape=[1, 1, None, 1])
return blured.reshape((bs, ch, h, w))
def sobel(img, border_mode='zero'):
filter = np.array([
[1, 0, -1],
[2, 0, -2],
[1, 0, -1],
], dtype=floatx())
img, conv_border = add_border(img, border=1, mode=border_mode)
kernel_x = theano.shared(filter[np.newaxis, np.newaxis])
kernel_y = theano.shared(np.transpose(filter)[np.newaxis, np.newaxis])
conv_x = T.nnet.conv2d(img, kernel_x, border_mode=conv_border,
input_shape=(None, 1, None, None),
filter_shape=(1, 1, 3, 3))
conv_y = T.nnet.conv2d(img, kernel_y, border_mode=conv_border,
input_shape=(None, 1, None, None),
filter_shape=(1, 1, 3, 3))
return conv_x, conv_y
