def multiply_by_locgrad(path_value):
result = np.zeros(flatshape)
result[np.arange(result.shape[0]), idx] = 1
swapped_shape = list(a.shape)
swapped_shape[axis], swapped_shape[-1] = swapped_shape[
-1], swapped_shape[axis]
result = result.reshape(swapped_shape)
result = np.swapaxes(result, axis, -1)
return path_value * result
def cross_entropy(y_pred: Variable, y_true: np.array, axis=-1) -> Variable:
"""Cross entropy.
Args:
y_pred: A sp.Variable instance of shape [batch_size, n_classes]
y_true: A NumPy array, of shape [batch_size], containing the true class labels.
Returns:
A scalar, reduced by summation.
"""
indices = (np.arange(len(y_true)), y_true)
return neg(sum(log(getitem(y_pred, indices))))
def patches_index(imheight, imwidth, kernheight, kernwidth, stride_y,
stride_x):
"Index to get image patches, e.g. for 2d convolution."
max_y_idx = imheight - kernheight + 1
max_x_idx = imwidth - kernwidth + 1
row_major_index = np.arange(imheight * imwidth).reshape(
[imheight, imwidth])
patch_corners = row_major_index[0:max_y_idx:stride_y, 0:max_x_idx:stride_x]
elements_relative = row_major_index[0:kernheight, 0:kernwidth]
index_of_patches = patch_corners.reshape(
[-1, 1]) + elements_relative.reshape([1, -1])
index_of_patches = np.unravel_index(index_of_patches,
shape=[imheight, imwidth])
outheight, outwidth = patch_corners.shape
n_patches = outheight * outwidth
return index_of_patches, outheight, outwidth, n_patches
def onehot(y: np.ndarray, n_classes: int) -> np.array:
"Onehot encode vector y with classes 0 to n_classes-1."
result = np.zeros([len(y), n_classes])
result[np.arange(len(y)), y] = 1
return result