def numgrad(func: Callable, a: np.ndarray, delta: float = 1e-6) -> np.ndarray:
"Numerical gradient of func(a) at `a`."
grad = np.zeros(a.shape, a.dtype)
for index, _ in np.ndenumerate(grad):
delta_array = np.zeros(a.shape, a.dtype)
delta_array[index] = delta / 2
grad[index] = np.sum(
(func(a + delta_array) - func(a - delta_array)) / delta)
return grad
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 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
def multiply_by_locgrad_b(path_value):
return np.where(
condition,
np.zeros(path_value.shape, a.array.dtype),
path_value,
)
def multiply_by_locgrad(path_value):
result = np.zeros(a.array.shape, a.array.dtype)
if axis: # Expand dims so they can be broadcast.
path_value = np.expand_dims(path_value, axis)
return result + path_value
def multiply_by_locgrad(path_value): # TODO: a faster method
result = np.zeros(a.shape, a.dtype)
np_add_at(np_strided_sliding_view(result, window_shape, strides), None,
path_value)
return result
def multiply_by_locgrad(path_value):
"(Takes into account elements indexed multiple times.)"
result = np.zeros(a.array.shape, a.array.dtype)
np_add_at(result, indices, path_value)
return result
def multiply_by_locgrad_b(path_value):
path_value = np.sum(path_value,
axis=b_repeatdims).reshape(b.array.shape)
return np.zeros(b.array.shape, b.array.dtype) + path_value