def grad(self, inputs, output_grads):
a, axis = inputs
inp_grad = theano.gradient.grad_not_implemented(
self, 0, axis,
"Currently, we only implement the gradient on sort for vector"
" and matrix (and axis is None)")
if a.ndim == 1:
idx = argsort(*inputs, kind=self.kind, order=self.order)
# rev_idx = numpy.where(idx[None, :]==numpy.arange(5)[:,None])[1]
rev_idx = theano.tensor.eq(idx[None, :],
arange(a.shape[0])[:, None]).nonzero()[1]
inp_grad = output_grads[0][rev_idx]
elif a.ndim == 2:
if (axis is None or
(isinstance(axis, theano.Constant) and axis.data is None)):
idx = argsort(*inputs, kind=self.kind, order=self.order)
rev_idx = theano.tensor.eq(idx[None, :],
arange(a.shape[0]*a.shape[1])[:, None]).nonzero()[1]
inp_grad = output_grads[0][rev_idx].reshape(a.shape)
elif (axis == 0 or
(isinstance(axis, theano.Constant) and axis.data == 0)):
idx = argsort(*inputs, kind=self.kind, order=self.order)
# not working: numpy.where(idx[None, :]==numpy.arange(2)[:, None, None])
pass
axis_grad = theano.gradient.grad_undefined(
self, 1, axis,
"sort is not defined for non-integer axes so"
" sort(x, axis+eps) is undefined")
return [inp_grad, axis_grad]
def __get_expanded_dim(self, a, axis, i):
index_shape = [1] * a.ndim
index_shape[i] = a.shape[i]
# it's a way to emulate
# numpy.ogrid[0: a.shape[0], 0: a.shape[1], 0: a.shape[2]]
index_val = arange(a.shape[i]).reshape(index_shape)
return index_val
def __get_expanded_dim(self, a, axis, i):
index_shape = [1] * a.ndim
index_shape[i] = a.shape[i]
# it's a way to emulate
# numpy.ogrid[0: a.shape[0], 0: a.shape[1], 0: a.shape[2]]
index_val = arange(a.shape[i]).reshape(index_shape)
return index_val
def L_op(self, inputs, outputs, out_grads):
x, k = inputs
k_grad = grad_undefined(self, 1, k, "topk: k is not differentiable")
if not (self.return_indices or self.return_values):
x_grad = grad_undefined(
self,
0,
x,
"topk: cannot get gradient"
" without both indices and values",
)
else:
x_shp = theano.tensor.shape(x)
z_grad = out_grads[0]
ndim = x.ndim
axis = self.axis % ndim
grad_indices = [
arange(x_shp[i]).dimshuffle([0] + ["x"] * (ndim - i - 1))
if i != axis else outputs[-1] for i in range(ndim)
]
x_grad = x.zeros_like(dtype=z_grad.dtype)
x_grad = set_subtensor(x_grad[tuple(grad_indices)], z_grad)
return [x_grad, k_grad]
def L_op(self, inputs, outputs, out_grads):
x, k = inputs
k_grad = grad_undefined(self, 1, k, 'topk: k is not differentiable')
if not (self.return_indices or self.return_values):
x_grad = grad_undefined(
self, 0, x, 'topk: cannot get gradient'
' without both indices and values')
else:
x_shp = theano.tensor.shape(x)
z_grad = out_grads[0]
ndim = x.ndim
axis = self.axis % ndim
grad_indices = [
arange(x_shp[i]).dimshuffle([0] + ['x'] * (ndim - i - 1))
if i != axis else outputs[-1] for i in range(ndim)]
x_grad = x.zeros_like(dtype=z_grad.dtype)
x_grad = set_subtensor(x_grad[tuple(grad_indices)], z_grad)
return [x_grad, k_grad]