def test_topk_sanity(self, dtype, axis, sorted):
x = vector(name="x", dtype=dtype)
fn = aesara.function([x],
topk(x, 1, axis=axis, sorted=sorted),
mode=self.mode)
assert any(
isinstance(n.op, self.op_class)
for n in fn.maker.fgraph.apply_nodes)
xval = np.asarray([1]).astype(dtype)
yval = fn(xval)
assert yval == xval
assert yval.dtype == xval.dtype
def test_grad(self, shp, k_, sorted):
ndim = len(shp)
for axis in range(-ndim, ndim):
if isinstance(k_, str):
k = eval(k_.replace("n", str(shp[axis])))
else:
k = k_
if k == 0:
continue
# make input away from undefined gradient (where some inputs are equal)
xval = gen_unique_vector(reduce(int.__mul__, shp),
dtype=aesara.config.floatX).reshape(shp)
utt.verify_grad(lambda x: topk(x, k, axis=axis, sorted=sorted),
[xval],
eps=1e-2)
def test_topk_1d(self, size, k, dtype, sorted):
if isinstance(k, str):
k = eval(k.replace("n", str(size)))
x = vector(name="x", dtype=dtype)
y = topk(x, k, sorted=sorted)
fn = aesara.function([x], y, mode=self.mode)
assert any(
isinstance(n.op, self.op_class)
for n in fn.maker.fgraph.apply_nodes)
# assert local_useless_topk opt is done properly
assert 1 == len(fn.maker.fgraph.outputs[0].owner.outputs)
# generate a all-unique array
xval = gen_unique_vector(size, dtype)
yval = fn(xval)
idx = slice(-k, None) if k > 0 else slice(-k)
goal = np.sort(xval)[idx]
assert yval.dtype == goal.dtype
utt.assert_allclose(goal, np.sort(yval))