def test_stochasticoptimization():
# this optimization alternates between triggering and not triggering.
last_time_replaced = [False]
@local_optimizer([add])
def insert_broken_add_sometimes(fgraph, node):
if node.op == add:
last_time_replaced[0] = not last_time_replaced[0]
if last_time_replaced[0]:
return [off_by_half(*node.inputs)]
return False
edb = EquilibriumDB()
edb.register("insert_broken_add_sometimes", insert_broken_add_sometimes,
"all")
opt = edb.query("+all")
a = dvector()
b = dvector()
with pytest.raises(StochasticOrder):
aesara.function(
[a, b],
add(a, b),
mode=DebugMode(
optimizer=opt,
check_c_code=True,
stability_patience=max(2, config.DebugMode__patience),
),
)
def test_softmax_optimizations_w_bias2(self):
x = matrix("x")
b = vector("b")
c = vector("c")
one_of_n = lvector("one_of_n")
op = crossentropy_categorical_1hot
fgraph = FunctionGraph([x, b, c, one_of_n],
[op(softmax_legacy(add(x, b, c)), one_of_n)])
assert fgraph.outputs[0].owner.op == op
optdb.query(OPT_FAST_RUN).optimize(fgraph)
assert len(fgraph.toposort()) == 2
assert fgraph.outputs[
0].owner.op == crossentropy_softmax_argmax_1hot_with_bias
def test_logsoftmax_grad_true_div_elemwise(self):
# Checks that the gradient of an expression similar to a log(softmax)
# but with a different elemwise operation than true_div is not
# optimized.
x = matrix("x")
y = log(softmax(x))
g = grad(y.sum(), x)
softmax_grad_node = g.owner
assert softmax_grad_node.op == softmax_grad_legacy
true_div_node = softmax_grad_node.inputs[0].owner
assert true_div_node.op == true_div
# We replace the elemwise true_div op by an elemwise add.
new_g = softmax_grad_legacy(add(*true_div_node.inputs),
softmax_grad_node.inputs[1])
fgraph = FunctionGraph([x], [new_g])
optdb.query(OPT_FAST_RUN).optimize(fgraph)
assert softmax_grad_legacy in [n.op for n in fgraph.toposort()]