Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[RELAY][GRAD] Fix first-order AD on tuple arguments #6827

Merged
merged 1 commit into from
Nov 6, 2020
Merged

[RELAY][GRAD] Fix first-order AD on tuple arguments #6827

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
20 changes: 19 additions & 1 deletion src/relay/transforms/gradient.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -181,6 +181,22 @@ struct FirstOrderReverseAD : ExprFunctor<ADValue(const Expr&)> {
return ret;
}

Expr UpdateGrad(const Type& t, const Expr& arg, const Expr& grad, LetList* ll) {
if (t.as<TensorTypeNode>()) {
return ll->Push(Add(arg, grad));
} else if (auto* tt = t.as<TupleTypeNode>()) {
Array<Expr> updates;
for (size_t i = 0; i < tt->fields.size(); ++i) {
updates.push_back(this->UpdateGrad(tt->fields[i], ll->Push(GetField(arg, i)),
ll->Push(GetField(grad, i)), ll));
}
return ll->Push(Tuple(updates));
} else {
LOG(FATAL) << "unsupported arg type of operator: " << t;
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can we try and do diagnostics here? we could put into improve AD with diagnostics

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I agree on this but it will probably need some refactoring, we might as well do it for the whole pass (first-order and higher-order). I think a separate PR will be ideal.

throw;
}
}

ADValue VisitExpr_(const OpNode* op) final {
Op op_ref = GetRef<Op>(op);
ICHECK(rev_map.count(op_ref)) << op->name << " does not have reverse mode defined";
Expand All @@ -198,8 +214,10 @@ struct FirstOrderReverseAD : ExprFunctor<ADValue(const Expr&)> {
tvm::Array<Expr> rev = rev_map[op_ref](orig, ret->reverse);
ICHECK(args.size() == rev.size());
for (size_t i = 0; i < args.size(); ++i) {
auto ad_arg = args[i]->get<ADTensor>();
auto ad_arg_type = ad_arg.forward->checked_type();
args[i]->get<ADTensor>().reverse =
ll->Push(Add(args[i]->get<ADTensor>().reverse, rev[i]));
this->UpdateGrad(ad_arg_type, ad_arg.reverse, rev[i], ll);
}
});
return ret;
Expand Down
33 changes: 33 additions & 0 deletions tests/python/relay/test_pass_gradient.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -255,6 +255,29 @@ def _test_tuple(mode):
tvm.testing.assert_allclose(grad_z.asnumpy(), -1 * np.ones_like(grad_z.asnumpy()))


def _test_tuple_argument(mode):
shape = (2, 3)
dtype = "float32"
tensor_type = relay.TensorType(shape, dtype)
fields = 3
tuple_type = relay.TupleType([tensor_type] * fields)
tup = relay.var("tup", type_annotation=tuple_type)
body = relay.TupleGetItem(tup, 0)
for i in range(1, fields):
body = relay.add(body, relay.TupleGetItem(tup, i))
func = relay.Function([tup], body)
func = run_infer_type(func)
back_func = run_infer_type(gradient(func, mode=mode))
xs = [rand(dtype, *shape) for _ in range(fields)]
xs_np = np.array([x.asnumpy() for x in xs])
expected_forward = np.sum(xs_np, axis=0)
ex = create_executor()
forward, grad = ex.evaluate(back_func)(tuple(xs))
tvm.testing.assert_allclose(forward.asnumpy(), expected_forward)
for field in grad[0]:
tvm.testing.assert_allclose(field.asnumpy(), np.ones_like(field.asnumpy()))


def test_tuple():
_test_tuple("higher_order")

Expand All @@ -263,6 +286,16 @@ def test_tuple_first_order():
_test_tuple("first_order")


@pytest.mark.xfail(raises=tvm.error.TVMError)
def test_tuple_argument():
# fails until we add support for top-level tuple arguments in higher-order AD
_test_tuple_argument("higher_order")


def test_tuple_argument_first_order():
_test_tuple_argument("first_order")


def test_pow():
mod = tvm.IRModule()
p = Prelude(mod)
Expand Down