zhangyiss/mlx
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx.git synced 2025-06-24 17:31:16 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

scatter_max vjp + bindings + tests (#431)

Browse Source 
Co-authored-by: DjamelMesbah <djamel.mesbah@adservio.fr>
This commit is contained in:
Tristan Bilot 2024-01-14 23:12:15 +01:00 committed by GitHub
parent 4bc446be08
commit 6022d4129e
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 53 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

30
mlx/primitives.cpp
View File

@ -2129,6 +2129,7 @@ std::vector<array> Scatter::vjp(
switch (reduce_type_) {
case Scatter::None:
case Scatter::Sum:
case Scatter::Max:
break;
default:
throw std::runtime_error(
@ -2139,6 +2140,17 @@ std::vector<array> Scatter::vjp(
const array& updates = primals.back();
const std::vector<array> indices(primals.begin() + 1, primals.end() - 1);
// Store result of scatter if needed for reuse in vjp
auto get_result = [&]() {
switch (reduce_type_) {
case Scatter::Max:
return scatter_max(values, indices, updates, axes_, stream());
default:
return array({});
}
};
array result = get_result();
std::vector<array> vjps;
for (auto num : argnums) {
// Gradient wrt to the input array
@ -2157,6 +2169,11 @@ std::vector<array> Scatter::vjp(
// The input array values are kept so they all get gradients
vjps.push_back(cotangents[0]);
break;
case Scatter::Max: {
auto mask = where(result == values, array({1}), array({0}));
vjps.push_back(multiply(cotangents[0], mask));
break;
}
default:
// Should never reach here
throw std::invalid_argument("");
@ -2174,6 +2191,19 @@ std::vector<array> Scatter::vjp(
gather(cotangents[0], indices, axes_, slice_sizes, stream()));
break;
}
case Scatter::Max: {
auto slice_sizes = cotangents[0].shape();
for (auto ax : axes_) {
slice_sizes[ax] = 1;
}
auto gathered_cotan =
gather(cotangents[0], indices, axes_, slice_sizes, stream());
auto gathered_result =
gather(result, indices, axes_, slice_sizes, stream());
vjps.push_back(
multiply(gathered_cotan, gathered_result == updates, stream()));
break;
}
default: {
// Should never reach here
throw std::invalid_argument("");

23
python/tests/test_autograd.py
View File

@ -293,6 +293,29 @@ class TestAutograd(mlx_tests.MLXTestCase):
self.assertTrue(mx.array_equal(dfdx, mx.array([1.0])))
self.assertEqual(dfdx.dtype, mx.float32)
def test_scatter_max_vjp(self):
def fun(src, updates):
x = src.at[1].maximum(updates)
return x
cotan = mx.array([4.0, 5.0, 6.0])
_, vjps = mx.vjp(fun, [mx.array([1.0, 2.0, 3.0]), mx.array([[3.0]])], [cotan])
mx.eval(vjps)
# Update larger than value
self.assertTrue(mx.allclose(vjps[0], mx.array([4.0, 0.0, 6.0])))
self.assertTrue(mx.allclose(vjps[1], mx.array([5.0])))
cotan = mx.array([[4.0], [5.0], [6.0]])
_, vjps = mx.vjp(
fun, [mx.array([[1.0], [2.0], [3.0]]), mx.array([[[2.0]]])], [cotan]
)
mx.eval(vjps)
# Update and value are equal
self.assertTrue(mx.allclose(vjps[0], mx.array([[4.0], [5.0], [6.0]])))
self.assertTrue(mx.allclose(vjps[1], mx.array([[[5.0]]])))
def test_vjp_types(self):
def fun(x):
return x