Fix select assign backward (#1739)

crates/burn-autodiff/src/ops/tensor.rs

@@ -1058,7 +1058,7 @@ impl<B: Backend, C: CheckpointStrategy> FloatTensorOps<Self> for Autodiff<B, C>
         }
 
         impl<B: Backend, const D: usize> Backward<B, D, 2> for IndexSelectDimAssign<D> {
-            type State = (usize, IntTensor<B, 1>, Shape<D>, Shape<D>, B::Device);
+            type State = (usize, IntTensor<B, 1>);
 
             fn backward(
                 self,
@@ -1066,21 +1066,14 @@ impl<B: Backend, C: CheckpointStrategy> FloatTensorOps<Self> for Autodiff<B, C>
                 grads: &mut Gradients,
                 _checkpointer: &mut Checkpointer,
             ) {
-                let (dim, indices, shape_lhs, shape_rhs, device) = ops.state;
-                let [indices_4lhs, indices_4rhs] = duplicate(&ops.parents, Some(indices));
+                let (dim, indices) = ops.state;
 
                 binary::<B, D, D, D, _, _>(
                     ops.parents,
                     ops.node,
                     grads,
-                    |grad| {
-                        let zeros = B::float_zeros(shape_lhs, &device);
-                        B::float_select_assign(grad, dim, indices_4lhs.unwrap(), zeros)
-                    },
-                    |grad| {
-                        let zeros = B::float_zeros(shape_rhs, &device);
-                        B::float_select_assign(zeros, dim, indices_4rhs.unwrap(), grad)
-                    },
+                    |grad| grad,
+                    |grad| B::float_select(grad, dim, indices),
                 );
             }
         }
@@ -1098,13 +1091,7 @@ impl<B: Backend, C: CheckpointStrategy> FloatTensorOps<Self> for Autodiff<B, C>
             .stateful()
         {
             OpsKind::Tracked(prep) => prep.finish(
-                (
-                    dim,
-                    indices.clone(),
-                    B::float_shape(&tensor.primitive),
-                    B::float_shape(&value.primitive),
-                    B::float_device(&value.primitive),
-                ),
+                (dim, indices.clone()),
                 B::float_select_assign(tensor.primitive, dim, indices, value.primitive),
             ),
             OpsKind::UnTracked(prep) => prep.finish(B::float_select_assign(

crates/burn-autodiff/src/tests/select.rs

@@ -54,4 +54,23 @@ mod tests {
             Data::from([[64., 64., 64.], [19., 19., 19.]])
         );
     }
+
+    #[test]
+    fn test_select_assign_grad_different_shapes() {
+        let device = Default::default();
+
+        let indices: Tensor<TestAutodiffBackend, 1, Int> = Tensor::from_ints([1], &device);
+        let x: Tensor<TestAutodiffBackend, 2> = Tensor::ones([1, 1], &device).require_grad();
+        let y = Tensor::ones([2, 1], &device).require_grad();
+
+        let w = y.clone().select_assign(0, indices, x.clone());
+        let w = w.matmul(y.clone().transpose());
+
+        let grads = w.backward();
+        let x_grad = x.grad(&grads).unwrap();
+        let y_grad = y.grad(&grads).unwrap();
+
+        assert_eq!(x_grad.into_data(), Data::from([[2.0]]));
+        assert_eq!(y_grad.into_data(), Data::from([[5.0], [5.0]]));
+    }
 }

crates/burn-tch/src/ops/base.rs

@@ -125,18 +125,12 @@ impl<E: tch::kind::Element + Copy + Default> TchOps<E> {
     pub fn select_assign<const D: usize>(
         tensor: TchTensor<E, D>,
         dim: usize,
-        indices_tensor: TchTensor<i64, 1>,
+        indices: TchTensor<i64, 1>,
         value: TchTensor<E, D>,
     ) -> TchTensor<E, D> {
-        let mut indices = Vec::with_capacity(D);
-        for _ in 0..D {
-            indices.push(None);
-        }
-        indices[dim] = Some(indices_tensor.tensor);
-
-        tensor.unary_ops(
-            |mut tensor| tensor.index_put_(&indices, &value.tensor, true),
-            |tensor| tensor.index_put(&indices, &value.tensor, true),
+        tensor.clone().unary_ops(
+            |mut tensor| tensor.index_add_(dim as i64, &indices.tensor, &value.tensor),
+            |tensor| tensor.index_add(dim as i64, &indices.tensor, &value.tensor),
         )
     }