Merge pull request #19760 from Blair-Johnson:fix-pytree-grads-sparse

PiperOrigin-RevId: 657258194

jax/experimental/sparse/ad.py

@@ -67,8 +67,12 @@ def flatten_fun_for_sparse_ad(fun, argnums: int | tuple[int, ...], args: tuple[A
     return f_recons(grad_out)
 
   def postprocess_gradients(grads_out):
-    out = [reconstruct(*args) for args in safe_zip(argnums_flat1, grads_out)]
-    return out[0] if isinstance(argnums, int) else out
+    leaf_grads = [None] * tree1.num_leaves
+    for i, grad in safe_zip(argnums_flat1, grads_out):
+      leaf_grads[i] = reconstruct(i, grad)
+    grad_tree = tree_util.tree_unflatten(tree1, leaf_grads)
+    grad_tree = tuple(filter(lambda x: jax.tree.leaves(x), grad_tree))
+    return grad_tree[0] if len(grad_tree) == 1 else grad_tree
 
   return fun_flat, argnums_flat, args_flat, postprocess_gradients
 

tests/sparse_test.py

@@ -712,6 +712,97 @@ class SparseGradTest(sptu.SparseTestCase):
       self.assertAllClose(jac_dense(f, argnums=1, has_aux=has_aux)(X, y),
                           jac_sparse(f, argnums=1, has_aux=has_aux)(Xsp, y), rtol=rtol)
 
+  @jtu.sample_product(has_aux=[True, False],
+                      deep=[True,False],
+                      arg0=[True,False],
+                      bias=[True,False])
+  def test_sparse_pytree_grad(self, has_aux, deep, arg0, bias):
+    rng_sparse = sptu.rand_sparse(self.rng())
+    rng = jtu.rand_default(self.rng())
+
+    y = rng(5, "float32")
+    X = rng_sparse((10, 5), "float32")
+    b = rng(10, "float32")
+    Xsp = sparse.BCOO.fromdense(X)
+    Xtree_sp = {'deep':{'X':Xsp},
+                'X':Xsp,
+                'list':[None,(b,None)]}
+    Xtree_de = {'deep':{'X':X},
+                'X':X,
+                'list':[None,(b,None)]}
+
+    def f(Xtree, y):
+      if deep:
+        out = Xtree['deep']['X'] @ y
+      else:
+        out = Xtree['X'] @ y
+      # Other grad variables
+      if bias:
+        out += Xtree['list'][1][0]
+      out = jnp.sum(out)
+      if has_aux:
+        return out, {'y': y.shape}
+      else:
+        return out
+
+    def g(y, Xtree):
+      if deep:
+        out = Xtree['deep']['X'] @ y
+      else:
+        out = Xtree['X'] @ y
+      # Other grad variables
+      if bias:
+        out += Xtree['list'][1][0]
+      out = jnp.sum(out)
+      if has_aux:
+        return out, {'y': y.shape}
+      return out
+
+    with self.subTest("wrt sparse"):
+      # Argument ordering
+      if arg0:
+        grad_de = jax.grad(f, argnums=0, has_aux=has_aux)(Xtree_de, y)
+        grad_sp = sparse.grad(f, argnums=0, has_aux=has_aux)(Xtree_sp, y)
+      else:
+        grad_de = jax.grad(g, argnums=1, has_aux=has_aux)(y, Xtree_de)
+        grad_sp = sparse.grad(g, argnums=1, has_aux=has_aux)(y, Xtree_sp)
+
+      if has_aux:
+        grad_de, aux_de = grad_de
+        grad_sp, aux_sp = grad_sp
+        self.assertAllClose(aux_de, aux_sp)
+
+      # Pytree structure
+      is_bcoo = lambda x: isinstance(x, sparse.bcoo.BCOO)
+      grad_densified = jax.tree_util.tree_map(sparse.todense, grad_sp,
+                                              is_leaf=is_bcoo)
+      self.assertEqual(jax.tree_util.tree_structure(grad_de),
+                       jax.tree_util.tree_structure(grad_densified))
+
+      # Depth in tree
+      if deep:
+        grad_sp_arr = grad_sp['deep']['X']
+        grad_de_arr = grad_de['deep']['X']
+      else:
+        grad_sp_arr = grad_sp['X']
+        grad_de_arr = grad_de['X']
+      self.assertIsInstance(grad_sp_arr, sparse.BCOO)
+      self.assertAllClose(grad_sp_arr.data,
+                          sparse_bcoo._bcoo_extract(grad_sp_arr.indices,
+                                                    grad_de_arr))
+      # Other grad variables
+      if bias:
+        self.assertAllClose(grad_sp['list'][1][0],
+                            grad_de['list'][1][0])
+
+    with self.subTest("wrt dense"):
+      # Argument ordering
+      if arg0:
+        self.assertAllClose(jax.grad(f, argnums=1, has_aux=has_aux)(Xtree_de, y),
+                            sparse.grad(f, argnums=1, has_aux=has_aux)(Xtree_sp, y))
+      else:
+        self.assertAllClose(jax.grad(g, argnums=0, has_aux=has_aux)(y, Xtree_de),
+                            sparse.grad(g, argnums=0, has_aux=has_aux)(y, Xtree_sp))
 
 class SparseObjectTest(sptu.SparseTestCase):
   @parameterized.named_parameters(