[feat] relax backward function to handle indexing as last operation (#10)

Author: zitongzhan

bae/autograd/graph.py

@@ -105,7 +105,18 @@ def backward(output_):
         index = output_.optrace[id(output_)][1]
         arg = output_.optrace[id(output_)][2]
 
-        # check if upstream index exists
+        # If the last operation is indexing, there is no downstream map op to
+        # populate Jacobian values. In this case, the Jacobian block values are
+        # identity matrices placed at the indexed columns.
+        if not hasattr(output_, 'jactrace'):
+            if output_.ndim == 1:
+                eye_blocks = torch.ones((output_.shape[0], 1, 1), device=output_.device, dtype=output_.dtype)
+            else:
+                block_dim = output_.shape[-1]
+                eye = torch.eye(block_dim, device=output_.device, dtype=output_.dtype)
+                eye_blocks = eye.unsqueeze(0).repeat(output_.shape[0], 1, 1)
+            output_.jactrace = (None, eye_blocks)
+
         if type(output_.jactrace) is tuple:
             if output_.jactrace[0] is not None:
                 upstream_index = output_.jactrace[0]
@@ -124,7 +135,7 @@ def backward(output_):
 
 
 def jacobian(output, params):
-    assert output.optrace[id(output)][0] == 'map', "The last operation in compute graph being indexing transform is not meaningful"
+    assert output.optrace[id(output)][0] in ('map', 'index'), "Unsupported last operation in compute graph"
     backward(output)
     res = []
     for param in params:

tests/autograd/test_graph_jacobian.py

@@ -71,3 +71,32 @@ def f(A: torch.Tensor, B: torch.Tensor) -> torch.Tensor:
 
     assert torch.equal(J_sparse[0].col_indices(), idx_a[sel])
     assert torch.equal(J_sparse[1].col_indices(), idx_b[sel])
+
+
+@pytest.mark.parametrize("device", ["cpu", "cuda"])
+def test_sparse_jacobian_last_op_indexing_is_identity(device: str):
+    if device == "cuda" and not torch.cuda.is_available():
+        pytest.skip("CUDA not available")
+
+    torch.manual_seed(0)
+    dtype = torch.float64
+
+    num_a = 6
+    n = 8
+    dim = 4
+
+    A0 = torch.randn(num_a, dim, device=device, dtype=dtype, requires_grad=True)
+    idx_a = torch.randint(0, num_a, (n,), device=device, dtype=torch.int32)
+
+    model = nn.Parameter(Track(A0))
+    out = model[idx_a]
+
+    (J_sparse,) = sparse_jacobian(out, [model])
+    assert J_sparse.layout == torch.sparse_bsr
+
+    def f(A: torch.Tensor) -> torch.Tensor:
+        return A[idx_a]
+
+    (JA,) = jacrev(f, argnums=(0,))(A0)
+    torch.testing.assert_close(J_sparse.to_dense(), _flatten_jac(JA), rtol=1e-10, atol=1e-10)
+    assert torch.equal(J_sparse.col_indices(), idx_a)