Merge pull request #132 from silent567/master

Debugging problemtic behaviors of scatter_logsumexp for -inf

test/composite/test_logsumexp.py

@@ -3,19 +3,18 @@ from torch_scatter import scatter_logsumexp
 
 
 def test_logsumexp():
-    src = torch.tensor([0.5, 0, 0.5, -2.1, 3.2, 7, -1, -100])
-    src.requires_grad_()
-    index = torch.tensor([0, 1, 0, 1, 1, 2, 4, 4])
+    inputs = torch.tensor([
+        0.5, 0.5, 0.0, -2.1, 3.2, 7.0, -1.0, -100.0,
+        float('-inf'),
+        float('-inf'), 0.0
+    ])
+    inputs.requires_grad_()
+    index = torch.tensor([0, 0, 1, 1, 1, 2, 4, 4, 5, 6, 6])
+    splits = [2, 3, 1, 0, 2, 1, 2]
 
-    out = scatter_logsumexp(src, index)
+    outputs = scatter_logsumexp(inputs, index)
 
-    out0 = torch.logsumexp(torch.tensor([0.5, 0.5]), dim=-1)
-    out1 = torch.logsumexp(torch.tensor([0, -2.1, 3.2]), dim=-1)
-    out2 = torch.logsumexp(torch.tensor(7, dtype=torch.float), dim=-1)
-    out3 = torch.logsumexp(torch.tensor([], dtype=torch.float), dim=-1)
-    out4 = torch.tensor(-1, dtype=torch.float)
+    for src, out in zip(inputs.split(splits), outputs.unbind()):
+        assert out.tolist() == torch.logsumexp(src, dim=0).tolist()
 
-    expected = torch.stack([out0, out1, out2, out3, out4], dim=0)
-    assert torch.allclose(out, expected)
-
-    out.backward(torch.randn_like(out))
+    outputs.backward(torch.randn_like(outputs))

torch_scatter/composite/logsumexp.py

@@ -23,18 +23,19 @@ def scatter_logsumexp(src: torch.Tensor, index: torch.Tensor, dim: int = -1,
         if dim_size is None:
             dim_size = int(index.max()) + 1
 
-    size = src.size()
+    size = list(src.size())
     size[dim] = dim_size
     max_value_per_index = torch.full(size, float('-inf'), dtype=src.dtype,
                                      device=src.device)
-    scatter_max(src, index, dim, max_value_per_index, dim_size)[0]
+    scatter_max(src, index, dim, max_value_per_index, dim_size=dim_size)[0]
     max_per_src_element = max_value_per_index.gather(dim, index)
-    recentered_scores = src - max_per_src_element
+    recentered_score = src - max_per_src_element
+    recentered_score.masked_fill_(torch.isnan(recentered_score), float('-inf'))
 
     if out is not None:
-        out = out.sub_(max_per_src_element).exp_()
+        out = out.sub(max_per_src_element).exp()
 
-    sum_per_index = scatter_sum(recentered_scores.exp_(), index, dim, out,
+    sum_per_index = scatter_sum(recentered_score.exp_(), index, dim, out,
                                 dim_size)
 
     return sum_per_index.add_(eps).log_().add_(max_value_per_index)