broadcasting capabilities

README.md

@@ -35,7 +35,7 @@ The package consists of the following operations:
 * [**Scatter Min**](https://pytorch-scatter.readthedocs.io/en/latest/functions/min.html)
 * [**Scatter Max**](https://pytorch-scatter.readthedocs.io/en/latest/functions/max.html)
 
-All included operations work on varying data types, are implemented both for CPU and GPU and include a backwards implementation.
+All included operations are broadcastable, work on varying data types, and are implemented both for CPU and GPU with corresponding backward implementations.
 
 ## Installation
 

docs/source/index.rst

@@ -6,7 +6,7 @@ PyTorch Scatter Documentation
 This package consists of a small extension library of highly optimized sparse update (scatter) operations for the use in `PyTorch <http://pytorch.org/>`_, which are missing in the main package.
 Scatter operations can be roughly described as reduce operations based on a given "group-index" tensor.
 
-All included operations work on varying data types, are implemented both for CPU and GPU and include a backwards implementation.
+All included operations are broadcastable, work on varying data types, and are implemented both for CPU and GPU with corresponding backward implementations.
 
 .. toctree::
    :glob:

test/test_broadcasting.py

+import pytest
+import torch
+from torch_scatter import scatter_add
+
+from .utils import devices
+
+
+@pytest.mark.parametrize('device', devices)
+def test_broadcasting(device):
+    B, C, H, W = (4, 3, 8, 8)
+
+    src = torch.randn((B, C, H, W), device=device)
+    index = torch.randint(0, H, (B, 1, H, W)).to(device, torch.long)
+    out = scatter_add(src, index, dim=2, dim_size=H)
+    assert out.size() == (B, C, H, W)
+
+    src = torch.randn((B, 1, H, W), device=device)
+    index = torch.randint(0, H, (B, C, H, W)).to(device, torch.long)
+    out = scatter_add(src, index, dim=2, dim_size=H)
+    assert out.size() == (B, C, H, W)
+
+    src = torch.randn((B, 1, H, W), device=device)
+    index = torch.randint(0, H, (B, 1, H, W)).to(device, torch.long)
+    out = scatter_add(src, index, dim=2, dim_size=H)
+    assert out.size() == (B, 1, H, W)
+
+    src = torch.randn((B, C, H, W), device=device)
+    index = torch.randint(0, H, (H, )).to(device, torch.long)
+    out = scatter_add(src, index, dim=2, dim_size=H)
+    assert out.size() == (B, C, H, W)

torch_scatter/add.py

@@ -24,6 +24,8 @@ def scatter_add(src, index, dim=-1, out=None, dim_size=None, fill_value=0):
     :attr:`dim` = `i`, then :attr:`out` must be an n-dimensional tensor with
     size :math:`(x_0, ..., x_{i-1}, y, x_{i+1}, ..., x_{n-1})`. Moreover, the
     values of :attr:`index` must be between `0` and `out.size(dim) - 1`.
+    Both :attr:`src` and :attr:`index` are broadcasted in case their dimensions
+    do not match.
 
     For one-dimensional tensors, the operation computes
 

torch_scatter/utils/gen.py

+from __future__ import division
+
 from itertools import repeat
 
 
@@ -16,6 +18,18 @@ def gen(src, index, dim=-1, out=None, dim_size=None, fill_value=0):
         index_size[dim] = src.size(dim)
         index = index.view(index_size).expand_as(src)
 
+    # Broadcasting capabilties: Expand dimensions to match.
+    if src.dim() != index.dim():
+        raise ValueError(
+            ('Number of dimensions of src and index tensor do not match, '
+             'got {} and {}').format(src.dim(), index.dim()))
+
+    expand_size = []
+    for s, i in zip(src.size(), index.size()):
+        expand_size += [-1 if s == i and s != 1 and i != 1 else max(i, s)]
+    src = src.expand(expand_size)
+    index = index.expand_as(src)
+
     # Generate output tensor if not given.
     if out is None:
         out_size = list(src.size())