YouTip

* * Pytorch torch Reference Manual](#) `torch.scatter_add` is a function in PyTorch used to add values from a source tensor to specified positions. It adds the values of `src` to `input` at the positions specified by `index`. ### Function Definition torch.scatter_add(input, dim, index, src) **Parameters**: * `input` (Tensor): The input tensor. * `dim` (int): The dimension along which to scatter. * `index` (Tensor): The index tensor, specifying the positions in `input` to which the values in `src` will be added. * `src` (Tensor): The source tensor, containing the values to be added. **Return Value**: * `torch.Tensor`: Returns the modified tensor. * * * ## Usage Examples ## Example import torch # Create input tensor input= torch.zeros(3,5) # Create index and source index = torch.tensor([[0,1,2,0,0], [1,2,0,1,2], [2,0,1,2,0]]) src = torch.tensor([[1,1,1,1,1], [2,2,2,2,2], [3,3,3,3,3]]) # Scatter and add along dim=0 output = torch.scatter_add(input, dim=0, index=index, src=src) print("Input:") print(input) print("nIndex:") print(index) print("nSource:") print(src) print("nResult:") print(output) Output result: Input: tensor([[0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.], [0., 0., 0., 0., 0.]])Index: tensor([[0, 1, 2, 0, 0], [1, 2, 0, 1, 2], [2, 0, 1, 2, 0]])Source: tensor([[1., 1., 1., 1., 1.], [2., 2., 2., 2., 2.], [3., 3., 3., 3., 3.]])Result: tensor([[4., 1., 2., 4., 4.], [2., 1., 2., 2., 2.], [3., 3., 1., 3., 3.]]) ## Example import torch # Using dim=1 input= torch.zeros(3,5) index = torch.tensor([[0,1,2,1,0], [1,2,0,2,1], [0,1,1,0,2]]) src = torch.arange(1,6).float() output = torch.scatter_add(input, dim=1, index=index, src=src) print("Scatter along dim=1:") print(output) Output result: Scatter along dim=1: tensor([[ 6., 3., 3., 0., 0.], [ 3., 6., 3., 0., 0.], [ 2., 4., 5., 0., 0.]]) ## Example import torch # Application scenario for aggregating values from multiple positions # For example, accumulating neighbor node features in graph neural networks # Simulate initial features of nodes node_features = torch.zeros(4,3) # Simulate edge connections (source nodes pointing to target nodes) edge_index = torch.tensor([0,1,2,3,0,1])# Source nodes of edges edge_weights = torch.tensor([1.0,2.0,3.0,1.5,2.5,0.5]) # Create weighted source node features for each edge src_features = torch.randn(6,3) * edge_weights.unsqueeze(1) # Accumulate features to target nodes (simplified here, actual implementation needs to follow edge target nodes) target_nodes = torch.tensor([0,0,1,1,2,3]) index = target_nodes output = torch.scatter_add(node_features,0, index.unsqueeze(1).expand_as(src_features), src_features) print("Node feature shape:", node_features.shape) print("Accumulated features:", output) * * * Note: `torch.scatter_add` does not modify the original input tensor, but returns a new tensor. Multiple indices can point to the same position, and their values will be accumulated. This function is the inverse operation of `torch.gather`. * * Pytorch torch Reference Manual](#)