How to implement tf.scatter_nd() function of tensorflow in pytorch

1

I want to implement the function called tf.scatter_nd() in pytorch. And I do not find a suitable function in pytorch. Does someone know how to do this?
Thank you!

2 Likes
2

Hi,

You can use the scatter_() function.

2 Likes
3

However, scatter_() is just a tensor operation. It does not have a autograd function. So maybe it cannot be used in network training.

4

All operations are actually autograd functions. You’ll get an error if the backward function is not implemented.

5

Hi, can you help me with a minimal examples of converting a tf.scatter_nd using torch.Tensor.scatter_add_.
The closest I’ve found is using the fancy indexing in pytorch, https://gist.github.com/airalcorn2/c7846d6fcb58a30b25ea6d97e16fe025.
But the fancy indexing technique does not match with tf.scatter_nd when there is duplication in the index tensor, because in tf the behavior is accumulating, while fancy indexing will just replace the value at that index.
Thanks.

6

index_add_ might be of help. I spent several hours looking for scatter_nd like function in pytorch, but all the links pointed to scatter_add_. The tensorflow like scatter_nd is more similar to index_add_.

1 Like