Skip to content

Support special case of broadcasting during Merge layers

Javier Duarte requested to merge github/fork/jmduarte/broadcast into master
  • Support special case of broadcasting during Merge layers
  • The case is Tensor(shape=[H, W, C]) + Tensor(shape=[1, 1, C]) = Tensor(shape=[H, W, C])
  • Was used for skip connections in small ResNet models
  • Only for io_stream

We also have a use case for io_parallel and Tensor(shape=[N, C]) + Tensor(shape=[N, 1]) = Tensor(shape=[N, C]) so it'd be great to generalize this PR to handle that as well

Merge request reports