search

bermanmaxim / superpixPool

Superpixel Pooling implemented in PyTorch and Chainer
MIT License
136 stars 26 forks source link

Some bugs in your pytorch code #1

Closed lxtGH closed 5 years ago

lxtGH commented 6 years ago

I found that your code didn't support bachtchsize>1 training ,in the kernel function spx_max_pooling_backward_kernel, you didn't allocate block dimension for batch

lxtGH commented 6 years ago

This means that you only allow for batchsize=1 for training ? @bermanmaxim

bermanmaxim commented 5 years ago

I think this means you need to split your batch first into individual images before sending it to the pooling layer. @Mathijssch do you confirm?

lxtGH commented 5 years ago

however when training in segmentation task, batch size is a key factor for performance, set batchsize=1 is not a correct choice.

Mathijssch commented 5 years ago

Yes, @bermanmaxim's suggestion is correct, that would be the only option in this implementation, which is indeed a bit limiting. This limitation is not present in the Pytorch version, though.

bermanmaxim commented 5 years ago

wontfix as we do not want to invest time in this project right now, PR welcome however! Thanks for your understanding.