Holger Roth

@unalakunal, any updates on this?

WSL shouldn't be the problem. Others have used it as well https://github.com/Project-MONAI/tutorials/issues/870. @siomvas might be able to confirm.

Maybe it's a general issue with limited computing resources (e.g. limited GPU memory). Can you try running with just one client? These are the steps 1. Set "[min_clients](https://github.com/Project-MONAI/tutorials/blob/1675efec26dd6e15dd9ca63e62900c678590cdb6/federated_learning/breast_density_challenge/code/configs/mammo_fedavg/config/config_fed_server.json#L4)" to 1...

Okay. So, this points to a resource issue when running all three clients. 1 client seems to be running fine. What's the GPU memory usage when you run just one...

That's a very good point. We are using `CacheData`. @unalakunal, you can try reducing the `cache_rate` [here](https://github.com/Project-MONAI/tutorials/blob/2b3d9a8901e56596f1efb2fb038daa6f64c88ee5/federated_learning/breast_density_challenge/code/pt/learners/mammo_learner.py#L242).

The patch size is really restricted only by the memory size of your GPU card. If you have enough memory, you can train on larger patches (including the whole image).

Please follow the information on the challenge website http://www.visceral.eu/benchmarks/anatomy3-open/

You can still contact the challenge organizers to get a download link for the data, I believe.

This is common problem with Caffe and/or hdf5, see https://github.com/BVLC/caffe/issues/3084. How large are your images and how many are you saving in a hdf5 file?

Can you share your prototxt file as well please?