Selman Özleyen

> yeah but seems like `apply` requires just `batch_size_sparse` times more memory, which means that `apply` and `solve` requires equally much when we `apply` only to one vector. Hence, `vmap`...

yep, here are the results ``` [50.00%] ··· Running (benchmarks.PointCloud.time_apply1--). [75.00%] ··· benchmarks.PointCloud.peakmem_apply1 ok [75.00%] ··· =============== ======= ======= ======= -- batch_size_sp --------------- ----------------------- batch_size_pc 1200 600 120 =============== =======...

> Hence, it's faster id we have a large batch size in `point_cloud`, and a small batch size in `sparsify`, is this correct? Thus, there doesn't seem to be a...

Btw I tried this comparison as you told me. The difference is so much and the one with cpu begining takes so much that it times out of 6 minutes...

closed with https://github.com/theislab/moscot/pull/696

I tried all of your suggestions and it only worked when I set the tau's to be 1.0 which is balanced. I checked again and it fails after this https://github.com/ott-jax/ott/commit/41906a2a1ade19aa154189fabd7c159a160c9bf3...

> agree, can you maybe try to run the example without moscot, but directly with ott-jax @selmanozleyen ? This might help @michalk8 @MUCDK ok I created it here https://github.com/ott-jax/ott/issues/519

hi, I overlooked this issue from a long time ago. I don't understand if this is a design problem or a feature etc.

@MUCDK do you think GENOT code in ottjax is final? My version in my thesis project is very different from the one in ottjax . If you like maybe we...

@michalk8, as you said when I normalize it works. But when they don't sum to 1 it still doesn't work in many cases. For example see the cases below. I'd...