Remove the one hot encoding

theislab / chemCPA

Code for "Predicting Cellular Responses to Novel Drug Perturbations at a Single-Cell Resolution", NeurIPS 2022.

https://arxiv.org/abs/2204.13545

MIT License

88 stars 23 forks source link

Remove the one hot encoding #23

Closed siboehm closed 2 years ago

siboehm commented 2 years ago

Adds indices to dataset + model, to remove the need for creating large OHEs when training on LINCS. For now the adversarial classification loss is just a Cross Entropy loss over all drugs, this should be fixed with #22

Closes #20

siboehm commented 2 years ago

I did a run on LINCS, 50 epochs take ~30minutes, evaluation still takes ~100 minutes.

No idea why the evaluation is still so slow, don't know if it's worth investigating again. It feels like there's no good reason why it should take so long.

There have been some issues with NaN loss but I'm having a hard time recreating them, I've only seen it appear once on LINCS, never on Trapnell.