Adjust sweeps for better adv loss

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

Adjust sweeps for better adv loss #60

Closed siboehm closed 2 years ago

siboehm commented 2 years ago

Closes #57

I looked at the main adversary parameter differences between the models with the best (=lowest) perturbation disentanglement score and those with the worst (=highest) score and adjusted them.

MxMstrmn commented 2 years ago

  model.hparams.adversary_lr:
    min: 1e-5 -> min: 5e-5
  model.hparams.adversary_wd:
    min: 1e-6 -> min: 1e-5
  model.hparams.reg_adversary:
    new: min: 5    max: 100
  model.hparams.penalty_adversary:
    min: 1e-2 -> min: 1

Basically, everything is set higher with identical maxima.