search

TheMatrixMaster / omics-guided-gfn

Using gflownets to guide molecule search with multi-omic proxy models
Apache License 2.0
9 stars 0 forks source link

Pickle load issue -> nan loss #1

Closed Arkkienkeli closed 2 months ago

Arkkienkeli commented 2 months ago

Hi, very nice work, very interesting utilization of PUMA dataset :+1: I am trying to run the code like in ReadMe: python gflownet/src/gflownet/tasks/morph_frag.py with mmc_proxy and I tried a couple of samples, though I hit a couple of issues: First, if I start the code as is I get an error:

AttributeError: module 'multimodal_contrastive.networks.components' has no attribute 'GNEpropGIN'

after, I renamed MoleculeEncoder in multimodal_contrastive.networks.components to GNEpropGIN (and in the corresponding gmc_sm.yaml config as well). After that change I get:

Traceback (most recent call last):
  File "/omics-guided-gfn/gflownet/src/gflownet/trainer.py", line 223, in train_batch
    loss, info = self.algo.compute_batch_losses(self.model, self.critic1, self.critic2,
  File "/omics-guided-gfn/gflownet/src/gflownet/algo/soft_actor_critic.py", line 260, in compute_batch_losses
    raise ValueError("loss is not finite")
ValueError: loss is not finite

Thank you!

TheMatrixMaster commented 2 months ago

Hi @Arkkienkeli, thanks for your interest in our work and for helping identify these issues. I've gone in and fixed some of the things you mentioned and it should hopefully work ok now. Can you give it another try and let me know if anything isn't resolved?

Arkkienkeli commented 2 months ago

Yes, it starts now, thank you!