search

maxweissenbacher / latentRL

Extracting the latent space of chaotic systems to accelerate convergence of RL controllers.
0 stars 0 forks source link

Model exchange/ save CAEs #4

Open eliseoe opened 5 months ago

eliseoe commented 5 months ago

This issue is to discuss how we want to exchange and save models. As far as I know, these are some options:

  1. Sweep through models on wandb. A) Determine configuration of the best model and train it again from scratch locally. B) Determine configuration of the best model and upload weights of that specific model on github.
  2. Save all trained models on wandb. Make a research team, then we can share models directly. Downside: As a user, you can only have one research team.
  3. Explore alternative platforms?
maxweissenbacher commented 5 months ago

Thanks Elise! I think using WandB is a great option - I already have a team there but I'll just make another account to use here :)

For sure we don't need to upload all the models we train atm - for initial experimentation I think it will be enough to pick a model that works well (using wandb) and then share the model on OneDrive?

I've made a shared folder here

eliseoe commented 5 months ago

Great, thanks a lot @maxweissenbacher. For me, the link above is broken but this one here works.

maxweissenbacher commented 5 months ago

Nice - once you have a trained model you could just put it there - I'll get around to setting up the RL code tomorrow probably so we can already do a little test run and see what happens! :)

eliseoe commented 5 months ago

I uploaded data used for L=22 and a CAE model, see details here.

OneDrive seems a good solution for now!

maxweissenbacher commented 4 months ago

I've added some more data here.

eliseoe commented 3 months ago

I added some cae models trained on the symlog loss here.

This AE takes unnormalized data and is trained based on u_SAC_NU0.05_A20_NUMENVS5_BURNIN5000.dat (see ks.json). The wrapper on branch rl was updated accordingly. Let's start with model solar-sweep-1.