dreamerv2 VS polyaxon

In fact, they found that just reinforce gradients work in DM control now too: Dreamerv2 GitHub (they just needed to turn off gradients through the action path - which I guess was being passed back with straight-through estimation? I'm actually having a difficult time telling how the gradient is different on the action vs policy.log_prob(action)).

This is my implementation of Hafner et al. DreamerV2 algorithm. I found the PlaNet/Dreamer/DreamerV2 paper series to be some of the coolest RL research in recent years, showing convincingly that MBRL (model-based RL) does work and is competitive with model-free algorithms. And we all know that AGI will be model-based, right? :)

Last I checked, for running off a single node, the state of the art was Dreamerv2 https://github.com/danijar/dreamerv2

If you're not concerned about self-hosting, WandB is one of the more fully featured training monitoring tools (I've used it in the past without any issues but the lack of data and training privacy and lack of self-hosting possibilities makes it a hard no for anything that isn't scholastic). Polyaxon is an alternative but rewriting all your variable logging to conform to their requirements makes it very difficult to switch to it in the middle of a project so you have to commit to it from the get-go.

We use Polyaxon and it’s pretty good

Disclosure - I'm the author of Polyaxon.

I just came across https://github.com/polyaxon/polyaxon because mlflow gives me a hard time and costs my company money by the day because it is not working as expected.

For running experiments, http://polyaxon.com/ is a really good free open-source package that has lots of nice integrations so you can quickly run experiments in k8s but it might be overkill in some cases.

Polyaxon : https://polyaxon.com