Ankesh Anand

Hi @avnishn, Do you know if there's an update on v2 benchmark results?

Hi Alex, That's quite strange, those loss curves don't match what we saw in our runs. There definitely seems to be a bug or a numerical error in your runs...

We used pytorch 1.1.0 for our runs back then

So check if an older version of pytorch gives the correct results, this does look like a numerical stability bug. Also try increasing Adam's eps value. If that doesn't fix...

Thanks @DuaneNielsen! We've read the Keypoints paper, it's a really cool work! I don't know if we will be able to get to benchmarking it soon enough though. If you...

By environment features, do you mean the subset of RAM that we identified to be meaningful? If yes, then check out https://github.com/mila-iqia/atari-representation-learning/issues/40 which has some details on how to extract...

You need to run it for more steps. See Figure 2 of the paper: https://arxiv.org/pdf/1610.06918v1.pdf

I believe `async_sample=True` corresponds to asynchronous sampling *and* optimization. If you just want asynchronous sampling, you can use one of the parallel CPU sampler classes without setting `async_sample=True` in affinity

This is what we use for logging to wandb, basically creating a new class that inherits from existing runners: ```python class MinibatchRlEvalWandb(MinibatchRlEval): def log_diagnostics(self, itr, eval_traj_infos, eval_time): cum_steps = (itr...

Any updates on this @Miffyli? Support for headless rendering Also curious if you were able to find a workaround @danijar?