Ilya Kostrikov

Hi, cov_a and cov_g are implemented this way in order to be consistent the original KFAC code from OpenAI. It true that it can be implemented in a different way....

The GAIL implementation should work with convolutions as well. There were no specific reasons to force observations to be 1D. I will accept a pull request that adds ConvNets support...

Hi, I think I should update the pre-trained models anyway (also load the right parameters). Which environments do you need? Best regards, Ilya

Hi, Please install baselines by compiling from the source code as described in readme.

I did it mostly to match the baselines setup. I don't have enough resources to run a massive hyper parameter search myself and I assume that they compare results with...

Works on my linux machine. Can you provide more details?

Great! Can you submit a PR with the fix? > On Aug 5, 2019, at 7:25 AM, dmitrySorokin wrote: > > Ubuntu 18.04 > I am using latest commit from...

Sorry @miguelsuau, I've just noticed the issue. Yes, it backpropagates through the whole sequence (128 in this case).

I will accept a PR.

Hi @naivety77 ! Did you set the flag cuda_deterministic to True? Asynchronous execution on GPU is usually the main reason under getting different results for different runs.