Reproductibility SODA and SVEA conv

nicklashansen / dmcontrol-generalization-benchmark

DMControl Generalization Benchmark

MIT License

166 stars 39 forks source link

Reproductibility SODA and SVEA conv #13

Closed AdilZouitine closed 2 years ago

AdilZouitine commented 2 years ago

Hi Nicklas, Thank you for your high-quality repo. We have trouble reproducing your results on finger spin with SODA and SVEA (we have between 500 and 600). Even in training, we don't achieve the performance shown. Are there any special settings or configurations for this environment?

Best regards

nicklashansen commented 2 years ago

Hi Adil, did you figure out the problem? We use an action repeat of 2 for this task. Depending on the number of seeds that you have run, I also wouldn't rule out that there's just a bad seed, it's RL after all.

AdilZouitine commented 2 years ago

Hi, I figured out the problem! Thanks!

tungts1101 commented 4 months ago

@AdilZouitine May I know what is the problem? We are trying to reproduce the result but the loss does not update after 290K steps.