Yasuhiro Fujita

BTW I sent Will Dabney an email asking the values of N and N' just a month ago, still don't have a reply. Anyone knows the values?

> N and N' should be as in the implicit_quantile_icml.gin file. Thank you for the information! > Also, out of curiosity -- did you figure out what was wrong? I...

I got a reply from Georg Ostrovski and confirmed that N=N'=64. He said the weight initialization was as below: - for all linear layers, all weights are drawn uniformly from...

Dopamine uses 2D convolutions with `padding=SAME`, which makes the number of activations after the three convolutions be `11*11*64=7744`, but it should be `padding=VALID`, thus `7*7*64=3136` (confirmed by Georg Ostrovski).

I tried `padding=VALID` for the same set of games by changing these lines: ``` --- a/dopamine/agents/implicit_quantile/implicit_quantile_agent.py +++ b/dopamine/agents/implicit_quantile/implicit_quantile_agent.py @@ -121,13 +121,16 @@ class ImplicitQuantileAgent(rainbow_agent.RainbowAgent): state_net = tf.div(state_net, 255.) state_net =...

FYI, I have pasted the plots of SAME vs VALID for the six games here, although they are all single runs. https://docs.google.com/document/d/1fsYzmNhfLvtPP4Cm-dbtp_MviH5WLo8qeiUJYMgRVio/edit?usp=sharing

> possibly due to the way Adam handles step size adaption Could you elaborate this?

@cathera I haven't checked differences between with and without 30-noop.

I do not know why but `--no-sandbox` has been commented out https://github.com/dexplo/dataframe_image/commit/40f0d8f18e4ce3e6012682ff8120b566aeca059c. I am facing the same issue with dataframe-image==0.1.11.

@PaleNeutron I missed that conditional `--no-sandbox`, thanks. Now I understand why it has been commented out. However, I still needed `--no-sandbox` to make it work as a non-root user when...