zchuning / repo Goto Github PK

Hello, I want to do a comparison on standard DMC, but I don't know the hyperparameters of the standard task. Could you please provide the final convergence performance of standard DMC, which is the performance shown in Figure 11 in the appendix of the paper?

Hi, thank you for publishing awesome work with the code. This is very helpful.

I have two questions regard DMC experiments.

1. I wonder what X-axis in Figure 5 means. Does it mean 'total environment steps with action repeat of 2 counted' or 'a number of env.step() call'? Based on the implementation, I assume it is the latter but I would like to make sure.
2. Could you make the performances used for creating plots (specifically Figure 5 and Figure 7) available? This would be helpful to compare RePo with other methods.

Again, thank you so much!

Hello, I ran the distraction DMC experiment. The hyperparameters were set to Table 2 in Appendix C, but the performance was much worse than what was shown in the paper, and the variance was large. Why is this? Are there other settings that need attention?

Hi, I'm interested in this work. I find that 'calibration_mode' has two forms, 'pair' and 'simple_pair', and the default is 'simple_pair'. What's the difference between them?

And I find that calibration loss in code seems to be a cross-entropy loss, rather than the L2 loss mentioned in the paper. Why do this? Will it improve performance?

Hi, I have another question. As mentioned in the title of the 4th section: "RePo: Parsimonious Representation Learning without Reconstruction", but in the repo.py, there is a "# Reconstruction loss for probing", why is that？