Raphaël Marinier

Hello, For my own needs of configuring kelvin more finely, I implemented a new way to configure the light schedule very similarly to (and inspired) by @EmilyBjoerk's proposal. See [here](https://github.com/stefanwichmann/kelvin/compare/master...RaphaelMarinier:new_schedule2)....

@stefanwichmann, any interest in getting this kind of changes merged (once cleaned up)?

Also note that it would be easy to replace this by a simple and less specialized implementation that uses a state of shape [batch_size, , stack_size - 1] to perform...

Good observation. In practice, stack_frames() is agnostic to the spatial dimensions of the observation. It probably works on observations of shape [time, batch_size, , channels] regardless what are. The test...

Hello Szymon, I don't see any obvious mistake in your setup. It does seem that you use the right hyperparameters, and that indeed the only difference is the number of...

Hi again. I've been running my fork with the change described above for more than a year with no issue. It seems that it's still an often requested feature, e.g....

Indeed, it's mostly a duplicate (I was not aware of it), except that the default I propose in the doc is probably a better tradeoff (typically 3x less CPU usage,...

Regarding taking up all of GPU memory, this is the default of tensorflow. If you want to change this behavior, see [here](https://www.tensorflow.org/guide/using_gpu#allowing_gpu_memory_growth). We also see on our end that map...

The path is there: https://github.com/google-research/episodic-curiosity/blob/master/third_party/dmlab/dmlab_min_goal_distance.patch Maybe you tried the wrong command to apply it? Maybe you could describe the steps you performed to apply it?

Note that the CPU time did not change:  So the regression is probably due to some reduction in computation parallelism, and not the computations themselves taking...