Depends on what sort of RL you are doing. If you are trying to train agents to play small games with vision the agent will need a small cnn to process images. This will need a gpu and what you have should be enough.
I was training on atari for a while with 1080ti. The games run on the cpu so you need a decent cpu as well.
I think it's a derivative of CFR, with a lot of optimizations and abstractions. They reduce the search space by grouping different card combinations together e.g. KK and QQ are treated same. They also limit the action space and only allow half-pot or full pot bets.
This is what I remember. I went through what they published briefly sometime back. So I could be wrong here.
Hi, I am one of the developers. This makes total sense, we were thinking of doing alerts. Do you have any specific alerts in mind? Do you think they should be customizable or general anomalies be enough?
I don't think you are missing a better workflow. Our current use case is monitoring the hardware while monitoring model training. In this case researchers tend to take a look at how their models are doing from time to time and they could check on hardware at the same time. This wouldn't be the case for general servers.
One of the problems with notebooks for literate programming is that it kind of breaks down when you define a class or a long function. The entire code has to be within a single block.
I haven't had time to really use it. Finished off coding a couple of days ago. I only used it with a bunch of old pytorch projects for testing on Lambda Labs cloud. I plan to use it for training some NLP models, that are too slow for my home GPU setup.
For now, starting servers, copying their ips/hostnames to the configuration files and stopping servers when training finishes seem is tedious. I think I will integrate it to the AWS (and other cloud) api so that you can automate starting and terminating servers.