ConnectionResetError: [Errno 104] Connection reset by peer about gym-city

[smearle]

You need to launch the visdom server as described in the Readme. I think only the second error message is relevant here. First is generic "I was multiprocessing and something went wrong", third is a warning that came with the original micropolis as well, I believe. Or, you can just ignore these messages. The training/evaluation reward graphs will be output as png files anyway.

[leckylao]

Thank you for your reply @smearle. Visdom is up and running fine on 8097.

These are the log output from Visdom when running the train.py script. So looks like it's able to connect. Any ideas?

As shown it can set up a new session at the beginning but then has the "Unable to init server"

FYI I am running docker with the Dockerfile. But I don't think it's related to docker as I can see Visdom is up running and there's connection messages

[smearle]

Strange. Does the code work otherwise?
I think I've been getting this bug too, but simply ignoring it and referring to the "png" files for training/eval reward graphs. Keep me in the loop if you'd like to try hunting down the bug here.

EDIT: actually, I can launch a visdom server and connect to it successfully. However, there is a bug where the graph is reproduced each time it is updated, leaving a bunch of duplicate graphs in visdom. I'm on the "pcgrl_mix" branch (you will need to comment out some imports -- of other external gym environments -- to use it) -- might be slightly different than master in these respects.
My best guess is either this bug is fixed on the pcgrl_mix branch, or, it is a docker thing.

[leckylao]

No, the trained_models folder is empty.

I am uploading my image built from master which is huge 5.57GB. Then could you have a test? And then I will try the pcgrl_mix branch.

EDIT: image has been uploaded to https://hub.docker.com/repository/docker/leckylao/gym-city 3.67GB. And you can run docker pull leckylao/gym-city:latest to download. Then you can test it with:

docker run leckylao/gym-city:latest python3 -m visdom.server
then
docker ps to get the container name
then
docker exec -ti {cointainer_name} bash
then
python3 train.py --experiment test_0 --algo a2c --model FullyConv --num-process 24 --map-width 16 --render

EDIT: Just tried pcgrl_mix branch but still got the same connection issue

[leckylao]

I am now running Ubuntu VM and can confirm this issue is docker specific. In docker Visdom server stopped with message INFO:root:Application Started and missing You can navigate to http://localhost:8097. And when I ctrl-c from Visdom in docker it shows You can navigate to http://dfdfaweadsss:8097. So I guess the issue is related to Visdom trying to get hostname from docker. And I reckon adding this to README that due to this running in docker is not supported.

[smearle]

Aha, thank you for getting to the bottom of this docker issue! For now, you're good to simply comment out the import I believe. The import is because I've been toying around with applying the training loop in this repo to the environments in the repo for "PCGRL" (and on this experimental branch): https://github.com/amidos2006/gym-pcgrl/tree/FractalNet
I hope you're able to get to training soon. Keep me in the loop of any further hiccups.