search

USGS-R / river-dl

Deep learning model for predicting environmental variables on river systems
Creative Commons Zero v1.0 Universal
21 stars 15 forks source link

rdl_torch_py env issues -> channel priority and long solve #164

Open janetrbarclay opened 2 years ago

janetrbarclay commented 2 years ago

Just adding some notes re: troubleshooting a really long (multiple hrs) build of the new rdl_torch_tf environment.

I tried removing conda-forge from the channel list (seems to be associated with slow environment solving - https://stackoverflow.com/questions/53250933/conda-takes-20-minutes-to-solve-environment-when-package-is-already-installed). That spend up the build, but gave me an environment with python 3.6 (though the default in my conda config is 3.9).

Setting the channel priority to strict conda config --env --set channel_priority strict and keeping conda-forge in the channel list worked (py = 3.9, pytorch = 10.2, cuda = 11.3, tf = 2.7).

Solving the environment still took awhile (~ 1hr), but it finished.

SimonTopp commented 2 years ago

Thanks for the update Janet! Just wanted to add that I the same build took me about ~35 minutes. I think all of this, particularly inconsistent results between users, is good motivation for updating the container for the repo. With that said, it might be tricky right now with the docker licensing and whatnot that's going on. What do people think? Has anyone else used the new environment.yml and if so have any issues come up?

jdiaz4302 commented 2 years ago

I have not used the new environment.yml. In the temperature/reservoir forecasting group, we've decided to drop Docker Desktop in favor of WSL2. I haven't needed to learn/use it yet, but Jesse acted like it's pretty easy if you're comfortable with Linux/the command line