search

TRI-AMDD / htp_md

Shared repo for trajectory analysis and infrastructure development
MIT License
13 stars 8 forks source link

Random forest pre-trained model #24

Closed shenggong1996 closed 3 years ago

shenggong1996 commented 3 years ago

Random forest models to predict polymer transport property

txie-93 commented 3 years ago

I don't know why the test failed on Windows. Seems to be related to commit 853da0e @HakyungKwon-TRI @danielschweigert-TRI

HakyungKwon-TRI commented 3 years ago

@txie-93 I pushed a constraint on sklearn version (that seemed to be the error causing fatal exception) It's now running into errors with the feature size vectors. Not sure why this is not an issue with the macos/ubuntu builds, but I am looking into it. Can you check if the errors that it's complaining about (you can see it under Actions > find the latest failed run > click through until you see Annotations > pytest) are real?

txie-93 commented 3 years ago

@HakyungKwon-TRI This looks odd. I am asking Sheng to see what he thinks.

txie-93 commented 3 years ago

@HakyungKwon-TRI How is mordred installed? I had a quick chat with Sheng, and we suspect the version of mordred on linux and windows could be different. As a result, the number of features may also be different

HakyungKwon-TRI commented 3 years ago

@txie-93 they're both on 1.2.0 version, from conda-forge. Installed via conda using the env.yml.

Windows: https://github.com/TRI-AMDD/htp_md/runs/3538718102?check_suite_focus=true MacOS/Linux: https://github.com/TRI-AMDD/htp_md/runs/3538718124?check_suite_focus=true

txie-93 commented 3 years ago

I did some checks. The number of features on linux is 1611. But https://github.com/TRI-AMDD/htp_md/runs/3538718102?check_suite_focus=true says the number of features on Windows is 1265. So I think this is probably the issue. But I don't have a windows computer to reproduce.

@HakyungKwon-TRI Where did you see the mordred version on Windows is 1.2.0 on windows? It seems that I cannot find it. It is odd that the number of features are different if the version is the same.

HakyungKwon-TRI commented 3 years ago

@txie-93 If you go to the active run (https://github.com/TRI-AMDD/htp_md/runs/3540539025?check_suite_focus=true), it's under Run Conda Install mkl=2018. Scroll, and you should see a list of all packages that were installed as part of conda install.

I'll try to see if there are other ways to debug (I also only have access to macos). We can consider logging this and moving on for now, since macos/linux tests pass.

txie-93 commented 3 years ago

I agree that we could merge this but leave an issue on the repo for the future. @shenggong1996 is on vacation now and we can ask him to explore more once he is back. Can you merge this pull request? @HakyungKwon-TRI