any plans to port the code?

[linminhtoo]

hello authors,

i came to know of this work through your team's study on using ESP for drug discovery projects: https://pubs.acs.org/doi/10.1021/acs.jmedchem.2c00164?fig=tgr1&ref=pdf

However, it appears that this repo was built several years back and was written in Python2 and uses Tensorflow 1, which is kind of outdated and will not work on newer GPUs. I was wondering if the team has any plans to upgrade the codebase? Or including a PyTorch version as well.

Thanks,
Min Htoo

[fredludlow]

Good question... We're not actively working on the codebase, but I'd like to keep on top of bit-rot where possible - what specific GPUs are causing problems?

[linminhtoo]

hmm yea, i think if I really force myself to use tensorflow1, it won't be compatible with the newer GPUs (like RTX3080 that I'm using) as the CUDA version supported by tensorflow1 is only up to 8 or 9 (we're already at 11.X)

I tried to use tensorflow2 but there's quite a lot of code to change in the repo to make it compatible and I spent some time trying to migrate the tensorflow code but I didn't succeed in running the model inference.

[fredludlow]

Gotcha.

There's a python 3.7 port that I know of (still tensorflow1 but python 3 probably makes updating other parts easier), let me find out if it's publicly available or can be made so.

If you just want inference working then our internal "ESP server" runs more or less the code in this repo on a fairly old Xeon and does the inference on the CPU and it's still more than fast enough for interactive use in a webapp. Obviously if you want to retrain or further develop the model that's not very helpful.

[linminhtoo]

thanks for this @fredludlow , yes, at least the python3 version would certainly be helpful, as some of the errors I ran into
seemed to be due to the Python2 vs 3 mismatch. Looking forward to your response!

Interesting that CPU is fast enough. If all else fails, I could try running the inference on CPU & using it that way, although like you mentioned, I will not be able to re-train it unless I re-write the code

[linminhtoo]

hello @fredludlow , it's been a while but I'm still interested in this work so was wondering if any progress has been made on this front. cheers.

[fredludlow]

There's a PR (#6) which I still haven't had a chance to look at properly, but that promises to be running correctly on 3.x

Apologies for the very overlong delay reviewing this (if you find it works for you that's a very good +1 for me just merging it :) )

[linminhtoo]

There's a PR (#6) which I still haven't had a chance to look at properly, but that promises to be running correctly on 3.x

Apologies for the very overlong delay reviewing this (if you find it works for you that's a very good +1 for me just merging it :) )

oh thanks, I'll check it out!