search

mdroste / stata-pylearn

Supervised learning algorithms in Stata
MIT License
44 stars 11 forks source link

Bug: using an Anaconda environment for Python #3

Open laszlosandor opened 4 years ago

laszlosandor commented 4 years ago

Hi Michael, thanks for this (also, OI rocks). But the actual issue:

My primary Python installation lives under a conda environment (called idp, just following the Intel Distro's default). sh conda activate idp obviously did not work within Stata, as pylearn, setup did not find anything it wanted, instead, it downloaded quite a few things from pip without asking me. The latter behavior is a bit unpleasant and would be easy to fix.

But more fundamentally, how could I make Stata find a Python distro (installation) living in an environment?

Thanks!

mdroste commented 4 years ago

László - thanks so much for your thoughts! I'm a big fan of your work on portfolio choice with Raj ;)

I agree with everything you raise, particularly (now that you mention it) -pylearn, setup- automatically downloading stuff. Python intentionally makes it hard to do this (install libraries if they don't exist) from within a script, probably for this reason, and I should respect that with user prompts. I'll work this in ASAP.

The question about Python distros inside environments is a good one. I don't know the answer offhand. I'll play around with this and let you know ASAP too - but my prior is that Stata 16.1 might not play well with Anaconda environments. There is (at this point) a well-documented bug that Anaconda's included numpy/scipy don't work on some versions of OSX that the Stata developers are aware of, but I don't think there is much push to fix it atm.

mdroste commented 4 years ago

Just out of curiosity: would it be a serious problem if the recommendation was to have a vanilla Python installation for Stata, at least for now? I'm interested in your use case. It's hard to replicate the compatibility issues some users (esp on older Macbooks) have with Anaconda, and I'm worried this could just be a limitation of Stata 16.

laszlosandor commented 4 years ago

Thanks, Michael, I might be out of my depth here, I'm not even aware of the Anaconda issue. That said, my focus was not to use Anaconda necessarily, but to use the Intel distro, particularly for sklearn. That has significant (up to 200x) speedups for some operations and models, especially on recent server chips like some on Odyssey.

Of course, Intel does recommend conda for their distro, and I already have it anyway, so an independent installation is not very tempting.

But you're welcome to ignore this issue, it's your repo!