Porting to Python 3.6 for Google Colab

[MilesCranmer]

Hi,

Great package, thanks for making this!

It just so happens I'm planning on teaching a tutorial on your framework to ~20 Cosmologists here at Princeton on Monday.

Now, my plan is to use colab.research.google.com for this; it's essentially Google Docs for Jupyter notebooks + you get free GPUs. So nobody needs to install this package on their machine, we just use the same in-browser notebook. (It's also great for letting users try out your package before installation).

However, colab uses Python 3.6 and this can't be changed. You state a requirement of 3.7+. Right now I see the error

File "/usr/local/lib/python3.6/dist-packages/sbi/inference/base.py", line 4
    from __future__ import annotations
                                      ^
SyntaxError: future feature annotations is not defined

So, I am curious: how important is 3.7 internally? If need be I can help put a bit of time to porting it to 3.6 - but is this doable and what do I need to change?

Thanks!
Miles

[alvorithm]

Hi Miles,
We went over this decision some time ago and we are trying to reconstruct that particular decision. Hopefully I can get you an answer tomorrow!

[alvorithm]

May I suggest to just remove the import? that won't break anything. I don't think you will encounter extensive breakage, if at all. Keep us informed of your progress, we'll try to support you.

[MilesCranmer]

Thanks so much @Meteore! I removed all lines with that statement in my fork here: https://github.com/MilesCranmer/sbi. So far so good with the annotations.

Now having issues with nflows (see below). I'll update this issue if I can fix it.

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-2-8dd34a6eaff4> in <module>()
----> 1 from sbi.examples.minimal import simple
      2 posterior = simple()
      3 print(posterior)

13 frames
/usr/local/lib/python3.6/dist-packages/nflows/utils/torchutils.py in <module>()
      1 """Various PyTorch utility functions."""
      2 
----> 3 import nflows.utils.typechecks as check
      4 import numpy as np
      5 import torch

AttributeError: module 'nflows' has no attribute 'utils'

[MilesCranmer]

Ah, so I think this issue is because pyknos uses an old version of nflows, where it has this mistaken line (should be from nflows.utils import typechecks as check):
https://github.com/bayesiains/nflows/blob/v0.11/nflows/utils/torchutils.py

[alvorithm]

Are you by chance using nflows master? There should be no problem if you install everything in a separate environment.

pyknos is using the latest released version of nflows (v0.11). We cannot depend on nflows/master.

[MilesCranmer]

I am using nflows v0.11. The only modification I have made in my fork was to remove the import __future__ statements.

So v0.11 of nflows looks to have this code: https://github.com/bayesiains/nflows/blob/278c81b393c48749eb6f01bc3118dd6fb075d32a/nflows/utils/torchutils.py#L3

import nflows.utils.typechecks as check

which produces the bug. This was fixed in master for python < 3.7 with this commit: bayesiains/nflows@84dc029#diff-b48fd0fab82b566d32e68a198d49d271.

I could try forcing setup.py to install this commit instead of v0.11?

By the way, if you would like to dig into this in-depth yourself, this issue is reproducible with this colab notebook (runnable in browser): https://colab.research.google.com/drive/1f-ztyOcAJRtDLxN6ElTB5evfssqGQ6lV?usp=sharing.

[MilesCranmer]

Okay, I fixed all of these in my fork, and am doing an upgrading of nflows to commit 84dc0298d99cc3bee9650a1501ef6aa578558080 after installing sbi with:

pip install --upgrade git+https://github.com/bayesiains/nflows@84dc0298d99cc3bee9650a1501ef6aa578558080

This prevents it complaining about the import syntax.

However, now am seeing what I think are issues with the annotations unfortunately:

---------------------------------------------------------------------------
NameError                                 Traceback (most recent call last)
<ipython-input-3-8dd34a6eaff4> in <module>()
----> 1 from sbi.examples.minimal import simple
      2 posterior = simple()
      3 print(posterior)

7 frames
/usr/local/lib/python3.6/dist-packages/sbi/user_input/user_input_checks_utils.py in MultipleIndependent()
    171     def __init__(
    172         self, dists: Sequence[Distribution], validate_args=None,
--> 173     ) -> MultipleIndependent:
    174         self._check_distributions(dists)
    175 

NameError: name 'MultipleIndependent' is not defined

I'll let you know if I manage to solve this

[MilesCranmer]

Okay so my guess is Python 3.7 added recursive type-checking, which lets you check for a type within the class definition of that type. But for 3.6 I'll just remove it. Does that sound correct?

Also, I'm not sure why but the import of some types, such as Dict, is not specified (e.g., in sbi/inference/posterior.py). I see an issue for Python 3.6, but maybe these are imported by default in some versions 3.7? Although it looks like default imports of Dict/List have been taken out of Python at some point at some point since they were added: python/mypy#2863. Not sure if it will work for future versions.

[MilesCranmer]

It's working! I'll submit a PR where we can discuss this more, if you'd like.

• also, typing.Sequence was also missing some imports in a couple files. The typing issues should all hopefully be fixed now.

Here's SBI running on the cloud in Google Colab with python 3.6:

Installation comes down to:

!pip install --upgrade git+https://github.com/MilesCranmer/sbi
!pip install --upgrade git+https://github.com/bayesiains/nflows@84dc0298d99cc3bee9650a1501ef6aa578558080