fix: Ensure TensorFlow backend Poisson compatibility with other backends

[matthewfeickert]

Description

Resolves #997

As TensorFlow Probability v0.11.0+ is needed for PR #817, and as Issue #997 makes it clear that TF/TFP minor releases can be as breaking as normal major releases, require TensorFlow v2.3.0+ in the v2.3.X range and TensorFlow Probability v0.11.0+ in the v0.11.X range.

Ensure compatibility of TensorFlow Poisson's behavior with PyTorch and JAX while Issue #293 is being resolved by using the differentiable TensorFlow operations tf.logical_or and tf.where to allow for checking if the situation in which Poisson(n=0 | lam=0) is being encountered. This provides a working solution for both the TF v2.2.X and TFP V0.10.X API behavior as well as for the TF v2.3.+ and TFP v0.11.0+ API behavior. Also disallow TF v2.3.0 given tensorflow/tensorflow#40789.

Checklist Before Requesting Reviewer

• Tests are passing
• "WIP" removed from the title of the pull request
• Selected an Assignee for the PR to be responsible for the log summary

Before Merging

For the PR Assignees:

• Summarize commit messages into a comprehensive review of the PR

* Use differentiable TensorFlow operations to determine if Poisson(n=0 | lam=0) is being encountered and ensure a return value compatible with PyTorch and JAX
* Use TensorFlow releases compatible with v2.X with a lower bound of v2.2.1
   - Disallow TensorFlow v2.3.0 for accidental pinning of SciPy
* Use TensorFlow Probability releases compatible with v0.X with a lower bound of v0.10.1
* Use tf.errors.InvalidArgumentError for compatibility across TensorFlow releases

[matthewfeickert]

It appears that in TensorFlow Probability v0.11.0 they decided to make Pois(0|0) = 1 instead of nan. This breaks the following test_tensor.py test

pyhf/tests/test_tensor.py

Lines 201 to 206 in 8a2f7f6

	# poisson(lambda=0) is not defined, should return NaN
	assert tb.tolist(
	tb.poisson(tb.astensor([0, 0, 1, 1]), tb.astensor([0, 1, 0, 1]))
	) == pytest.approx(
	[np.nan, 0.3678794503211975, 0.0, 0.3678794503211975], nan_ok=True
	)

as

import pyhf

backends = ["numpy", "pytorch", "tensorflow", "jax"]

for backend in backends:
    pyhf.set_backend(backend, precision="64b")
    tb = pyhf.tensorlib
    poisson_results = tb.tolist(
        tb.poisson(tb.astensor([0, 0, 1, 1]), tb.astensor([0, 1, 0, 1]))
    )
    print(f"{backend} backend gives: {poisson_results}")

gives

numpy backend gives: [nan, 0.36787944117144233, 0.0, 0.36787944117144233]
pytorch backend gives: [nan, 0.36787944117144233, 0.0, 0.36787944117144233]
tensorflow backend gives: [1.0, 0.3678794411714423, 0.0, 0.3678794411714423]
jax backend gives: [nan, 0.367879441171442, 0.0, 0.367879441171442

So it seems that before this PR can be resolved that Issue #293 will need to get addressed again.

[matthewfeickert]

This is going to be on hold for quite sometime as TensorFlow v2.3.0 has a bug where scipy==1.4.1 was incorrectly introduced somehow. This has been fixed and so the next patch release of TensorFlow should remove all SciPy dependency completely making this a non-issue moving forward.

This will mean that we will need to require

    'tensorflow': [
        'tensorflow~=2.3.1',  # TensorFlow minor releases are as volatile as major
        'tensorflow-probability~=0.11.0',
    ],

and by that time there might be a patch release of TFP out that we might want as well.

[matthewfeickert]

The pinning of scipy==1.4.1 has been fixed in TensorFlow v2.3.1 but we still need to deal with the fact that TensorFlow Probability v0.11.0 has Pois(0|0) = 1 instead of nan.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 97.70%. Comparing base (b7a2c65) to head (f7b2c1a).

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #1001   +/-   ##
=======================================
  Coverage   97.70%   97.70%           
=======================================
  Files          63       63           
  Lines        4047     4050    +3     
  Branches      576      576           
=======================================
+ Hits         3954     3957    +3     
  Misses         54       54           
  Partials       39       39           

Flag	Coverage Δ
contrib	25.43% <16.66%> (+<0.01%)	⬆️
unittests	97.48% <100.00%> (+<0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[matthewfeickert]

I don't know of any nan equivalent in TensorFlow, as they use np.nan in their test suites.