fix: Ensure tensorlib precision used throughout TensorFlow backend

[matthewfeickert]

Description

While investigating Issue #1398, it was noticed that for 64b precision for the TensorFlow backed the precision was not properly being used throughout. tf.float32 was used as a default instead of the result of self.dtypemap["float"] and if TensorFlow Probability is given floats as parameters it will default to using 32b precision, and so requires that a tensor of the desired precision be given.

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

* Ensure that the precision requested in the set_backend API is used throughout the TensorFlow backend calculations
   - Determine dtype from dtypemap
   - Ensure TensorFlow Probability operations are given a tensor of the correct precision

[codecov]

Codecov Report

Merging #1399 (a06e20e) into master (cbf8d03) will increase coverage by 0.00%.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##           master    #1399   +/-   ##
=======================================
  Coverage   97.53%   97.53%           
=======================================
  Files          63       63           
  Lines        3808     3817    +9     
  Branches      538      538           
=======================================
+ Hits         3714     3723    +9     
  Misses         55       55           
  Partials       39       39           

Flag	Coverage Δ
contrib	24.12% <0.00%> (-0.06%)	⬇️
unittests	97.53% <100.00%> (+<0.01%)	⬆️

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

Impacted Files	Coverage Δ
src/pyhf/tensor/tensorflow_backend.py	97.22% <100.00%> (+0.18%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update cbf8d03...a06e20e. Read the comment docs.

[kratsg]

I don't think we should be adding so many self.astensor calls, should we?

[matthewfeickert]

I don't think we should be adding so many self.astensor calls, should we?

What do you propose instead? We had to do something similar in PR #1351, and at the moment on master this is a problem as setting the precision to 64b doesn't propagate to TensorFlow Probability operations if they're given floats.

>>> import pyhf
>>> pyhf.set_backend("tensorflow", precision="64b")
>>> pyhf.tensorlib.precision
'64b'
>>> pyhf.tensorlib.normal_logpdf(0.5, 0., 1.)
<tf.Tensor: shape=(), dtype=float32, numpy=-1.0439385>
>>> pyhf.tensorlib.normal(0.5, 0., 1.)
<tf.Tensor: shape=(), dtype=float32, numpy=0.35206532>

[matthewfeickert]

We also made Issue #1352 based off of these comments on PR #1351.

[kratsg]

Yeah, this is the cleanest way forward for now. We need to check that the fit times for tflow don't change.

[matthewfeickert]

We need to check that the fit times for tflow don't change.

Yeah, from looking at the CI times there isn't any obvious uptick, but we should note this in the Issue as well.