Fixes numpy >2.0 compatibility for asr while replicating existing behavior

[andylamp]

What does this PR do ?

The PR attempts to restore functionality with recent numpy versions; numpy 2.0 removed sctypes and this is breaking asr functionality.

edit: closed the PR due to commits not being signed-off properly and rebasing caused issues :)

Collection: asr

Changelog

• As said above numpy from version 2.0 onwards removed sctypes functionality, which breaks asr functionality in various places (e.g. transcribe_speech, example notebooks etc).
• The changes reside in the _convert_samples_to_float32 function residing in both feature_loader.py and segment.py.
• There have been other issues relating to this (e.g. Replace usage of np.sctypes with np.issubdtype #10839 and Replace usage of np.sctypes with np.issubdtype #10180), however these do not replicate existing functionality.

Usage

The usage is that ASR tasks that call _convert_samples_to_float32 functions now succeed. As noted above, while this has been attempted to be addressed in the PRs mentioned above, I believe that the functionality is not replicated accurately.

More concretely, if you check the output of sctype for the types of int and float in supported versions of numpy you'd get the following:

print(np.sctypes["int"])
[<class 'numpy.int8'>, <class 'numpy.int16'>, <class 'numpy.int32'>, <class 'numpy.int64'>]
print(np.sctypes["float"])
[<class 'numpy.float16'>, <class 'numpy.float32'>, <class 'numpy.float64'>]

However, if we use the issubdtype to perform this the set will be wider, case on point for floating point:

import numpy as np
float128 = np.floating[128]
np.issubdtype(float128, np.floating)
Out[6]: True

And, also for integers the set mostly covers the output of sctype for signed ones. However any subclass from signedinteger or unsignedinteger will return true. A more concrete example would be,

import numpy as np
int96 = np.signedinteger[96]
np.issubdtype(int96, np.integer)
Out[4]: True
uint96 = np.unsignedinteger[96]
np.issubdtype(uint96, np.integer)
Out[5]: True

Therefore, in this case not only we consider int{8,16,32,64} but also uint{8,16,32,64} which is not the expected result when querying sctype['int'] leading to potentially unexpected behavior. Case on point,

import numpy as np
np.issubdtype(np.uint8, np.integer)
Out[3]: True

GitHub Actions CI

The Jenkins CI system has been replaced by GitHub Actions self-hosted runners.

The GitHub Actions CI will run automatically when the "Run CICD" label is added to the PR.
To re-run CI remove and add the label again.
To run CI on an untrusted fork, a NeMo user with write access must first click "Approve and run".

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests? No new tests were required.
• Did you add or update any necessary documentation? No documentation update required.
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc) No optional components are affected.
• Reviewer: Does the PR have correct import guards for all optional libraries?

PR Type:

• New Feature
• Bugfix
• Documentation

If you haven't finished some of the above items you can still open "Draft" PR.

Who can review?

Anyone in the community is free to review the PR once the checks have passed.
Contributor guidelines contains specific people who can review PRs to various areas.

Additional Information

The fix is backwards compatible with all supported numpy 1.x versions as well. Therefore, it should pose minimal risk wrt to integration.

[github-actions]

beep boop 🤖: 🙏 The following files have warnings. In case you are familiar with these, please try helping us to improve the code base.

---

Your code was analyzed with PyLint. The following annotations have been identified:

------------------------------------
Your code has been rated at 10.00/10

---

Thank you for improving NeMo's documentation!