[lhotse] Support for NeMo tarred manifests with offset field

[pzelasko]

What does this PR do ?

Support for NeMo tarred manifests with offset field in lhotse dataloading.

Collection: ASR

Changelog

• Add specific line by line info of high level changes in this PR.

Usage

• You can potentially add a usage example below

# Add a code snippet demonstrating how to use this 

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The Jenkins CI system has been replaced by GitHub Actions self-hosted runners.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?
• Did you add or update any necessary documentation?
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
  • 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

• Related to # (issue)

[zhehuaichen]

looks good to me . @KunalDhawan or @tango4j can either of you guys also take a look?
I will do a test with this PR

[KunalDhawan]

LGTM, thank you Piotr!

[vadimkantorov]

@pzelasko I checked, caching TarInfos with a single linear scan lets then extract in O(1):

def taropen_build(tar_path):
    import tarfile
    tar = tarfile.open(tar_path)
    tarinfo_index = {tarinfo.name : tarinfo for tarinfo in tar.getmembers()}
    return lambda member_path, mode: tar.extractfile(tarinfo_index[member_path]) if mode == 'rb' else None

    # instead for uncompressed tars (which makes sense for audio files) can do without caching the TarInfo structure and just cache two ints per string file name, this makes it more amenable to sharing between DataLoader threads and easier to sync in tensors
    #import io
    #tarinfo_index = {tarinfo.name : (tarinfo.offset_data, tarinfo.size) for tarinfo in tar.getmembers()}
    #return lambda member_path, mode: [io.BytesIO(f.seek(e[0]) and f.read(e[1])) for f in [open(tar_path, mode)] for e in [tarinfo_index[member_path]]][0] if mode == 'rb' else None

if __name__ == '__main__':
    taropen = taropen_build('myfile.tar')
    print(taropen('mymember.txt', 'rb').read())

Supporting .tars with wav/opus files inside would be very nice, as these small audio files can be in 10s or 100s of thousands which is always a pressure on the file system and adds time to unpacking

[pzelasko]

We already support tars with wav/opus. Linear scan is not feasible for URL data sources (cloud storage, AIStore).