SpeechLM2 : Add support for offset key in Multimodal conversation

[AudranBert]

Important

The Update branch button must only be pressed in very rare occassions.
An outdated branch is never blocking the merge of a PR.
Please reach out to the automation team before pressing that button.

What does this PR do ?

Add support for offset key in Multimodal conversation. Also make that it uses the duration key to load audios. Goal is to be able to have multiples segments per audio.

Collection: speechlm2

Changelog

• Offset key support (if present)
• Duration key now used (if present) for audio loading
• Export_conversations_to_tar.py now write segments (use duration and offset) in .tar instead of the full audio

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The Jenkins CI system has been replaced by GitHub Actions self-hosted runners.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 SpeechLM not using "offset" key #15080

[AudranBert]

Hi, @pzelasko what do you think of the code logic? Any concerns or warnings?
From my testings, it works with tar and jsonl and does not load the whole file like it did before.
I will write automated test when I can.

[pzelasko]

Thanks! Yeah I think that will work. Please make sure to cover both reading and writing + reading cases with the new offset field.

[AudranBert]

Thanks! Yeah I think that will work. Please make sure to cover both reading and writing + reading cases with the new offset field.

Except for NeMoMultimodalConversationTarWriter I do not see functions to write NeMoMultimodalConversation, is there some somewhere else?

[AudranBert]

Thanks! Yeah I think that will work. Please make sure to cover both reading and writing + reading cases with the new offset field.

Do I need to add offset/duration support to NeMoMultimodalConversationShareGPTJsonlAdapter ? It seems like an established format.

[AudranBert]

tests/collections/common/test_lhotse_multimodal_dataloading.py are passing @pzelasko. Is it good? For now, I didn't touch NeMoMultimodalConversationShareGPTJsonlAdapter but I can update it if needed (but it is missing the duration key already?). NeMoMultimodalConversationJsonlAdapter is fully working : loading jsonl or tar and writing tar with NeMoMultimodalConversationTarWriter.

The CI job Isort and Black Formatting / reformat_with_isort_and_black (pull_request_target) is not working, it seems to be because the PR is from a forked repository, I don't understand what I should do and if I should do something?
I can't add the ``run CICD` label too.

[pzelasko]

Except for NeMoMultimodalConversationTarWriter I do not see functions to write NeMoMultimodalConversation, is there some somewhere else?

I don't think there are - it's OK to just test it with TarWriter

Do I need to add offset/duration support to NeMoMultimodalConversationShareGPTJsonlAdapter ? It seems like an established format.

Could you do that while you're at it? It will be definitely helpful.

Don't worry about the CI job, I just approved it, will probably work on re-trigger.

[AudranBert]

Except for NeMoMultimodalConversationTarWriter I do not see functions to write NeMoMultimodalConversation, is there some somewhere else?

I don't think there are - it's OK to just test it with TarWriter

Do I need to add offset/duration support to NeMoMultimodalConversationShareGPTJsonlAdapter ? It seems like an established format.

Could you do that while you're at it? It will be definitely helpful.

Don't worry about the CI job, I just approved it, will probably work on re-trigger.

Hi @pzelasko, I added a line in the doc to acknowledge the offset param. I think the PR is finished, except if you have some feedbacks. Just a side question, the duration param is mandatory for NeMo multimodal conversations? It is used to estimate the number of tokens? Because it was unused in the text_adapters.py file (now it is).

[pzelasko]

Just a side question, the duration param is mandatory for NeMo multimodal conversations? It is used to estimate the number of tokens? Because it was unused in the text_adapters.py file (now it is).

Yes, it was unused, and present there as useful metadata for reading the manifest files in other contexts (EDA etc).

[AudranBert]

Hi @pzelasko, (sorry for the ping) the PR needs a review

[pzelasko]

Thanks for the contribution @AudranBert !