Use datum dataset for PyTorch

[CourchesneA]

I would like to use datumaro as a standard dataset format, but in order to use it for pytorch I need to create a pytorch.Dataset wrapper for it. The dataset interface requires to implement the __getitem__(idx:int) function, which is not available on a datumaro dataset.

I am currently using the code from datumaro.FrameworkConverter, which keeps a list of item ids in a list and use indexes from this list. However, once it is converted this way, I can no longer use datumaro dataset manipulation utilities such as split and merge.

Would it be possible to get access to an integer __getitem__ function directly on the datumaro dataset ?

[sooahleex]

Hi @CourchesneA, Sorry for late response.

We plan to incorporate the mentioned __getitem__ functionality into the Datumaro Dataset. Once this feature is implemented, you will be able to retrieve datasetitem within the dataset using an index as you desire. I will provide further updates once this enhancement is included in the release.

Thank you for your patience and understanding.

[CourchesneA]

That is great news. Thanks for the update !

[wonjuleee]

Dear @CourchesneA, as following your suggestion, we have implemented random accessing method with __getitem__ at #1247. This will be publicly released with Datumaro 2.0 at the mid of March. Sincerely thank you for your continued interests and suggestions.