Multi-Dataset Sampling for Speaker Recognition Fine-tuning

[ap48923]

I am fine-tuning an ECAPA-TDNN model for speaker verification using this provided script in the speaker_tasks examples directory.

From the 00_NeMo_Primer.ipynb tutorial I see that I can include multiple datasets in manifest_path by entering a comma-separated list of manifest file paths. Is there a way for me to specify a sampling approach for each dataset during training? For example, given that I have 2 datasets, I would like to ensure that each training batch contains 50% from each dataset.

[VahidooX]

I don't think we have such a feature now. @nithinraok may know a solution for this.