NeMo: problem with fine-tuning th e pre-trained `SpeakerNet` model for speaker recognition

Hi!

As an experiment, I was fine-tuning the pre-trained SpeakerNet model with the an4 dataset plus my own voice recordings. I used the example script speaker_reco_finetune.py and changed the configuration file according to the documentation and the colab notebook tutorial. But it appears something is always wrong with setting up the optimizer:

Traceback (most recent call last):
  File "/Users/xujinghua/speaker-recognition-nemo/scripts/experimental/fine_tune_an4.py", line 119, in <module>
    main()
  File "/Users/xujinghua/speaker-recognition-nemo/scripts/experimental/fine_tune_an4.py", line 108, in main
    speaker_model.setup_optimization(speaker_model._cfg.optim)
  File "/Users/xujinghua/miniconda3/lib/python3.7/site-packages/nemo/core/classes/modelPT.py", line 603, in setup_optimization
    optimizer=self._optimizer, scheduler_config=scheduler_config, train_dataloader=self._train_dl
  File "/Users/xujinghua/miniconda3/lib/python3.7/site-packages/nemo/core/optim/lr_scheduler.py", line 539, in prepare_lr_scheduler
    max_epochs, accumulate_grad_batches, num_workers, num_samples, batch_size, drop_last,
  File "/Users/xujinghua/miniconda3/lib/python3.7/site-packages/nemo/core/optim/lr_scheduler.py", line 587, in compute_max_steps
    sampler_num_samples = math.ceil(num_samples / num_workers)
ZeroDivisionError: division by zero

I tried a few ways to debug but nothing helped I have no clue what is going on wrong. Has anyone got an idea or run into such problems before? Help would be super appreciated!

Cheers, JX