CSVIterableDataset doesn't work in the same way as CSVDataset

[Nic-Ma]

Is your feature request related to a problem? Please describe.
Here is the feedback from @holgerroth :

When I am using CSVIterableDataset, I get this error

Traceback (most recent call last):
  File "run_classification_criteo_monai_mlp.py", line 227, in <module>
    main()
  File "run_classification_criteo_monai_mlp.py", line 143, in main
    f"{step}/{len(train_ds) // train_loader.batch_size}, "
TypeError: object of type 'CSVIterableDataset' has no len()

Also, trying to set shuffle=True for the training dataset gives

Traceback (most recent call last):
  File "run_classification_criteo_monai_mlp.py", line 227, in <module>
    main()
  File "run_classification_criteo_monai_mlp.py", line 87, in main
    train_loader = torch.utils.data.DataLoader(
  File "/home/hroth/.local/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 224, in __init__
    raise ValueError(
ValueError: DataLoader with IterableDataset: expected unspecified shuffle option, but got shuffle=True

[Nic-Ma]

Thanks for @holgerroth 's feedback, let me try to enhance the doc-string of CSVIterableDataset to make the usage much more clear. And may also enhance the CSVIterableDataset with more features.

Thanks.

[wyli]

I feel we need a demo for the class, instead of docstring updates. And implement a shuffled buffer option. I can look into the shuffling soon..

[Nic-Ma]

After online discussion with @wyli , we aligned the idea to add shuffle option, I will try to update the PR #3198 ASAP.

Thanks.