'CombinedLoader' object has no attribute 'batch_sampler' when using a CombinedLoader in predict_dataloader

[franciscojferrari]

🐛 Bug

When returning a CombinedLoader with a dictionary in a predict_dataloader throws AttributeError: 'CombinedLoader' object has no attribute 'batch_sampler'

predict_dataloader should behave the same as test_dataloader where this behavior doesn't happen in order to be consistent between dataloaders

To Reproduce

Please check colab link : https://colab.research.google.com/drive/1EUc_SoomMTNw148xo4Szvozh4tJwkWC1?usp=sharing

Expected behavior

predict_dataloader should behave the same as test_dataloader where this behavior doesn't happen in order to be consistent between dataloaders

Environment

• CUDA:
  - GPU:
  - Tesla P100-PCIE-16GB
  - available: True
  - version: 11.1* Packages:
  - numpy: 1.19.5
  - pyTorch_debug: False
  - pyTorch_version: 1.10.0+cu111
  - pytorch-lightning: 1.5.5
  - tqdm: 4.62.3
• System:
  - OS: Linux
  - architecture:
  - 64bit
  -
  - processor: x86_64
  - python: 3.7.12
  - version: Proposal for help #1 SMP Sat Jun 5 09:50:34 PDT 2021

Additional context

cc @justusschock @awaelchli @ninginthecloud

[awaelchli]

The issue here boils down to us having to ensemble a batch sampler for a combined dataloader (holding dataloaders of which a subset have batch samplers). This issue occurs mainly because in the prediction we require a batch sampler to keep track of the indices of each sample.

[rohitgr7]

should we drop support for combinedLoader during prediction, because of the incompatibility?

[justusschock]

I am not convinced that this would be the way to go. I talked to @awaelchli yesterday about different ways to make this possible. A hotfix would be to make the attribute access optional.

For a longterm solution we probably need to revisit the loader itself, but we should not drop support I think

[awaelchli]

I'll send that hotfix since I modified this code recently.