Replace legacy code in `TextCNN` example

[KickItLikeShika]

🚀 Feature

torchtext in their latest release 0.9.0 moved a lot of features to legacy code, and because of that we temporary updated our TextCNN example to compatible with the legacy code in #1890 (as quick fix for now).
Now we need to replace the legacy code in this example with the provided features in torchtext==0.9.0.

• torchtext==0.9.0 release note: https://github.com/pytorch/text/releases/tag/v0.9.0-rc5
• Migration guide from torchtext to show how to replace legacy code: https://github.com/pytorch/text/blob/master/examples/legacy_tutorial/migration_tutorial.ipynb

[cozek]

I did try to migrate it has have not been successful so far and haven't had time to go debug. Here is a link to the notebook
https://colab.research.google.com/drive/1LVyYWtSzY2Y65AO95gkG3Uo8rkiTRtA0?usp=sharing

I got stuck with the error:

ERROR:ignite.engine.engine.Engine:Engine run is terminating due to exception: Accuracy must have at least one example before it can be computed.
ERROR:ignite.engine.engine.Engine:Engine run is terminating due to exception: Accuracy must have at least one example before it can be computed.

IMO, perhaps we can keep the legacy notebook and add a fresh new one since it would require rewrite of the tutorial for the parts which used the legacy code.

[KickItLikeShika]

Thanks for the idea @cozek. @vfdev-5 what do you think about @cozek idea? he can send a PR to edit the name of existing TextCNN one to (for example TextCNNLegacy) and make a new notebook compatible with the new release of torchtext

[vfdev-5]

@cozek thanks for investigations. I think there is no value to keep legacy notebook as API changed.
In our notebook we can add a requirement to install newer torchtext package. Yes, I agree that it would require rewriting a bit the tutorial as well.

@KickItLikeShika could you please check the error in the colab that has @cozek ?
It should be certainly related to the input dataloaders...

[KickItLikeShika]

Thanks for your nice work @cozek! as @vfdev-5 said the issue is related to the dataloaders, I have edited the way we create the dataloader (if you want to keep the bucket iterators you have to edit the batch_sampler), please check this copy notebook: https://colab.research.google.com/drive/16_F6G_cBmp9RJRXwhRbof36pcWd7EZ44?usp=sharing .
This notebook solves the ERROR:ignite.engine.engine.Engine:Engine run is terminating due to exception: Accuracy must have at least one example before it can be computed. issue, but we need to edit the devices related code in the model (and other training and evaluation functions if needed) before sending a PR, because you will run into issues like Inputs and outputs are not on the same device while running this notebook on GPU.

[cozek]

@KickItLikeShika Thanks for your help. Please grant me access so that I can view the NB. :) I think I can fix the device related issues.

[KickItLikeShika]

@cozek check if you can access the NB now