76 ptd nlp classifier template

[AgatheZ]

Linked Issue(s)

#76

Summary of changes

I have implemented the code for the NLP template, as a standard multi class classifier leveraging a Bert based transformer.

This includes:
training script
datamodule
dataset
network
wrapper
configs

I created fake data (in tests/data/)

Key features I introduced:
KFOLD cross validation possibility (can be enabled/disabled in the config file)
Save the best model (can be enabled/disabled in the config file)
f1 score and other metrics are logged
A data manifest is logged as a mlflow artifact, logging the classes distributions in the training and validation test sets)

FYI: I had to change the Dockerfile for the default folder to be project_NLP, might incur some clashes in the future.

I have tested it on the DGX, it runs properly and nicely

[AgatheZ]

@heyhaleema Don't hesitate if you have any questions re. how to leverage this piece of code in a more elegant manner (e.g. Factory depending on the type of project)

[heyhaleema]

@heyhaleema Don't hesitate if you have any questions re. how to leverage this piece of code in a more elegant manner (e.g. Factory depending on the type of project)

@AgatheZ Firstly - this is amazing! 🔥 One (initial) question: is NLPDataset like a preprocessing step before NLPDataModule?

Also, I've just realised that since we run mlops run scripts/train.py -c config/config.cfg, some of the changes (related to #86) I've made to train.py (main.py in my local branch) might be better placed in the csc-mlops package e.g., mlops run local imaging, mlops run dgx nlp, etc. Having said that, not yet sure about this approach either...

@mikewoodward94 If you have some time for a quick chat about this next week, that'd be great 🙏🏼 My thought is that I can then have the diagram for #86 be the proposed updated workflow, and potentially have this used to update this diagram in the csc-mlops repo or have it added as another, separate asset.

[AgatheZ]

@heyhaleema NLPDataset is about wrapping the data and labels in a PyTorch Dataset (more info here). This step is also performed for image-based projects (but I think they do it in the Datamodule class instead of having a dedicated Datset class)

About your second question, we could run mlops run scripts/train.py -c config/config_image.cfg or mlops run scripts/train.py -c config/config_NLP.cfg and only modify the config files. This would not require any update from the csc-mlops package.