Refactor training and evaluation with Lightning

[philgzl]

Changes include:

• Training routines are now methods of the model to comply with Lightning and. Training script now uses Lightning.
• Moved model evaluation to model GCN.evaluate methods. Evaluation method returns a float equal to accuracy over entire test dataset. Also added a new prediction method GCN.predict to predict a single observation. The prediction method takes as argument the item index in the dataset and returns a string equal to the paper category.
• Added and updated tests.
• Updated README.md. Renamed from readme.md to README.md.
• Renamed scripts to shorter names. Example: src/models/train_model.py -> src/models/train.py
• Renamed model checkpoint form trained_model.pt to checkpoint.pt
• Fixed weight decay from config file not being used as the optimizer was initialized twice with no arguments the second time. The weight decay was changed from 5e4 to 0 in the config file since 5e4 was extremely high and resulted in weights not updating during training.
• Changed the number of epochs from 100 to 250. 100 was not leading to convergence.
• Added validation loss monitoring. The output model from the training script is now the model with the best validation loss. Training curves now look like this:

[codecov-commenter]

Codecov Report

:exclamation: No coverage uploaded for pull request base (main@6b5663d). Click here to learn what that means. The diff coverage is n/a.

@@           Coverage Diff           @@
##             main      #39   +/-   ##
=======================================
  Coverage        ?   28.57%           
=======================================
  Files           ?        6           
  Lines           ?      161           
  Branches        ?        0           
=======================================
  Hits            ?       46           
  Misses          ?      115           
  Partials        ?        0           

:mega: We’re building smart automated test selection to slash your CI/CD build times. Learn more