This repo is the original code of a Kaggle notebook for SMS Spam Collection Dataset.
What you can additionally do on this code;
- manage arguments, callbacks, and other various features of pl.Trainer with a single yaml file
- configure all the hyperparameters of an experiment with a single yaml file
- check the past config through logs of hydra
- trace model training by wandb
| F1 score | Accuracy | |
|---|---|---|
| ham | 99.77 | - |
| spam | 98.65 | - |
| total | 99.61 | 99.61 |
The model is tested on 517 samples selected as 10% of the entire dataset. The validation dataset is randomly sampled each time the training runs and it does not involve in the model's learning.
As an objective loss, I adopted focal loss to deal with positive/negative imbalance. Instead of following the paper, I implemented it in the form of multi-class classification, which makes training more stable. In addition, for better stability of training, I adopted regularization by AdamW, warmup start, and linear-decreasing lr scheduler.
To train the model on this repo, run the command line;
python train.py —config-name exp_0
If you want to test the checkpoints, enter checkpoints file path on exp_0.yaml and run the command line;
python test.py —config-name exp_0
BSD 3-Clause License Copyright (c) 2022