Light-Weighted CNN for Text Classification

Dataset Used

Tobacco-3482

Categories in the dataset are:

['ADVE', 'Email', 'Form', 'Letter', 'Memo', 'News', 'Note', 'Report', 'Resume', 'Scientific']

Train Model

To train on Optimized Text CNN

python ADAM_optmized_train.py

Lightweight Text CNN

python singleADAM_LW_train.py

Lightweight TextCNN with Dual Optimizer

Switches from Adam to SGD when a triggering condition is satisfied. python SWAT_LW_train.py

Optional arguments:

python train.py --help

Evaluate Model

To evaluate, run below command

python eval.py --eval_train --checkpoint_dir="./runs/trained_model/checkpoints/"

To use your own data, change the eval.py script to load your data.

Test Model

To prediction on new test data, make sure evaluate model is working , Then run below :

python test.py --out_test --checkpoint_dir="./runs/trained_model/checkpoints/" --test_dir="path to test data"

Link to the paper

For more details please go through my paper at link: https://arxiv.org/pdf/2004.07922.pdf

References :

1. https://github.com/dennybritz/cnn-text-classification-tf