web199195 / CNN-text-from-scratch

Keras implementation of the paper 'Text Understanding from Scratch'

CNN-text-from-scratch

This is the keras implementation of the CNN model given in the paper 'Text Understanding from Scratch' (https://arxiv.org/pdf/1502.01710.pdf). Toy datasets are used currently, but the repo will be updated soon to use the original datasets used by the authors of the paper.

Implementation:

Python3, keras API with tensorflow as the backend

Run using:

python3 -W ignore main.py dataset_train.tsv dataset_test.tsv

Files:

1. main.py - Main file, contains code for model implementation, training and testing
2. data_handling.py - Auxillary file for dataset handling
3. dataset_train.tsv - TSV file with toy training data. The character sequences have been generated randomly and don't have any meaning.
4. dataset_test.tsv - TSV file with toy test data. The character sequences have been generated randomly and don't have any meaning.

In both the data files, the format used for each sample is: Character_sequence \TAB Class_label

Notes:

1. All output to console
2. The program generates a file called 'model_structure_new.png' (I've already placed it in this folder), in which it prints the entire structure of the ConvNet
3. The program generates a .hdf5 file in which it saves the best model during training