Grad-CAM for Text

This is implementation of Grad-CAM for text classification model. The model used is 1D-CNN for text classification which was introduced in Yoon Kim (2014). The data used are re-fined version of Stanford Sentimental Treebank which re-labeled for binary classification. Input feature is slim version of word2vec.

It especially requires python>=3, tensorflow>=1.4,<2.0 (requires tf.data.Dataset and tf.estimator.Estimator, but not support TF2).

Installation

>>> pip3 install -r requirements.txt

Feature

Before training, it requires word2vec binary file. Download by word2vec.sh, it will download in directory word2vec/.

>>> ./word2vec.sh

All word2vec binary have to be located on word2vec/ directory.

• word2vec/GoogleNews-vectors-negative300-SLIM.bin

Stanford Sentimental Treebank Dataset will be download automatically when training.

Training

usage: train.py [-h] [--epoch EPOCH] [--batch-size BATCH_SIZE]
                [--learning-rate LEARNING_RATE] [--gpu-index GPU_INDEX]

optional arguments:
  -h, --help            show this help message and exit
  --epoch EPOCH         Train Epoch
  --batch-size BATCH_SIZE
                        Train Batch Size
  --learning-rate LEARNING_RATE
                        Train Learning Rate
  --gpu-index GPU_INDEX
                        GPU Index Number

Visulization

It shows which word impacts the result with heatmap of sentence. Strong impact factor gets brighter color, otherwise darker color.

Fill CKPT variable with your ckeckpoint path in visualize.ipynb.