Introduction

This is the implementation of Kim's Convolutional Neural Networks for Sentence Classification paper in PyTorch.

1. Kim's implementation of the model in Theano: https://github.com/yoonkim/CNN_sentence

2. Denny Britz has an implementation in Tensorflow: https://github.com/dennybritz/cnn-text-classification-tf

3. Alexander Rakhlin's implementation in Keras; https://github.com/alexander-rakhlin/CNN-for-Sentence-Classification-in-Keras

Requirement

• python 3

• pytorch > 0.1

• torchtext > 0.1

• numpy

Result

I just tried two dataset, MR and SST.

I haven't adjusted the hyper-parameters for SST seriously.

Usage

./main.py -h

or

python3 main.py -h

You will get:

CNN text classificer

optional arguments:
  -h, --help            show this help message and exit
  -batch-size N         batch size for training [default: 50]
  -lr LR                initial learning rate [default: 0.01]
  -epochs N             number of epochs for train [default: 10]
  -dropout              the probability for dropout [default: 0.5]
  -max_norm MAX_NORM    l2 constraint of parameters
  -cpu                  disable the gpu
  -device DEVICE        device to use for iterate data
  -embed-dim EMBED_DIM
  -static               fix the embedding
  -kernel-sizes KERNEL_SIZES
                        Comma-separated kernel size to use for convolution
  -kernel-num KERNEL_NUM
                        number of each kind of kernel
  -class-num CLASS_NUM  number of class
  -shuffle              shuffle the data every epoch
  -num-workers NUM_WORKERS
                        how many subprocesses to use for data loading
                        [default: 0]
  -log-interval LOG_INTERVAL
                        how many batches to wait before logging training
                        status
  -test-interval TEST_INTERVAL
                        how many epochs to wait before testing
  -save-interval SAVE_INTERVAL
                        how many epochs to wait before saving
  -predict PREDICT      predict the sentence given
  -snapshot SNAPSHOT    filename of model snapshot [default: None]
  -save-dir SAVE_DIR    where to save the checkpoint

Train

./main.py

You will get:

Batch[100] - loss: 0.655424  acc: 59.3750%
Evaluation - loss: 0.672396  acc: 57.6923%(615/1066)

Test

If you has construct you test set, you make testing like:

/main.py -test -snapshot="./snapshot/2017-02-11_15-50-53/snapshot_steps1500.pt

The snapshot option means where your model load from. If you don't assign it, the model will start from scratch.

Predict

• Example1

   ./main.py -predict="Hello my dear , I love you so much ." 
             -snapshot="./snapshot/2017-02-11_15-50-53/snapshot_steps1500.pt"

  You will get:

   Loading model from [./snapshot/2017-02-11_15-50-53/snapshot_steps1500.pt]...
   
   [Text]  Hello my dear , I love you so much .
   [Label] positive

• Example2

   ./main.py -predict="You just make me so sad and I have to leave you ."
             -snapshot="./snapshot/2017-02-11_15-50-53/snapshot_steps1500.pt"

  You will get:

   Loading model from [./snapshot/2017-02-11_15-50-53/snapshot_steps1500.pt]...
   
   [Text]  You just make me so sad and I have to leave you .
   [Label] negative

Your text must be separated by space, even punctuation.And, your text should longer then the max kernel size.

Reference

• Convolutional Neural Networks for Sentence Classification