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 >= 3.8
• pytorch >= 2.0

No longer depends on torchtext or numpy. Data loading uses standard PyTorch Dataset and DataLoader.

Result

Two datasets were tested: MR and SST.

Dataset	Class Size	Best Result	Kim's Paper Result
MR	2	76.5%(CNN-rand)	76.1%(CNN-rand)
SST	5	45.6%(CNN-rand)	45.0%(CNN-rand)

Both results are consistent with Kim's paper. The SST result uses phrase-level training data, Adadelta optimizer, and embedding dimension 300, matching the original paper's setup:

python3 main.py -dataset SST -embed-dim 300 -batch-size 50 -optimizer adadelta -lr 1.0 -epochs 15 -early-stop 10000 -test-interval 500

Usage

python3 main.py -h

You will get:

CNN text classificer

optional arguments:
  -h, --help            show this help message and exit
  -lr LR                initial learning rate [default: 0.001]
  -epochs N             number of epochs for train [default: 256]
  -batch-size N         batch size for training [default: 64]
  -log-interval N       how many steps to wait before logging training status [default: 1]
  -test-interval N      how many steps to wait before testing [default: 100]
  -save-interval N      how many steps to wait before saving [default: 500]
  -save-dir DIR         where to save the snapshot
  -early-stop N         iteration numbers to stop without performance increasing [default: 1000]
  -save-best BOOL       whether to save when get best performance [default: True]
  -dataset DATASET      dataset to use: MR or SST [default: MR]
  -no-phrases           SST: use sentence-level only (no phrase data)
  -shuffle              shuffle the data every epoch
  -dropout DROPOUT      the probability for dropout [default: 0.5]
  -max-norm FLOAT       l2 constraint of parameters [default: 3.0]
  -embed-dim N          number of embedding dimension [default: 128]
  -kernel-num N         number of each kind of kernel [default: 100]
  -kernel-sizes STR     comma-separated kernel size to use for convolution [default: 3,4,5]
  -static               fix the embedding
  -optimizer OPTIMIZER  optimizer: adam or adadelta [default: adam]
  -device DEVICE        device to use for iterate data, -1 mean cpu [default: -1]
  -no-cuda              disable the gpu
  -snapshot FILE        filename of model snapshot [default: None]
  -predict TEXT         predict the sentence given
  -test                 train or test

Train

python3 main.py

You will get:

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

Test

python3 main.py -test -snapshot="./snapshot/2017-02-11_15-50-53/best_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

   python3 main.py -predict="Hello my dear , I love you so much ." \
                   -snapshot="./snapshot/2017-02-11_15-50-53/best_steps1500.pt"
  

  You will get:

   Loading model from ./snapshot/2017-02-11_15-50-53/best_steps1500.pt...
  
   [Text]  Hello my dear , I love you so much .
   [Label] positive
  

• Example2

   python3 main.py -predict="You just make me so sad and I have to leave you ."\
                   -snapshot="./snapshot/2017-02-11_15-50-53/best_steps1500.pt"
  

  You will get:

   Loading model from ./snapshot/2017-02-11_15-50-53/best_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 be longer than the max kernel size.

Changes from Original

• Removed dependency on deprecated torchtext.data.Field, BucketIterator, Example APIs
• Replaced with standard PyTorch Dataset, DataLoader, and custom BucketSampler
• Fixed feature.data.t_() / target.data.sub_(1) RuntimeError on modern PyTorch
• Fixed loss.data[0] IndexError (use loss.item() instead)
• Fixed size_average deprecation (use reduction='sum')
• Fixed early stopping (now actually stops training)
• Added L2 weight constraint on fc layer (matching Kim's original implementation)
• Added SST dataset support with phrase-level training data
• Added Adadelta optimizer option (-optimizer adadelta)
• Embedding initialization with uniform[-0.25, 0.25] (matching Kim's paper)
• Added torch.no_grad() context for evaluation and prediction
• Compatible with PyTorch 2.0+

Reference

• Convolutional Neural Networks for Sentence Classification