Skip to content

Variants of Multi-Perspective Convolutional Neural Networks

Notifications You must be signed in to change notification settings

tuzhucheng/MP-CNN-Variants

Repository files navigation

MP-CNN Variations

This is a PyTorch implementation of MP-CNN as a base model with modifications and additions such as attention and sparse features.

Here is the MP-CNN paper:

The datasets are available in https://git.uwaterloo.ca/jimmylin/Castor-data, as well as the GloVe word embeddings.

Directory layout should be like this:

├── MP-CNN-Variants
│   ├── README.md
│   ├── ...
├── Castor-data
│   ├── README.md
│   ├── ...
│   ├── msrvid/
│   ├── sick/
│   └── GloVe/

Note the original paper doesn't use dropout, so dropout=0 mimics this behaviour to allow for fair comparison in the results reported below.

To visualize the training process, just add --tensorboard to use TensorBoard.

SICK Dataset

To run MP-CNN on the SICK dataset mimicking the original paper as closely as possible, use the following command:

python main.py mpcnn.sick.model --dataset sick --epochs 19 --dropout 0 --lr 0.0005
Implementation and config Pearson's r Spearman's p MSE
Paper 0.8686 0.8047 0.2606
PyTorch using above config 0.8692 0.8145 0.2520

TrecQA Dataset

To run MP-CNN on TrecQA, you first need to run the get_trec_eval.sh script in utils.

Then, you can run:

python main.py mpcnn.trecqa.model --arch mpcnn --dataset trecqa --epochs 5 --holistic-filters 200 --lr 0.00018 --regularization 0.0006405 --dropout 0
Implementation and config map mrr
Paper 0.762 0.854
PyTorch using above config 0.774 0.836

The paper results are reported in Noise-Contrastive Estimation for Answer Selection with Deep Neural Networks.

WikiQA Dataset

You also need trec_eval for this dataset, similar to TrecQA.

Then, you can run:

python main.py mpcnn.wikiqa.model --arch mpcnn --dataset wikiqa --epochs 5 --holistic-filters 100 --lr 0.0001 --regularization 0.0002 --dropout 0
Implementation and config map mrr
Paper 0.693 0.709
PyTorch using above config 0.699 0.714

The paper results are reported in Noise-Contrastive Estimation for Answer Selection with Deep Neural Networks.

Other Datasets

MSRVID

To run MP-CNN on the MSRVID dataset, use the following command:

python main.py mpcnn.msrvid.model --dataset msrvid --batch-size 16 --lr 0.0005 --epsilon 1e-7 --epochs 32 --dropout 0 --regularization 0.001

You should be able to obtain Pearson's p to be 0.8980 (untuned), for reference the performance in the paper is 0.9090.

MSRP Dataset

To run MP-CNN on the MSRP dataset, use the following command:

python main.py mpcnn.msrp.model --dataset msrp --epochs 15

To see all options available, use

python main.py --help

Experimental

There are some scripts in this repo for hyperparameter optimization using watermill with some hacks since the library is in alpha. Hence, the imports in hyperparameter_tuning_{random,hyperband}.py and utils/hyperband.py will not work for you at the moment.

References

For results, please see my Master's thesis here:

@mastersthesis{tu2018experimental,
  title={An Experimental Analysis of Multi-Perspective Convolutional Neural Networks},
  author={Tu, Zhucheng},
  year={2018},
  school={University of Waterloo}
}