An efficient RNN-inspired sequence modelling architecture called SparseRNN, which employs dilation to reduce the cost of dense-recurrence without compromising on performance.
This was made as the final course project for the Deep Learning[DL2023] course at Indian Institute of Technology, Jodhpur.
The research report(link) contains in-depth explanations of our research and the results obtained on three benchmarks.
Refer to the demo video(link) for a barebones presentation and commentary on our work.
- Python >= 3.6
- PyTorch >= 1.9
- Torchvision >= 0.10
- wandb >= 0.12.1
- cuda >= 10.2
- part-of-speech tagging. Datasets used-
- Penn Treebank
- conll2000
- brown
- Sentiment Analysis. Dataset used-
- MDB Dataset of 50K Movie Reviews
- SparseRNN
- LSTM
- RNN
To run a model on for a task, run the following command:
python main.py --yaml <path to yaml file in config folder>
For example, to run sparsernn for sentiment analysis, run the command:
python main.py --yaml ./config/sentiment/sparsernn.yaml