trevor-richardson/rnn_zoo

This repository tests various recurrent neural network architectures on baseline datasets SeqMNIST and pMNIST.

rnn_zoo

This repository tests various recurrent neural network architectures on baseline datasets SeqMNIST and pMNIST.
The network architectures chosen were those deemed to be the most effective currently available.

Architectures tested include:

• RNN

  

• LSTM - http://www.bioinf.jku.at/publications/older/2604.pdf

• GRU - https://arxiv.org/pdf/1406.1078v3.pdf

    

• IRNN - https://arxiv.org/abs/1504.00941

    

• Peephole LSTM - https://ieeexplore.ieee.org/document/861302/

  

• UGRNN - https://arxiv.org/pdf/1611.09913.pdf

    

• Intersection RNN - https://arxiv.org/pdf/1611.09913.pdf

    

Results

The following results are were generated using the architectures listed above.
Hyperparameters used: layers 3, num neurons 50, optimizer Adam, learning rate .0001 and batch 64.

---

---

---

Running the code

  python train.py --model-type=irnn --task=seqmnist --layers=2 --batch-size=64 --epochs=10

Installing

Update BASE_DIR in config.ini with the absolute path to the current directory.
Packages needed to run the code include:

• numpy
• python
• PyToch
• argparse
• configparser