MUFold-ss-pytorch

Introduction

MUFold-ss (https://arxiv.org/pdf/1709.06165.pdf) protein secondary structure prediction implementation in PyTorch. Note that the Conv11 in the final Struct2Struct network is not implemented.

Run python mode.py exp1 to train the network and validate after each epoch. My data is not shareable but you can follow the links provided in the paper to download the Cullpdb data. Modify data_loading.py to properly load that version data.

Dependency

1. Python 2.7
2. Pytorch 0.2.0
3. tensorboard-pytorch

To do

• Improve the argparse and file management in model.py.
  • Now max_seq_len for training data, learning rate, epochs, batch size, etc. can be controlled by command-line.
• Properly set up saving and restoring of the model.
  • Saving the best model based on validation accuracy for each model
• Try out different hyperparameters and optimizers.

Misc

Several bugs in PyTorch. Conv1d only accepts DoubleTensor and BatchNorm1d only accepts float.