Reproducing the PredFull model

This repo contains code to reproduce the PredFull model as per its paper. The model provided by the authors in their repository (linked above) differs from the paper, so the provided model is used as the basis for this reproduction.

Preparations

pip install -r requirements.txt

Reproduction

• Download data from the PredFull website
• Run predfull_data_to_h5.py infile outfile to preprocess the mgf file and save it in hdf5 format
• [Optional] Download the pm.h5 pretrained model from the PredFull repository
• Set DATA, PRETRAINED, USE_PRETRAINED, RESET_PRETRAINED and FMT parameters in the predfull_keras.py script as desired as per the instructions in the script
• Run predfull_keras.py PHASE where PHASE is the training phase number. To reproduce the paper, that's python predfull_keras.py 0 && python predfull_keras.py 1 && python predfull_keras.py 2

Omitting the PHASE parameter just trains the model endlessly with a set learning rate, it's meant to be customized as needed or used as an example for something more serious.

Current Status

Model summary is identical between the model from the repository and this model, and training proceeds similarly. Some activation functions may differ. A script was provided by the paper's first author for reference for model creation, although it may not completely match the released model. It is available as build_model_example.py.