Diet Networks: Thin Parameters for Fat Genomics

Unofficial implementation of diet networks in TensorFlow.

Requirements:

• Python 2.7 or 3.5
• TensorFlow >= 1.1.0. See installation instructions.
• numpy
• plinkio (only for preprocessing) Install via pip e.g. pip install plinkio
• plink2 (only for preprocessing) This can be easily installed via Bioconda e.g. conda install -c bioconda plink2
• pandas (only for preprocessing)

Usage:

• Install the requirements above. Use plink2numpy to preprocess PLINK files and run dietnet script with --prefix option.
• To reproduce 1000G results, install dietnet e.g. python setup.py install and run make command in 1000G folder. (~700MB file will be downloaded) Finally, run ./dietnet train 1000G/genotypes

TODO:

• Dropout
• Bias terms for We and Wd
• Summary ops and tensorboard screenshots. also misclass. err
• K-fold CV
• Make train and predict subcommands e.g. add placeholders
• Other embeddings: random projection, histogram
• SNP2vec