This dir contains the implementation of Diet Networks. The particular embedding we are going to use is the Histogram. The idea is to preprocess the data set by contructing a matrix s.t. each cell is the frequency of SNP k taking on value j for class i (where k is the SNP; j is 0,1,2; and i is 1,...,26). This is the embedding that is used for the auxilary networks which are used to construct the weight matrix in such a way that we greatly reduce the number of parameters. That is, we lean an embedding on the transpose of the data and construct a matrix that represents the first layer of the discrimative network (the net that makes predictions and optionally reconstructs using an autoencoder).

We assume you have {.panel file for labels, the 3 plinkfiles}. We also are working with Docker. (Dockerfile provided)

• Build environment image:

sudo docker build -t diet_code_env -f Dockerfile.gpu .

• launch the container and place add the volume with the data we are using:

sudo nvidia-docker run -it -p 81:6006 -v /path/to/4files:/usr/local/diet_code/1000G diet_code_env

• Preprocess the data. Assumes you have the files: affy_samples.20141118.panel, genotypes.bim/bed/fam in /usr/local/diet_code/1000G. In the container, we call:

 python preprocess.py

• We get these files:
  • hist3x26.npy: (px78) matrix of the freq of snp k taking on value j for class i. Built from the train/val set only.
  • train{}.npy: 75% of the remaining 80% of the data in the format where {} = X or Y for genomic data and labels resp..
  • valid{}.npy: 25% of the remaining 80% of the data in the format where {} = X or Y for genomic data and labels resp.
  • test.npy: 20% of the overall data (not used in constructing the histogram embedding.

• Grab the docker tf environment image if it doesnt build above:

docker pull ljstrnadiii/diet_code_env:0.1

• Then, run the preprocess described above.
• Finally, train the model like so:

sudo nvidia-docker run -it -p 81:6006 -v /path/to/this/repo:/usr/local/diet_code  diet_code_env

python train.py

• build the skeleton of the tensorflow model
• build numpy pipeline for data entry
• output the loss and scores to tensorboard