DIETNETWORK

Pytorch implementation of DietNetwork (https://arxiv.org/abs/1611.09340)

Training pipeline

Scripts

Main scripts

1. create_dataset.py : Create dataset and partition data into folds. The script takes snps.txt and labels.txt files as input to create dataset.npz and folds_indexes.npz
2. generate_embedding.py : Takes dataset.npz and folds_indexes.npz files created in the previous step and computes the embedding (genotypic frequency) of every fold. Embedding of each fold is saved in embedding.npz
   1. Missing values are -1 and are not included in the computation of genotypic frequencies embedding
   2. Embedding values are computed on train and valid sets
3. train.py : Whole training process. The data is divided in train/valid and test sets. Performance is reported on the test set.
   1. Data preprocessing of auxiliary net : Square Euclidean distance normalization
   2. Data preprocessing of discrim net: Missing values are replaced by the mean of the feature computed on training set. Data normalization (standardization) using mean and sd computed on training set.
4. test_external_dataset.py : Test model on an external set, ie on individuals that are not part of dataset.npz
5. evaluate.py : Utilities to visualize the model performance such as confusion matrix

Helper scripts

• dataset_utils.py : Data related functions (shuffle, partition, split, get_fold_data, replace_missing_values, normalize, ...)
• model.py : Model definition of feature embedding (auxiliary) and discriminative (main) networks.
• mainloop_utils.py : Function used in the training loop (get_predictions, compute_accuracy, eval_step, ...)
• log_utils.py : Utilities to save data (model summary and parameters, experiment parameters, predictions, etc.)
• test_utils.py : Utilities related to testing a trained model on an external set

Files

Raw files provided by user

• snps.txt : File of genotypes in additive encoding format and tab-separated.
• labels.txt : File of samples and their label.

Files created before training

• dataset.npz : Dataset created from the parsed snps.txt and labels.txt files.
• folds_indexes.npz : Array index (arrays are in dataset.npz) for each fold. The indexes are those of the data points to use as test.
• embedding.npz : Computed embeddings of each fold.

Files returned after training

• exp_params.log : Experiment parameters (fixed seed, learning rate, number of epochs, etc.)
• model_summary.log : Model information (number of hidden layers, number of neurons in each layers, activation functions, etc.)
• model_params.pt : Model parameters of final trained model
• model_predictions.npz: Scores and predictions returned by the trained model for test samples
• additional_data.npz : Some more information used at training time (mus and sigmas values used for normalization, feature names, label names, training samples ids, validation samples ids, etc.)

To do

• Embedding
• Data preprocessing : Missing values
• Data preprocessing : Data normalization
• Dataset class (for dataloader)
• Auxiliary and Main networks models
• Training loop
• Loss/Accuracy monitoring of train and valid
• Early stopping
• Test for in-sample data
• Test in-sample with missing values rates
• Test for out-of-sample data
• Save model params, results

Packages

• Python 3.6
• torch 1.5.0+cu101
• numpy 1.19.1
• pandas 1.1.0
• matplotlib 3.2.1
• captum 0.2.0 (https://captum.ai/)
• h5py 2.10.0