GAN-based Training of Semi-Interpretable Generators for Biological Data Interpolation and Augmentation

authors: Anastasios Tsourtis, Georgios Papoutsoglou, Yannis Pantazis

Paper: https://doi.org/10.3390/app12115434

This repository hosts the official TensorFlow implementation of our paper, published in MDPI.

Environment Setup

We assume that conda is already installed on your system. After cloning the repository run setup.sh script (while sourcing it):

. ./setup.sh

It will create and activate a python 3.9 environment called tf2_env_py39 using conda and will install all required packages.

Please refer to https://www.tensorflow.org/install/pip#linux_1 for more information on how to set up TensorFlow with GPU acceleration. GPU acceleration is preferable but not mandatory.

Input data

Please find the data corresponding to each experiment under \input_data . For further information refer to the published paper and data_loading.py

Training

python main_example_name.py --help
usage: main_example_name.py [-h] [--data_fname] [--steps] [--d] [--mb] [--alpha] 
                             [--K] [--K_lip] [--lam_gp] [--Z_dim] [--y_dim] [--spen]
                             [--lr]  [--saved_model] [--output_fname] [--resume_from_iter]
                             [--missing_labels] [--generate]


optional arguments:
  -h, --help            show this help message and exit
  --data_fname          directory of input data
  --steps               number of training steps      
  --d                   number of dimensions of input data
  --mb                  mini-batch size (default: 1024)
  --alpha               alpha parameter of cumulant loss (default: 0.5)
  --K                   number of GMM modes (default: 30)
  --K_lip               Lipschitz constant K (default: 1.0)
  --lam_gp              gradient penalty coefficient (default: 1.0)
  --Z_dim               noise dimension of generator
  --y_dim               dimension of label embedding (default: 10)
  --spen                use Sigma Penalty on generator (default: 0.001) 
  --lr                  learning rate (default: 0.0002)
  --saved_model         name of the saved model checkpoints
  --output_fname        name of the output file directory, for this experiment
  --resume_from_iter    steps corresponding to last checkpoint, needed to resume training
  --missing_labels      Missing labeled data in the training set. Options: 'none' (default), '0.4_0.6', 'state_2'
  --generate            Generate samples, provided training has taken place (default: False)

• Example 1: Train Lipschitz GAN using a 2-D SwissRoll dataset

We provide two options for the generator network:

1. Train using the conditional feedforward neural network (cFNN) generator for 100K steps:

python main_swiss_LIP.py  --steps 100000

2. Train using the conditional Gaussian Mixture Model (cGMM) generator and sigma penalty coefficient to be 0.002:

python main_swiss_LIP_GMM.py -spen 0.002

The experiment parameters are saved in \output_files\experiment_name\commandline_args.txt.

• Example 2: Synthetic RNA-seq data

1. Train using the conditional feedforward neural network (cFNN) generator with gating on the output layer. Training data labels lie uniformly in [0, 1]:

python main_synth_data_LIP.py --steps 200000 --missing_labels none

2. Train using the conditional zero-inflated Mixture Model (cZIMM) generator, where interval [0.4:0.6] is missing from the training data set:

python main_synth_data_LIP_ZIMM.py --missing_labels 0.4_0.6

• Example 3: Real single-cell mass cytometry data

1. Train using the conditional feedforward neural network (cFNN) generator with gating on the output layer. Here we continue training from a previous run of 100K steps for another 200K steps:

python main_real_data_LIP.py --steps 300000 --resume_from_iter 100000

2. Train using the conditional zero-inflated Mixture Model (cZIMM) generator:

python main_real_data_LIP_ZIMM.py 

Inference

During training, generated data, plots and running losses are being exported in the corresponding subdirectories under \output_files\experiment_name\. Assuming training has already taken place (checkpoint files containing model weights have been created), we can further generate data by:

python main_example_name.py --generate

The generated data are:

• .csv file containing generated points (rows: samples, columns: dimensions)
• corresponding plots (similar to the ones during training)

Further training can be carried out as shown in the previous section.

Questions

Please don't hesitate to send me an email.