Longitudinal Mixed Effects VAEGAN

Description

Here is a description a brief description about the data, model and purpose.

Training

python LVAEGAN.py param_file.txt

This executable script takes the experiment parameters from param_file.txt, creating the following project folders/files.

• Models/
• Latent Parameters/
• Logs/
• project_name_loss.txt

Parameter Name	Description
NAME	The name of the project.
PROJECT_DIR	Specify the directory to save the project in.
IMAGE_DIR	Specify the directory where the images are located.
SUBJ_DIR	Specify the directory where the subject information (.csv file) is located.
VERSION	Options are 1 or 2.
Z_DIM	The size of the size of the latent dimension. (Default 64)
MIXED_MODEL	If True then the IGLS estimator will be applied to the latent space. (Default False)
GAN	If True then a discriminator will be used to train the model (like a GAN). (Default False)
IGLS_ITERATIONS	The number of IGLS iterations used in the estimation. (Default 1)
SLOPE	This specifies whether or not to estimate the individual gradient, a1. (Default False).
INCLUDE_A01	This specifies whether or not to estimate the covariance, a01. (Default False).
SAVE_LATENT	If True then the latent variable z_ijk will be saved and overwritten every minibatch. (Default False)
USE_SAMPLER	If True or MIXED_MODEL is True then the custom sampler will be used forces each batch to have subjects with a number of time points specified by SAMPLER_PARAMS. (Default False)
SAMPLER_PARAMS	Two integer values defining boundaries for the number of time points sampled from each subject. (Default 4, 6)
MIN_DATA	This specifies the minimum number of time points that will be sampled from each subject. If a subject has less than this number of time points then they will not be included in training. (Default 4)
BATCH_SIZE	The size of the batch. (Default 100)
SHUFFLE_BATCHES	If True then the batches will be randomly sampled. (Default False)
H_FLIP	The probability of the image being horizontally flipped. (Default 0.)
V_FLIP	The probability of the image being vertically flipped. (Default 0.)
EPOCHS	The number of epochs to train for, starting from 0 if the project is new, or the latest save model. (Default 100)
LR	This is the learning rate of the main model. (Default 1e-5)
D_LR	When the GAN is true, the discriminator has it's own optimizer with it's own learning rate. (Default 1e-5).
RECON_LOSS	If True then the reconstruction loss between the input image and output image will be used in training (Default True)
ALIGN_LOSS	If True then the loss between z_ijk from the encoder and z_hat from the mixed effect model will be used in training. (Default False)
KL_LOSS	This will only work with VERSION 2. If True then use KL loss. (Default False)
D_LOSS	If True then the GAN discriminator will be trained. (Default False).
BETA	This parameter specifies the factor by which the KL loss will be multiplied to contribute to the total loss. (Not recommended. Default 1).
GAMMA	This parameter specifies the factor by which the align loss contributes to the total loss. (Default 1)
NU	This parameter specifies the factor by which the discriminator loss contributes to the total loss. (Default 1)

Plotting Reconstructions

python output_comparisons param_file.txt

Given a set of parameters, specifying a model and subject, a plot of the input image will be displayed adjacent to the reconstruction.

The following parameters are specified and can remain in the param_file.txt used for training without effecting the training. Many of the parameters used for training are also required for the visualisation of the reconstruction.

Parameter Name	Description
PLOT_EPOCH	When using this parameter file for evaluation, this is the number of epochs corresponding to the model being evaluation. (Default 100)
IMAGE_NO	The specific subject you want to visualise. (Default 0)
TIMEPOINT	The specific time point index from the subject to visualise (Default 0)
AXIS	As the image is 3D, specify the axis to view the 2D slice from. (Default 2)

Other Stuff to describe later.

Below is a description of the files and their intended method of running.

1. python testing_code_conv.py

   This code can be run as it is. Within the file you can change the number of epochs. It already has the path to save the model to and so will save the future models with the same name in the same directory but with the number of epochs. For example, D:\\models\\vae_1000.h5. You can change within the file after how many epochs the model should be saved.

2. python output_comparison.py $EPOCHS

   This code loads in the data, passes it through a trained model and then view it next to the ground truth. Assuming that the model directory and prefix are correctly specified in the file, the only argument that is required to be specified is the epoch number for the model you would like to use.

3. python get_latent.py $EPOCHS

   This code passes all the data through the model to get mu and log_var. Then it reparameterises and returns the latent vector z. For each subject this becomes a row and is saved as a csv. If specified then this latent space can be reduced to 2d and plotted to see the distribution.

4. python plot_loss.py loss_file.txt

   This code takes the file containing the losses through training and plots them. It assumes the loss file is formatted as train: train: loss_val\nval: loss_val\n

5. python extract_data.py

   Within the file, you specify the path for the images and the directory that you would like to save the new files. The files are reformatted from .gz to .pt, and you can specify whether to scale them down or not before saving.

6. python create_train_test.py

   This takes all data specified, splits these into k cross validation folds as specified within the file. It splits the validation in half to be used as validation and test. It then saves this as a csv.

7. python data_exploration.py

   A throwaway file that we used to understand how the data looked.