The generative model is based on VAE using TensorFlow. Scripts are at /projects/piml_inversion/DCVD/. Everything for training the model is in vae.py, including data loading and processing, model structure, training, and saving model. Python script to reconstruct is vae_recon.py, and to sample from latent space is vae_gen.py. The training script with style loss is vae_style.py (Not implemented yet)

03.10.2021 Add python script to reconstruct and sample from latent space. Add number of iterations to 1,000, still not converge.

• Python 3
• Numpy
• PIL
• TensorFlow 1.15 (tested) or latest TensorFlow 2 (not tested)

To train VAE on Darwin using TensorFlow, using Power nodes is suggested, since GPU cannot be used when using x86 nodes with GPU. Creating environment for Power architecture in Anaconda: https://www.ibm.com/support/knowledgecenter/SS5SF7_1.6.0/navigation/pai_install.htm. Shell script for submitting jobs is vae.sh. Please edit the scripts accordingly, like file name, output files, job name, data folder and so on. Before submitting a job to Power node, please load anaconda module and activate the environment for power architecture.

The model using current hyperparameters cannot converge! Maybe more iterations or smaller learning rate is needed, but may lead to overfitting.