iamvibhorsingh / novel-metric-to-calculate-similarity-between-original-and-wae-synthesized-image

This is the code implementation of our research paper in AICAI '19 Dubai.

Requirements

• Python 3.x (tested with Python 3.6)
• TensorFlow v1.x (tested with 1.8)
• tqdm (for the progress bar)

Doing it yourself

Training

Just run this command

python main.py

How to run Inference

Open the MNIST Plot.ipynb with Jupyter Notebook/Lab.

A pretrained model for MNIST is included in the repository here. Please download the zip file and decompress it on assets/pretrained_models/celeba/last*. Or, you can easily modify a path at the first cell on the notebook.

Results

MNIST

• Learning statistics
• Reconstruction Results (top): original images from MNIST validation set, (bottom): reconstructed image
• As can be seen in the MNIST jupyter notebook, our novel implementation of quasi-euclidean metric gives a similarity score of 1 to WAE-MMD synthesized images and the ground truth (original) image. It is justifiable since WAE-MMD do a very well job at synthesizing MNIST data especially since the images only have a few amount of pixels as well as not a lot of features that the model has to recognize.
• Random Sampled Images