Welcome to the C-GAN / UNIT Project.
Multiple computer vision problems can be posed astransforming one input image into an output one (edge de-tection, image enhancement, colorization, restoration ...).As such, the Image to Image translation refers to the map-ping from one domain to another and it’s an active area ofresearch given the many possibilities that it has. For exam-ple, let’s consider super-resolution, which consists of up-sampling a low resolution image to a realistic higher resolu-tion one. We already know that going to a lesser resolutionis a trivial problem but upsampling requires inferring un-known details. Finding a framework that can perform sucha task would allow you to watch your favorite movie fromthe 70s in a higher quality or retrieve a picture from thatvery old phone you had 20 years ago and make it nice. The problem can be approached both as a supervised and unsupervised problem.
We chose to reproduce the results from two papers who approached the image-to-image translation problem from both settings : Image-to-Image Translation with Conditional Adversarial Networks and UNIT: UNsupervised Image-to-image Translation Networks
Make sure you have PyTorch installed on your virtual environment. This can be done by pasting the following command on your terminal:
python -c "import torch; print(torch.__version__)"
Create a file named env.py in the project root dir. Inside it set the following environment variables:
DISCRIMINATOR_PATH— the path from root directory to the discriminator .pth fileGENERATOR_PATH— the path from root directory to the generator .pth file
Create a directory containing the inputs in the project root dir.
Make the script executable the first time you run it:
chmod +x demo.py
-h— show this help message and exit-in— path to input file
-n— number of data to display, 1 by default--save— store the data on a specific folder, False by default