Deep Learning Model Evaluation as a Service

Inspiration

More and more companies start to use deep learning models in their business. Evaluating deep learning models is a challenge.

What it does

We want to help companies with limited resource and time constraint to evaluate their machine learning models.

Business flow

After companies provide us with their their trained model and sample data, we help them to expand data set by collecting data and using GAN to generate more data sets. Then we run inference and evaluate the quality of the model.

User experience

We provide a easy-to-use platform for companies to

provide/edit inference API and datasource
view results of evaluation

What we built

For this hackathon, we are tackling the most difficult part of this business - using GAN to expand data sets. we prototyped serval types of image translation GANs:

CycleGan(https://github.com/junyanz/CycleGAN): Torch implementation for learning an image-to-image translation (i.e. pix2pix) without input-output pairs

pix2pix(https://phillipi.github.io/pix2pix/): Image-to-Image Translation with Conditional Adversarial Nets

StarGAN(https://github.com/yunjey/stargan): StarGAN can flexibly translate an input image to any desired target domain using only a single generator and a discriminator