Unsupervised Domain Translation with Alignment Guarantees

Code for 'Unsupervised Domain Translation with Alignment Guarantees'

Dependencies

tensorflow
skimage
matplotlib
python 3.5

Handwriting-to-Label Translation

Running the code

Here are commands to reproduce the results of full model / without shuffling / without pairwise discriminator

python main_mnist.py --data zipcode
python main_mnist.py --data zipcode --model nodop
python main_mnist.py --data zipcode --model nopd

Visualizing the Training Process

Here is the visualization of samples of G: Z -> X (sampled handwritten digits given the label, left) and confusion matrix of F: X -> Z (C_[i,j] is the probability that our model will predict digit j-1 when the true digit is i-1, diagonal is perfect, right) every 100 iters. There is a discrete optimization every 1000 iters.

The results is obtained using CPU (accuracy is 98.6% at iteration 12,500), the results of GPU will be a little different due to the randomness GPU introduced in.

State-to-Image Translation for RL Environments

To be coming