A PyTorch implementation of this paper. It was implemented in Tensorflow1 here
Use train.py
for training where the parameters are taken from config.yaml
.
Another version of train.py
, train-fabric.py
, uses lightning Fabric with bf16-mixed
precision and accomplishes about 25% speedup on NVIDIA RTX3070.
The transform.py
transforms the gaze of an input image. The generator file and the input are required. The gaze is given as --angles H V
where H and V are the horizontal and vertical angles respectively. If no angles are specified the V is set to 0 and H in the set [-15,-10,-5,0,5,10,15]