(On-going) My implementation of PARTS: Unsupervised segmentation with slots, attention and independence maximization.
I'm still debugging the model (help is welcomed of course).
The model is currently significantly downscaled (model with the original setting is very large) and trained on the Sprites-MOT dataset from here which is an easier dataset than Clevrer.
Training can be unstable and inconsistent: sometimes representations will be disentangled, sometimes not; sometimes colors and shapes can be learned, sometimes not; sometimes the model will suddenly collapse to nonsense outputs. It is also very sensitive to beta. Below is trained from beta=0.3.
Ground truth
Reconstruction
