Boolean Structured Autoencoder Convolutional Deep Learning Network (BSautoconvnet)
- Our model has only 44000+ parameters instead of millions parameters in U-net
- Use monotone Boolean algebra convolutional layers, so it does not overfit
- Achieved near perfect result of masked image reconstruction on CIFAR10 dataset
- Able to be trained on online using Google Colab with GPU
- No data augmentations, no regularization such as weight decay and dropout
Download the jupyter notebook
Open using Google Colab Colab It can also be run using a jupyter notebook
Follow the steps in the notebook, run each block of codes starting from the top to the bottom Model can be trained in 50 epoches.
If you do not want to train the model, you can load the pretrained model (1 megabytes only).
Input
Ground Truth
Reconstructured Output
Discrete Markov Random Field Relaxation
That's it. Have a Nice Day!!!