zeroQiaoba/ladder

Implementation of Ladder Network in PyTorch.

ladder

Implementation of Ladder Network and Stacked Denoising Autoencoder in PyTorch.

Requirements

• PyTorch

Training ladder

1. Run python utils/mnist_data.py to create the MNIST dataset.

2. Run the following command to train the ladder network:

• python ladder/ladder.py --batch 100 --epochs 20 --noise_std 0.2 --data_dir data

Status: The unsupervised loss starts at a high value because of which the network overfits the unsupervised loss and the supervised performance is bad. Current best accuracy on MNIST validation set using 3000 labelled and 47000 unlabelled examples: 98.33%.