This is (an intentionally) simple implementation of constrained de-noising auto-encoder. And auto-encoder is an unsupervised learning model, which takes some input, runs it though "encoder" part to get encodings of the input. Then it attempts to reconstruct original input based only on obtained encodings.
More details and explanations can be found here: DeepLearning@Home
The idea is that encodings will encode the most important information in the data.
File | Description |
---|---|
autoencoder.py | contains AutoEncoder class. Is not application specific, multilayer architecutre. |
autoencider_use_mnist.py | Example of using the AutoEncoder for MNIST image compression/representation |
utils.py | different helper functions |
Example run:
python ~/repos/autoencoder/autoencoder_use_mnist.py --encoder_network=784,128,10 --noise_level=0.0 --batch_size=64 --num_epochs=60 --logdir=LD_784_128_10_N0
Then start Tensorboard:
tensorboard --logdir=LD_784_128_10_N0