README.md

Denoising Autoencoder as TensorFlow estimator

This repository contains an implementation of a (Denoising) Autoencoder using TensorFlow's Estimator and Dataset API. You can find a more detailed description in my blog post.

Two flavors of autoencoders are currently implemented

1. fully_connected: Uses fully-connected layers with 128, 64, and 32 units, respectively.
2. convolutional: Uses 2D convolutional layers with 3x3 kernels and 16, 8, and 8 filters, respectively.

Training

You can train an autoencoder on the MNIST dataset using train_mnist_autoencoder.py. The first argument specifies which autoencoder to train (fully_connected or convolutional), the remaining arguments are listed below.

Argument	Description
--model_dir	Output directory for model and training stats.
--data_dir	Directory to download the data to.
--noise_factor	Amount of noise to add to input (default: 0)
--learning_rate	Learning rate (default: 0.001)
--batch_size	Batch size (default: 256)
--epochs	Number of epochs to perform for training (default: 50)
--weight_decay	Amount of weight decay to apply (default: 1e-5)
--dropout	The probability that each element is kept in dropout layers (default: 1)
--save-images	Path to directory to store pairs of input and reconstructed images after each epoch (default: disabled)

If --noise_factor is non-zero, a denoising autoencoder will be trained.

Prediction

After training an autoencoder, you can inspect the model using test_mnist_autoencoder.py. You can visualize a model's predictions on the MNIST test dataset or save the output of the encoder to display it in TensorBoard. The first argument specifies which autoencoder to evaluate (fully_connected or convolutional), the remaining arguments are listed below.

Argument	Description
--model_dir	Output directory for model and training stats.
--data_dir	Directory to download the data to.
--noise_factor	Amount of noise to add to input (default: 0)
--dropout	The probability that each element is kept in dropout layers (default: 1)
--images	Number of test images to reconstruct or save embedding for (default: 10)
--what	How to visualize a trained model. reconstruction displays input and reconstructed images next to each other. embedding creates a checkpoint that can be used with TensorBoard to visualize the low-dimensional embedding of the encoder (default: reconstruction)

Make sure you pass the same values you used during training.