LSTM Autoencoder and LSTM Future Predictor in Tensorflow.

This is a simple implementation based on this paper : https://arxiv.org/abs/1502.04681

Requirement

1. Tensorflow 1.4.0
2. Python 3.5.4
3. Python packages : numpy, matplotlib, os, argparse, scipy

Usage

Implement reconstruction or future prediction after data generation

• Data generation using rotation and shift : Image_generation.ipynb
  • Reconstruct data using LSTM autoencoder : Autoencoder.ipynb
  • Input Reconstruction and Future Prediction : AE_with_Predictor

Results

From Composite Model

Data sequence(left to right) : 0,1,...,t-1,t

Rotated Data

After entering sequence data in LSTM model, reconstruct the data and predict how much rotate

Input Sequence

Input Reconstruction

Future Prediction

Shifted Data

After entering sequence data in LSTM model, reconstruct the data and predict how much shift

Input Sequence

Input Reconstruction

Future Prediction

Reference Implementations

• https://github.com/iwyoo/LSTM-autoencoder
• https://github.com/peytonhong/LSTMAutoencoder
• https://arxiv.org/abs/1502.04681