An Implementation of the Original Approach of ConvLSTM by Shi et. al.

Summary

A fully functional and explicitely followed implementation from Shi et. al's work on "Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting."

Note: While this nn.Module is fully functional, it's parameter count is quite high from repeated convolutional matrices, that could be simplified. This original approach shown in conv_lstm_shi-et-al.py is meant to be an explicitely followed implementation of that work. conv_lstm_efficient.py is the more efficient and practical implementation of ConvLSTMCell.

Models

• conv_lstm_shi-et-al.py contains the original implementation for a ConvLSTMCell
• conv_lstm_efficient.py contains the efficient one convolultional layer imlementation of ConvLSTMCell
• ConvLSTMLayer.py contains an implementation for a ConvLSTMLayer to create reccurency and processing of temporal data for a ConvLSTMCell

Example Implementation of ConvLSTMLayer

layer = ConvLSTMLayer(64,64)

Dependencies:

torch