SpatioTemporalNormalisation

This repository consists of 3 jupyter notebooks that come together in order to display the optimal normalisation strategy for SpatioTemporal Graphs.

The conclusion from running our experiments were that the UnitedNorm technique that utilises a coalescence of LayerNorm, InstanceNorm, BatchNorm, GraphNorm and our novel TemporalNorm (transposition of existing normalisation strategies over the temporal layer, which employs a learnable shift) is most effective when dealing with SpatioTemporal Graphs when normalizing the inputs to smoothen out training and speed up convergence.

Run Instructions:

The experiments can be run by opening the jupyter notebooks in Google Colab and following these steps:

1. Save the SpatioTemporalNorm.ipynb and the SpatioTemporalNorm.ipynb files in Google Drive within the same directory.
2. Change the following line of code in 5th cell of SpatioTemporalNorm_Experiments.ipynb to point to the Google Drive directory that you chose in Step 1 : %cd "mnt/My Drive/Colab Notebooks/GDL/"
3. Run the SpatioTemporalNorm_Experiments.ipynb file.