Implementation and enhancement of the model "Learning Graph Neural Networks for Image Style Transfer"

Description:

This project is an simple Pytorch implementation of the paper "Learning Graph Neural Networks for Image Style Transfer"

Architecture:

Re-implementation:

Blocks of code that we have implemented :

• Encoder : VGG19
• Image2patch
• Knn graph constructor
• GATv2
• Patch2feat
• AdaIN
• Pyramid feature

Improvements:

• Using GATv2 instead of GAT since the graph structure is similar to graph structure of “SYNTHETIC BENCHMARK: DICTIONARY LOOKUP” in GATv2.
• We define a threshold to decide to construct an edge or not instead of KNN.
• Improving Patch2Feat with Feature Pyramid network.

Results: NOT THE DESIRED OUTPUT

Conclusion:

• The operation proposed required a tremendous amount of GPU mem
• Even though our implementation didn’t get the desired results, We think that if we have enough GPU for smaller patch stride the output would be better.

Future works:

• Invest on GPU power
• Finding improvement in order to reduce the memory requirements (different encoder).
• Implement deformable graph convolutional operation