Implementation of Spatial Contrasting Network in Keras. Included are three models:
-
Baseline Convolutional Neural Network
- Uses the same architecture reported in paper
- Able to replicate results of ~73% accuracy when using 4000 training examples
-
Spatial Contrasting Network
- Trained to embed patches from same image closer in deep space than patches from other images
-
Pre-trained Convolutional Neural Network
- Same architecture as baseline conv-net except early layers are initialized with weights learned from Spatial Contrasting Network
- Performance falls short of baseline conv-net
Images of each network and the weights from their first layers are in /saved (/models and /images sub-directories, respectively).
Please help me reproduce the results from the paper!