keras-triplet-center-loss
A simple Keras implementation of Triplet-Center Loss on the MNIST dataset. As a reference in this repository also implementations of other two similar losses, Center-Loss and Triplet-Loss are included.
The Center-Loss implementation is from shamangary: https://github.com/shamangary/Keras-MNIST-center-loss-with-visualization
The Triplet-Loss implementation is from KinWaiCheuk: https://github.com/KinWaiCheuk/Triplet-net-keras
Triplet-Center Loss
Triplet-Center Loss has been introduced by He et al. in https://arxiv.org/abs/1803.06189. It is an "hybrid" loss between Center Loss and Triplet Loss that allows to maximise inter-class distance and minimize intra-class distance.
Details
In this repository a simple implementation on the MNSIT or alternatively Fashion MNIST is shown.
Running main.py will start sequentially 4 training routines with 4 different losses:
- Categorical Crossentropy only
- Center-loss + Categorical Crossentropy
- Triplet-loss + Categorical Crossentropy
- Triplet-Center loss + Categorical Crossentropy
In Folder runs there will be the results of those models, including Tensorboard summaries. Also T-SNE is run on the embeddings to visualize how the network internal representation changes as the loss is changed.
triplet-center loss, T-SNE on internal representation (Train Data):
Center loss, T-SNE on internal representation (Train Data):
Triplet loss, T-SNE on internal representation (Train Data):
As it can be seen the triplet-center loss maximises the inter-class distance as the Triplet Loss while keeping the Center-loss characteristic of minimizing intra-class distance. Another advantage of Triplet-Center loss is that it does not need advanced batching and triplet selection mining techniques as the Triplet-Loss does.