Keras-MNIST-center-loss-with-visualization
Fig. (left) Softmax only. (right) Softmax with center loss
Update (2018/03/02)
Update (2017/11/10)
Remove the one-hot inputs for Embedding layer and replace it by single value labels.
There are two kinds labels: single value for center loss, and one-hot vector labels for softmax term.
Every classes are visually seperated now :)
How to run?
isCenterloss = True
#isCenterloss = False
KERAS_BACKEND=tensorflow python TYY_mnist.py
Dependencies
conda install -c anaconda scikit-learn
conda install -c conda-forge matplotlib
References: