Modification of SoftMax layer to distribute features on a hypersphere

Tutorial: Medium Story

VGG16 Training using MNIST dataset to show the difference in feature PCA visualization between SoftMax and modified SoftMax.

Versions for Tensorflow 1.14 and 2.0

Contents:

1. SoftMax training
2. modified SoftMax training
3. fine-tuning of SoftMax with modified SoftMax
4. Tensorboard visualizations

Tensorflow 1.14:

python3 mnist.py

Visualization of scalars and features using: tensorboard --logdir path/to/logs_mnist

Tensorflow 2.0:

python3 mnist.py

Visualization of scalars using: tensorboard --logdir path/to/logs_mnist

For feature visualization upload the two .tsv files generated under "logs_mnist" directory to Embedding Projector

2D Feature visualization(PCA) with SoftMax training	2D Feature visualization(PCA) with modified SoftMax training

Resources

• ArcFace: Additive Angular Margin Loss for Deep Face Recognition paper
• Official implementation of deepinsight/insightface on github