The Pytorch Implementation of L-Softmax

this repository contains a new, clean and enhanced pytorch implementation of L-Softmax proposed in the following paper:

Large-Margin Softmax Loss for Convolutional Neural Networks By Weiyang Liu, Yandong Wen, Zhiding Yu, Meng Yang [pdf in arxiv] [original CAFFE code by authors]

L-Softmax proposes a modified softmax classification method to increase the inter-class separability and intra-class compactness.

this re-implementation is based on the earlier pytorch implementation here by jihunchoi and borrowing some ideas from its TensorFlow implementation here by auroua. Generally the improvements are as follows:

• Now features visualization as depicted in the original paper using the vis argument in the code.
• Cleaner and more readable code
• More comments in lsoftmax.py file for future readers
• Variable names are now in better correspondence with the original paper
• Using the updated PyTorch 0.4.1 syntax and API
• Two models to produce visualization in paper's fig 2 and the original MNIST model is provided
• The lambda (beta variable in code) optimization missing in the earlier PyTorch code has been added (refer to section 5.1 in the original paper)
• The numerical error of torch.acos has been addressed
• Provided training logs in the Logs folder
• Some other minor performance improvements

Version compatibility

This code has been tested in Ubuntu 18.04 LTS using PyCharm IDE and a NVIDIA 1080Ti GPU. Here is a list of libraries and their corresponding versions:

python = 3.6
pytorch = 0.4.1
torchvision = 0.2.1
matplotlib = 2.2.2
numpy = 1.14.3
scipy = 1.1.0

Network parameters

• batch_size = 256
• max epochs = 100
• learning rate = 0.1 (0.01 at epoch 50 and 0.001 at epoch 65)
• SGD with momentum = 0.9
• weight_decay = 0.0005

Results

Here are the test set visualization results of training the MNIST for different margins:

• this plot has been generated using the smaller network proposed in the paper for visualization purposes only with batch size = 64, constant learning rate = 0.01 for 10 epochs, and no weight decay regularization.

And here is the tabulated results of training MNIST with the proposed network in the paper:

• the test accuracy values are the max test accuracy of running the code only once with the network parameters above!