• Pytorch Implementation of the algorithm.
• This repository includes a new proposed method for instance-based label smoothing in neural networks, where the target probability distribution is not uniformly distributed among incorrect classes. Instead, each incorrect class is going to be assigned a target probability that is proportional to the output score of this particular class relative to all the remaining classes for a network trained with vanilla cross-entropy loss on the hard target labels.

• The following figure summarizes the idea of our instance-based label smoothing that aims to keep the information about classes similarity structure while training using label smoothing.

• Python 3.x
• pandas
• numpy
• pytorch

• CIFAR10 / CIFAR100 / FashionMNIST

The project have a structure as below:

├── Vanilla-cross-entropy.py
├── Label-smoothing.py
├── Instance-based-smoothing.py
├── Models-evaluation.py
├── Network-distillation.py
├── utils
│   ├── data_loader.py
│   ├── utils.py
│   ├── evaluate.py
│   ├── params.json
├── models
│   ├── resnet.py
│   ├── densenet.py
│   ├── inception.py
│   ├── shallownet.py

Vanilla-cross-entropy.py is the file used for training the networks using cross-entropy without label smoothing.
Label-smoothing.py is the file used for training the networks using cross-entropy with standard label smoothing.
Instance-based-smoothing.py is the file used for training the networks using cross-entropy with instance-based label smoothing.
Models-evaluation.py is the file used for evaluation of the trained networks.
Network-distillation.py is the file used for distillation of trained networks into a shallow convolutional network of 5 layers.
models/ includes all the implementations of the different architectures used in our evaluation like ResNet, DenseNet, Inception-V4. Also, the shallow-cnn student network used in distillation experiments.
utils/ includes all utilities functions required for the different models training and evaluation.

Example

python Instance-based-smoothing.py --dataset cifar10 --model resnet18 --num_classes 10

--lr type = float, default = 0.1, help = Starting learning rate (A weight decay of $1e^{-4}$ is used).
--tr_size type = float, default = 0.8, help = Size of training set split out of the whole training set (0.2 for validation).
--batch_size type = int, default = 512, help = Batch size of mini-batch training process.
--epochs type = int, default = 100, help = Number of training epochs.
--estop type = int, default = 10, help = Number of epochs without loss improvement leading to early stopping.
--ece_bins type = int, default = 10, help = Number of bins for expected calibration error calculation.
--dataset, type=str, help=Name of dataset to be used (cifar10/cifar100/fashionmnist).
--num_classes type = int, default = 10, help = Number of classes in the dataset.
--model, type=str, help=Name of the model to be trained. eg: resnet18 / resnet50 / inceptionv4 / densetnet (works for FashionMNIST only).

• Results of the comparison of different methods on 3 datasets using 4 different architectures are reported in the following table.
  
• The experiments were repeated 3 times, and average $\pm$ stdev of log loss, expected calibration error (ECE), accuracy, distilled student network accuracy and distilled student log loss metrics are reported.

• A t-sne visualization for the logits of 3-different classes in CIFAR-10 can be shown below: