karanchahal / distiller

A large scale study of Knowledge Distillation.

Distiller

A clean Pytorch implementation to run quick distillation experiments. Our findings are available for reading in our paper "The State of Knowledge Distillation for Classification" linked here.

Python Dependencies

This codebase only supports Python 3.6+.

Required Python packages:

• torch torchvision tqdm numpy pandas seaborn

All packages can be installed using pip3 install --user -r requirements.txt.

This project is also integerated with Pytorch Lightning. Use the lightning branch to see Pytorch Lightning compatible code.

Run

The benchmarks can be run via python3 evaluate_kd.py and providing the respective command line parameters. For example:

python3 evaluate_kd.py --epochs 200 --teacher resnet18 --student resnet8 --dataset cifar10 --teacher-checkpoint pretrained/resnet18_cifar10_95260_parallel.pth --mode nokd kd

Runs basic student training and knowledge distillation for 200 epochs using a pretrained teacher. There are checkpoints of multiple models in the pretrained folder.

Supported distillation modes

Distillation Technique	Mode	Description
Baseline Accuracy	nokd	Plain training with no knowledge distillation.
Baseline Knowledge Distillation Accuracy	kd	Hinton loss to distill a student network.
Ensemble of Teachers	allkd	Distill from a list of teacher models and pick the best performing one.
Hyper parameter tuning of Hinton Loss	kdparam	Distill using varying combinations of temperature and alpha and pick the best performing combination.
Triplet Loss	triplet	Knowledge Distillation with a triplet loss using the student as negative example.
Ensemble of Students	multikd	Train a student under an ensemble of students that are picked from a list.
Unsupervised Data Augmentation Loss	uda	Run knowledge distillation in combination with unsupervised data augmentation.
Teacher Assistant Knowledge Distillation	takd	Run distillation using Teacher-Assistant distillation.
Activation Boundary Distillation	ab	Run feature distillation using Activation-Boundary distillation.
Overhaul Distillation	oh	Run feature distillation using the Feature Overhaul distillation.
Relational Knowledge Distillation	rkd	Run distillation using the Relational Knowledge distillation.
Patient Knowledge Distillation	pkd	Run feature distillation using the Patient Knowledge distillation.
Simple Knowledge Distillation	sfd	Runs a custom feature distillation distillation (Simple Feature Distillation) that just pools and flattens feature layers.

Results