Revisiting Distillation and Incremental Classifier Learning

Accepted at ACCV18. Pre-print is available at : http://arxiv.org/abs/1807.02802

Citing the paper :

@inproceedings{javed2018revisiting,
  title={Revisiting distillation and incremental classifier learning},
  author={Javed, Khurram and Shafait, Faisal},
  booktitle={Asian Conference on Computer Vision},
  pages={3--17},
  year={2018},
  organization={Springer}
}

Interface to Run Experiments

usage: runExperiment.py [-h] [--batch-size N] [--lr LR]
                        [--schedule SCHEDULE [SCHEDULE ...]]
                        [--gammas GAMMAS [GAMMAS ...]] [--momentum M]
                        [--no-cuda] [--random-init] [--no-distill]
                        [--distill-only-exemplars] [--no-random]
                        [--no-herding] [--seeds SEEDS [SEEDS ...]]
                        [--log-interval N] [--model-type MODEL_TYPE]
                        [--name NAME] [--outputDir OUTPUTDIR] [--upsampling]
                        [--pp] [--distill-step] [--hs]
                        [--unstructured-size UNSTRUCTURED_SIZE]
                        [--alphas ALPHAS [ALPHAS ...]] [--decay DECAY]
                        [--alpha-increment ALPHA_INCREMENT] [--l1 L1]
                        [--step-size STEP_SIZE] [--T T]
                        [--memory-budgets MEMORY_BUDGETS [MEMORY_BUDGETS ...]]
                        [--epochs-class EPOCHS_CLASS] [--dataset DATASET]
                        [--lwf] [--no-nl] [--rand] [--adversarial]

Default configurations can be used to run with same parameters as used by iCaRL. Simply run:

python run_experiment.py

Dependencies

Pytorch 0.3.0.post4
Python 3.6
torchnet (https://github.com/pytorch/tnt)
tqdm (pip install tqdm)

Please see requirements.txt for a complete list.

Setting up enviroment

The easiest way to install the required dependencies is to use conda package manager.

Install Anaconda with Python 3
Install pytorch and torchnet
Install tqdm (pip install progressbar2) Done.

Branches

iCaRL + Dynamic Threshold Moving is implemented in "Autoencoders" branch.

=======

Selected Results

Removing Bias by Dynamic Threshold Moving

Result of threshold moving with T = 2 and 5. Note that different scale is used for the y axis, and using higher temperature in general results in less bias.

Confusion Matrix with and without Dynamic Threshold Moving

Confusion matrix of results of the classifier with (right) and without (left) threshold moving with T=2. We removed the first five classes of MNIST from the train set and only distilled the knowledge of these classes using a network trained on all classes. Without threshold moving the model struggled on the older classes. With threshold moving, however, not only was it able to classify unseen classes nearly perfectly, but also its performance did not deteriorate on new classes

FAQs

How do I implement more models?

A. Add the model in model/ModelFactory and make sure the forward method of the model satisfy the API of model/resnet32.py

How do I add a new dataset?

A. Add the new dataset in DatasetFactory and specify the details in the dataHandler/dataset.py class. Make sure the dataset implements all the variables set by other datasets.

References

[1] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015

[2] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. Icarl: Incremental classifier and representation learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2001–2010, 2017.

[3] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017.

khurramjaved96/incremental-learning