/IAS

Exploring Class-incremental Learning Using Incremental Architecture Search

Primary LanguagePython

Exploring Class-incremental Learning Using Incremental Architecture Search

Link to paper: https://marcpickett.com/cl2018/CL-2018_paper_34.pdf

To run the experiments:

To replicate the MNIST experiment in the paper:

python -u runMLP_SAS.py |& tee output.txt

To run MNIST experiment with a fixed architecture:

python -u fixed_architecture.py |& tee output.txt

To replicate the incremental experiment of Fashion-MNIST and MNIST in the paper

python -u run_Incremental_SAS.py |& tee output.txt

Note that both experiments utilizes around 8 CPUs and might take long to run. output.txt reports the validation accuracy per epoch as well as the test accuracies of each class once training is finished

Dependencies:

Python 3.6.4

Keras 2.2.0

Tensorflow 1.8.0 (cpu version)

numpy 1.14.5

pathos 0.2.1