This is the main repository for the master's thesis Incremental Learning in the Mobile Scenario.

Average incremental accuracy

* Results of ResNet-backbone models are reported directly from [Douillard et al. 2020].

If conda is not installed yet, please install Miniconda or Anaconda first.

This pipeline was tested and run on Ubuntu and Windows WSL.

When you install the requirements on other OS, if you encounter ResolvePackageNotFound error, please first comment these packages in envrionment.yml and try again.

1. Change the working directory to the root of the project directory.

   cd <your_project_dirctory>

2. Setup the environment

   conda env create --name il_mobile --file environment.yml

3. Activate the environment

   conda activate il_mobile

4. Experiments on CIFAR-100

   • Run MobileNetV2 baseline

     python demo_baseline.py

   • Run PODNet using ResNet as backbone

     python3 -minclearn --options options/podnet/podnet_cnn_cifar100.yaml options/data/cifar100_3orders.yaml \
         --initial-increment 50 --increment 1 --fixed-memory \
         --device <GPU_ID> --label podnet_cnn_cifar100_50steps_resnet \
         --data-path <PATH/TO/DATA>

   • Run PODNet using MobileNetV2 as backbone

     python3 -minclearn --options options/podnet/podnet_cnn_cifar100_mobilenetv2.yaml options/data/cifar100_3orders.yaml \
         --initial-increment 50 --increment 1 --fixed-memory \
         --device <GPU_ID> --label podnet_cnn_cifar100_50steps_mobilenetv2 \
         --data-path <PATH/TO/DATA>

   • Run iCaRL or BiC, just change to corresponding --options and give a new --label

5. Deactivate the environment

   conda deactivate

6. [Optional] Delete the environment if you do not need it any more

   conda env list
   conda env remove --name il_mobile

Some of the code are from their original repositories and adapted to our protocol. To get full look of the original repositories please check incremental_learning.pytorch and pytorch-cifar.