This code was used in the SSLAD Track 3 Challenge. To facilitate easier use outside the challenge, a new repository has been created with updated and refactored code. Please use: CLAD repo from now on.
This repository provides the code required to participate in the thirth track of the SSLAD ICCV 2021 workshop, about Continual Learning. The code is build with the Avalanche framework, making it easy to compare and setup very customizable benchmarks.
Tested with python 3.9, pytorch 1.9.0, torchvision 0.10.0 and cuda toolkit 10.2
- Clone the Avalanche Fork here
git clone https://github.com/VerwimpEli/avalanche.git - Create a (new) conda environment with PyTorch and torchvision installed.
conda create -n sslad python=3.9
conda activate sslad
conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch -c conda-forge - Update your conda environment with the provided .yml file.
conda env update --file env.yml - Then, make sure the Avalanche directory is on your PYTHONPATH.
export PYTHONPATH=[Avalanche directory]:$PYTHONPATH
Any questions, problems... can be mailed to eli.verwimp(at)kuleuven.be, or by creating a git issue.
Implementing a new strategy in Avalanche works by implementing a
subclass of the StrategyPlugin. This plugin is passed through
when updating the model and its callback methods are called during training
of the model. For examples of how such strategies work, see the training/plugins
folder inside Avalanche for some well known methods. For a more in depth explanation,
see the tutorials on the avalanche website.
Don't hesitate to contact me at eli.verwimp(at)kuleuven.be if you don't immediately know how to implement your method in Avalanche. It's easy, but learning a new framework can be challenging sometimes.
-
Currently PyTorch with Python 3.9 is giving a warning about leaking thread pools when using multiple workers. This is expected. See this issue.
-
The current installation command on the PyTorch website doesn't work. You should add the
conda-forgechannel for installing.