This is an unofficial PyTorch implementation of the paper "Conditional Channel Gated Networks for Task-Aware Continual Learning" by Abati et al. (https://arxiv.org/abs/2004.00070).
At the moment, the code implements the task-incremental setting only. I have put some work into implementing class-incremental setup, but it is not finished yet.
I have tried to follow the paper as close as possible, but maybe there are some details I missed/got wrong. So please use this implementation with care and feel free to ask questions and open PRs.
- Python 3.8
- PyTorch 1.7.1 (was also tested on 1.5.0, may work on some older versions)
The rest required modules are listed in requirements.txt
-
Set up a proper configuration by editing config/cfg.py
-
Run the following command to train the model
make train
- All tensorboard logs, checkpoints, and task-incremental accuracies will be stored in results/experiment_tag
To visualize the sparseness of the model and see how many kernels were frozen by each task, check Sparse_visualization.ipynb
Due to some differences in implementation, the hyperparameters for ResNet18 mentioned in the paper didn't work well. I changed the optimizer and performed upscaling + augmentations to get better quality.
You can download checkpoint with the best results for ResNet18 so far here. For consistency with other experiments, you may unzip it to the results folder.
There are a couple of differences between my and the original authors' implementation:
-
To speed up the computation, it is assumed that the batch contains elements from one particular task only.
-
I also observed a strange sparsity pattern for all values of lambda_sparse I used (from 0.1 to 0.5). There is free kernels attrition in the first couple of layers when the first 2-3 tasks use all available capacity, but later layers show underuse. This may be a result of the bug, so let me know if you find one :)