Bilevel Continual Learning
This project contains the implementation of the paper: Bilevel Continual Learning (arXiv). A dual memory management strategy: a replay memory and an evaluation set. The model learns new samples by experience replay with the replay memory such that in can generalize to the evaluation set.
- Pytorch 1.5.0
- CUDA 10.2
All experiments in this work was run on a single K80 GPU with 12Gb memory.
1. Prepare data
This project uses the same data format as GEM, which includes benchmark such as Permutation MNIST, rotation MNIST, split CIFAR, etc.
To prepare the datasets, follow the GEM's instruction to create the
cifar100.pt in the
Then, run the
data/mnist_permutations.py scripts to create the corresponding benchmarks. Each benchmark will consists of two files: the
-val.pt file only contains 3 tasks used for hyper-parameter cross-validation and the
-cl.pt file contains the remaining tasks for actual continual learning.
2. Run experiments
To replicate our results on the Permuted MNIST, Split CIFAR100, Split CUB, and Split miniImagenet, run
chmod 777 scripts/run.sh ./scripts/run.sh
The results will be put in the
3. Parameter Setting
The provided script
scripts/run.sh includes the best hyper-parameter cross-validated from the cross-validation tasks. The following is the list of parameters you can experiment with
|data_path||path where the data will be saved||e.g.
|data_file||name of the data file||e.g.
|use||randomly use a subset of data. When
|n_memories||number of data stored per task||e.g.
|memory_strength||value of the regularizer's coefficient||e.g.
|temperature||temperature of the softmax in knowledge distillation||e.g.
|lr||(inner) learning rate||e.g.
|beta||outer learning rate (BCL)||e.g.
|adapt||use adaptation at test time or not||e.g.
|adapt_lr||learning rate of the adaptation step||e.g.
|inner_steps||number of SGD udpates per samples||e.g.
|n_meta||number of meta update per samples||e.g.
|n_val||percentage of the total memory used for the evaluation set (BCL-Dual)||e.g.
|replay_batch_size||number of data in the memory used per experience replay step||e.g.
This project structure is based on the GEM repository with additional methods, metrics and implementation improvements.