cuongpxu / bilevel-continual-learning

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.

Cite

Requirements

• Pytorch 1.5.0
• CUDA 10.2

All experiments in this work was run on a single K80 GPU with 12Gb memory.

Benchmarks

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 mnist_permutations.pt and cifar100.pt in the data/raw/ folder. Then, run the data/cifar100.py and 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 resuts/ folders.

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

Parameter	Description	Values
data_path	path where the data will be saved	e.g. data/
data_file	name of the data file	e.g. mnist_permutations.pt
use	randomly use a subset of data. When use < 1, use% of the original data, when use > 1, select use samples from the data	e.g. 0.5 (select 50% of data), 1000 (select 1000 data samples)
n_memories	number of data stored per task	e.g. 65
memory_strength	value of the regularizer's coefficient	e.g. 100
temperature	temperature of the softmax in knowledge distillation	e.g. 5
lr	(inner) learning rate	e.g. 0.1
beta	outer learning rate (BCL)	e.g. 0.3
adapt	use adaptation at test time or not	e.g. no
adapt_lr	learning rate of the adaptation step	e.g. 0.001
inner_steps	number of SGD udpates per samples	e.g. 2
n_meta	number of meta update per samples	e.g. 2
n_val	percentage of the total memory used for the evaluation set (BCL-Dual)	e.g. 0.2
replay_batch_size	number of data in the memory used per experience replay step	e.g. 64

Acknowledgement

This project structure is based on the GEM repository with additional methods, metrics and implementation improvements.

• https://github.com/facebookresearch/GradientEpisodicMemory