This project contains the implementation of the following ICLR 2021 paper:
Title: Contextual Transformation Networks for Online Continual Learning (ICLR 2021). [openreview], [pdf].
Authors: Quang Pham, Chenghao Liu, Doyen Sahoo, and Steven Hoi
CTN proposes a novel network design with a controller that can efficiently extract task-specific features from a base network. Both the base network and the controller have access to their own memory units and are joinly trained via a bilevel optimization strategy.
- Pytorch 1.5.0
- CUDA 10.2
All experiments in this work was run on a single K80 GPU with 12Gb memory.
Follow the instructions in the data/
folders to prepare the benchmarks.
To replicate our results on the Permuted MNIST, Split CIFAR100, CORE50, and Split miniImagenet, run
chmod 777 scripts/run.sh
./scripts/run.sh
The results will be put in the resuts/
folders.
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 sets are 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. 50 |
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 | e.g. 0.3 |
inner_steps | number of SGD udpates per samples | e.g. 2 |
n_meta | number of outer updates per samples | e.g. 2 |
n_val | percentage of the total memory used for the semantic memory (in CTN) | e.g. 0.2 |
replay_batch_size | number of data in the memory used per experience replay step | e.g. 64 |
If you found CTN useful for your research, please consider citing.
@inproceedings{pham2020contextual,
title={Contextual transformation networks for online continual learning},
author={Pham, Quang and Liu, Chenghao and Sahoo, Doyen and Steven, HOI},
booktitle={International Conference on Learning Representations},
year={2021}
}
This project structure is based on the GEM repository with additional methods, metrics and implementation improvements. For the CORe50 benchmark, we modify the data loader from MIR.