jovialio / sslad

3rd Place Solution for ICCV 2021 Workshop SSLAD Track 3A - Continual Learning Classification Challenge

Online Continual Learning via Multiple Deep Metric Learning and Uncertainty-guided Episodic Memory Replay

3^rd Place Solution for ICCV 2021 Workshop SSLAD Track 3A - Continual Learning Classification

Description

Official implementation of our solution (3^rd place) for ICCV 2021 Workshop Self-supervised Learning for Next-Generation Industry-level Autonomous Driving (SSLAD) Track 3A - Continual Learning Classification using "Online Continual Learning via Multiple Deep Metric Learning and Uncertainty-guided Episodic Memory Replay".

How to run

First, install dependencies

# clone project   
git clone https://github.com/mrifkikurniawan/sslad.git

# install project   
cd sslad 
pip3 install -r requirements.txt   

Next, preparing the dataset via links below.

• Train/val: SODA10M Labeled Trainval (Google Drive). Then, replace the original annotations with this new annotations with timestamps: [train][val].
• Test: SODA10M 3A Tesetset (Google Drive)

Next, run training.

# run training module with our proposed cl strategy
python3.9 classification.py \
--config configs/cl_strategy.yaml \
--name {path/to/log} \
--root {root/of/your/dataset} \
--num_workers {num workers} \
--gpu_id {your-gpu-id} \
--comment {any-comments} 
--store \

or see the train.sh for the example.

Experiments Results

Method	Val AMCA	Test AMCA
Baseline (Uncertainty Replay)*	57.517	-
+ Multi-step Lr Scheduler*	59.591 (+2.07)	-
+ Soft Labels Retrospection*	59.825 (+0.23)	-
+ Contrastive Learning*	60.363 (+0.53)	59.68
+ Supervised Contrastive Learning*	61.49 (+1.13)	-
+ Change backbone to ResNet50-D*	62.514 (+1.02)	-
+ Focal loss*	62.71 (+0.19)	-
+ Cost Sensitive Cross Entropy	63.33 (+0.62)	-
+ Class Balanced Focal loss*	64.01 (+1.03)	64.53 (+4.85)
+ Head Fine-tuning with Class Balanced Replay	65.291 (+1.28)	62.58 (-1.56)
+ Head Fine-tuning with Soft Labels Retrospection	66.116 (+0.83)	62.97 (+0.39)

^*Applied to our final method.

File overview

classification.py: Driver code for the classification subtrack. There are a few things that can be changed here, such as the model, optimizer and loss criterion. There are several arguments that can be set to store results etc. (Run classification.py --help to get an overview, or check the file.)

class_strategy.py: Provides an empty plugin. Here, you can define your own strategy, by implementing the necessary callbacks. Helper methods and classes can be ofcourse implemented as pleased. See here for examples of strategy plugins.

data_intro.ipynb: In this notebook the stream of data is further introduced and explained. Feel free to experiment with the dataset to get a good feeling of the challenge.

Note: not all callbacks have to be implemented, you can just delete those that you don't need.

classification_util.py & haitain_classification.py: These files contain helper code for dataloading etc. There should be no reason to change these.

Citation

Please consider citing our paper in your publications if the project helps your research. BibTeX reference is as follow.

@article{Kurniawan2021OnlineCL,
  title={Online Continual Learning via Multiple Deep Metric Learning and Uncertainty-guided Episodic Memory Replay - 3rd Place Solution for ICCV 2021 Workshop SSLAD Track 3A Continual Object Classification},
  author={Muhammad Rifki Kurniawan and Xing Wei and Yihong Gong},
  journal={ArXiv},
  year={2021},
  volume={abs/2111.02757}
}