by Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu and Stella X. Yu at UC Berkeley, ICSI and NTU
International Conference on Learning Representations (ICLR), 2021. Spotlight Presentation
Project Page | PDF | Preprint | OpenReview | Slides | Citation
This repository contains an official re-implementation of RIDE from the authors, while also supports several other works on long-tailed recognition. Further information please contact Xudong Wang and Long Lian.
This repo has RIDE on ResNet and ResNeXt. For RIDE on Swin-Transformer, see it here.
If you find our work inspiring or use our codebase in your research, please consider giving a star ⭐ and a citation.
@inproceedings{wang2020long,
title={Long-tailed Recognition by Routing Diverse Distribution-Aware Experts},
author={Wang, Xudong and Lian, Long and Miao, Zhongqi and Liu, Ziwei and Yu, Stella},
booktitle={International Conference on Learning Representations},
year={2021}
}
- RIDE
- Cross-Entropy (CE) Loss
- Focal Loss
- LDAM Loss
- Decouple: cRT (limited support for now)
- Decouple: tau-normalization (limited support for now)
[05/2022] We release an additional RIDE repo with Swin-Transformer. Check it out.
[04/2022] We add additional iNaturalist checkpoints with more experts and longer training in model zoo.
[04/2021] Pre-trained models are avaliable in model zoo.
[12/2020] We added an approximate GFLops counter. See usages below. We also refactored the code and fixed a few errors.
[12/2020] We have limited support on cRT and tau-norm in load_stage1 option and t-normalization.py, please look at the code comments for instructions while we are still working on it.
[12/2020] Initial Commit. We re-implemented RIDE in this repo. LDAM/Focal/Cross-Entropy loss is also re-implemented (instruction below).
- Python >= 3.7, < 3.9
- PyTorch >= 1.6
- tqdm (Used in
test.py) - tensorboard >= 1.14 (for visualization)
- pandas
- numpy
8 GPUs with >= 11G GPU RAM are recommended. Otherwise the model with more experts may not fit in, especially on datasets with more classes (the FC layers will be large). We do not support CPU training, but CPU inference could be supported by slight modification.
CIFAR code will download data automatically with the dataloader. We use data the same way as classifier-balancing. For ImageNet-LT and iNaturalist, please prepare data in the data directory. ImageNet-LT can be found at this link. iNaturalist data should be the 2018 version from this repo (Note that it requires you to pay to download now). The annotation can be found at here. Please put them in the same location as below:
data
├── cifar-100-python
│ ├── file.txt~
│ ├── meta
│ ├── test
│ └── train
├── cifar-100-python.tar.gz
├── ImageNet_LT
│ ├── ImageNet_LT_open.txt
│ ├── ImageNet_LT_test.txt
│ ├── ImageNet_LT_train.txt
│ ├── ImageNet_LT_val.txt
│ ├── test
│ ├── train
│ └── val
└── iNaturalist18
├── iNaturalist18_train.txt
├── iNaturalist18_val.txt
└── train_val2018
We have a model zoo available.
python train.py -c "configs/config_imbalance_cifar100_ride.json" --reduce_dimension 1 --num_experts 3
Note: --reduce_dimension 1 means set reduce dimension to True. The template has an issue with bool arguments so int argument is used here. However, any non-zero value will be equivalent to bool True.
python train.py -c "configs/config_imbalance_cifar100_distill_ride.json" --reduce_dimension 1 --num_experts 3 --distill_checkpoint path_to_checkpoint
Distillation is not required but could be performed if you'd like further improvements.
python train.py -c "configs/config_imbalance_cifar100_ride_ea.json" -r path_to_stage1_checkpoint --reduce_dimension 1 --num_experts 3
Note: different runs will result in different EA modules with different trade-off. Some modules give higher accuracy but require higher FLOps. Although the only difference is not underlying ability to classify but the "easiness to satisfy and stop". You can tune the pos_weight if you think the EA module consumes too much compute power or is using too few expert.
python train.py -c "configs/config_imagenet_lt_resnet10_ride.json" --reduce_dimension 1 --num_experts 3
python train.py -c "configs/config_imagenet_lt_resnet50_ride.json" --reduce_dimension 1 --num_experts 3
python train.py -c "configs/config_imagenet_lt_resnext50_ride.json" --reduce_dimension 1 --num_experts 3
python train.py -c "configs/config_imagenet_lt_resnet10_distill_ride.json" --reduce_dimension 1 --num_experts 3 --distill_checkpoint path_to_checkpoint
python train.py -c "configs/config_imagenet_lt_resnet50_distill_ride.json" --reduce_dimension 1 --num_experts 3 --distill_checkpoint path_to_checkpoint
python train.py -c "configs/config_imagenet_lt_resnext50_distill_ride.json" --reduce_dimension 1 --num_experts 3 --distill_checkpoint path_to_checkpoint
python train.py -c "configs/config_imagenet_lt_resnet10_ride_ea.json" -r path_to_stage1_checkpoint --reduce_dimension 1 --num_experts 3
python train.py -c "configs/config_imagenet_lt_resnet50_ride_ea.json" -r path_to_stage1_checkpoint --reduce_dimension 1 --num_experts 3
python train.py -c "configs/config_imagenet_lt_resnext50_ride_ea.json" -r path_to_stage1_checkpoint --reduce_dimension 1 --num_experts 3
python train.py -c "configs/config_iNaturalist_resnet50_ride.json" --reduce_dimension 1 --num_experts 3
python train.py -c "configs/config_iNaturalist_resnet50_distill_ride.json" --reduce_dimension 1 --num_experts 3 --distill_checkpoint path_to_checkpoint
python train.py -c "configs/config_iNaturalist_resnet50_ride_ea.json" -r path_to_stage1_checkpoint --reduce_dimension 1 --num_experts 3
- Focal Loss: switch the loss to Focal Loss
- Cross Entropy: switch the loss to Cross Entropy Loss
To test a checkpoint, please put it with the corresponding config file.
python test.py -r path_to_checkpoint
Please see the pytorch template that we use for additional more general usages of this project (e.g. loading from a checkpoint, etc.).
We provide an experimental support for approximate GFLops calculation. Please open an issue if you encounter any problem or meet inconsistency in GFLops.
You need to install thop package first. Then, according to your model, run python -m utils.gflops (args) in the project directory.
Use python -m utils.gflops to see the documents as well as explanations for this calculator.
python -m utils.gflops ResNeXt50Model 0 --num_experts 3 --reduce_dim True --use_norm False
To change model, switch ResNeXt50Model to the ones used in your config. use_norm comes with LDAM-based methods (including RIDE). reduce_dim is used in default RIDE models. The 0 in the command line indicates the dataset.
All supported datasets:
- 0: ImageNet-LT
- 1: iNaturalist
- 2: Imbalance CIFAR 100
python -m utils.gflops ResNet50Model 1 --num_experts 3 --reduce_dim True --use_norm True
python -m utils.gflops ResNet32Model 2 --num_experts 3 --reduce_dim True --use_norm True
We provide a ea_percentage for specifying the percentage of data that pass each expert. Note that you need to switch to the EA model as well since you actually use EA model instead of the original model in training and inference.
An example:
python -m utils.gflops ResNet32EAModel 2 --num_experts 3 --reduce_dim True --use_norm True --ea_percentage 40.99,9.47,49.54
See FAQ.
If you have any general questions, feel free to email us at longlian at berkeley.edu and xdwang at eecs.berkeley.edu. If you have code or implementation-related questions, please feel free to send emails to us or open an issue in this codebase (We recommend that you open an issue in this codebase, because your questions may help others).
This is a project based on this pytorch template. The readme of the template explains its functionality, although we try to list most frequently used ones in this readme.
This project is licensed under the MIT License. See LICENSE for more details. The parts described below follow their original license.
This is a project based on this pytorch template. The pytorch template is inspired by the project Tensorflow-Project-Template by Mahmoud Gemy
The ResNet and ResNeXt in fb_resnets are based on from Classifier-Balancing/Decouple. The ResNet in ldam_drw_resnets/LDAM loss/CIFAR-LT are based on LDAM-DRW. KD implementation takes references from CRD/RepDistiller.
