🌠 Towards Stable Test-Time Adaptation in Dynamic Wild World
September 8, 2023 · View on GitHub
This is the official project repository for Towards Stable Test-Time Adaptation in Dynamic Wild World 🔗 by Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao and Mingkui Tan (ICLR 2023 Oral, Notable-Top-5%).
- 1️⃣ SAR conducts model learning at test time to adapt a pre-trained model to test data that has distributional shifts ☀️ 🌧 ❄️, such as corruptions, simulation-to-real discrepancies, and other differences between training and testing data.
- 2️⃣ SAR aims to adapt a model in dymamic wild world, i.e., the test data stream may have mixed domain shifts, small batch size, and online imbalanced label distribution shifts (as shown in the figure below).
Method: Sharpness-Aware and Reliable Entropy Minimization (SAR)
-
1️⃣ SAR conducts selective entropy minimization by excluding partial samples with noisy gradients out of online adaptation.
-
2️⃣ SAR optimizes both entropy and the sharpness of entropy surface simutaneously, so that the model update is robust to those remaining samples with noisy gradients.
Installation:
SAR depends on
Data preparation:
This repository contains code for evaluation on ImageNet-C 🔗 with ResNet-50 and VitBase. But feel free to use your own data and models!
-
Step 1: Download ImageNet-C 🔗 dataset from here 🔗.
-
Step 2: Put IamgeNet-C at "--data_corruption".
-
Step 3 [optional, for EATA]: Put ImageNet test/val set at "--data".
Usage:
import sar
from sam import SAM
model = TODO_model()
model = sar.configure_model(model)
params, param_names = sar.collect_params(model)
base_optimizer = torch.optim.SGD
optimizer = SAM(params, base_optimizer, lr=args.lr, momentum=0.9)
adapt_model = sar.SAR(net, optimizer, margin_e0=0.4*math.log(1000))
outputs = adapt_model(inputs) # now it infers and adapts!
Example: Adapting a pre-trained model on ImageNet-C (Corruption).
Usage:
python3 main.py --data_corruption /path/to/imagenet-c --exp_type [normal/bs1/mix_shifts/label_shifts] --method [no_adapt/tent/eata/sar] --model [resnet50_gn_timm/vitbase_timm] --output /output/dir
'--exp_type' is choosen from:
-
'normal' means the same test setting to prior mild data stream in Tent and EATA
-
'bs1' means single sample adaptation, only one sample comes each time-step
-
'mix_shifts' conducts exps over the mixture of 15 corruption types in ImageNet-C
-
'label_shifts' means exps under online imbalanced label distribution shifts. Moreover, imbalance_ratio indicates the imbalance extent
Note: For EATA method, you need also to set "--data /path/to/imagenet" of clean ImageNet test/validation set to compute the weight importance for regularization.
Experimental results:
The Table below shows the results under online imbalanced label distribution shifts. The reported average accuracy is averaged over 15 different corruption types in ImageNet-C (severity level 5).
| ResNet-50 (BN) | ResNet-50 (GN) | VitBase (LN) | |
|---|---|---|---|
| No adapt | 18.0 | 30.6 | 29.9 |
| MEMO | 24.0 | 31.3 | 39.1 |
| DDA | 27.2 | 35.1 | 36.2 |
| Tent | 2.1 | 22.0 | 47.3 |
| EATA | 0.9 | 31.6 | 49.9 |
| SAR (ours) | -- | 37.2 0.6 | 58.0 0.5 |
Please see our PAPER 🔗 for more detailed results.
Correspondence
Please contact Shuaicheng Niu by [niushuaicheng at gmail.com] if you have any questions. 📬
Citation
If our SAR method or wild test-time adaptation settings are helpful in your research, please consider citing our paper:
@inproceedings{niu2023towards,
title={Towards Stable Test-Time Adaptation in Dynamic Wild World},
author={Niu, Shuaicheng and Wu, Jiaxiang and Zhang, Yifan and Wen, Zhiquan and Chen, Yaofo and Zhao, Peilin and Tan, Mingkui},
booktitle = {Internetional Conference on Learning Representations},
year = {2023}
}