Skip to content

README.md

Latest commit

 

History

History
254 lines (206 loc) · 10.3 KB

README.md

File metadata and controls

254 lines (206 loc) · 10.3 KB
Raw

Image Classification Tutorial

Table of Contents

  1. Updates
  2. Features
  3. Dependencies
  4. Data
  5. Baseline Results
  6. Pretraining Results
  7. FixRes Results
  8. References

1. Updates 📰

  • [2025.06] Support DINOv2
  • [2025.02] Support Reliability Metric: AURC ( 👍 AURC = 👍 classification accuracy + 👍 reliability)
  • [2025.02] Add Deit-B and ConvNext-B Results
  • [2024.02] Add FixRes training strategy
  • [2024.01] Initial release with ResNet support

2. Features 🚥

  • Supporting DINOv2, 🔥 For DINOv2-base with register, CIFAR10: 99.3%; CIFAR100: 93.8%; CUB: 90.8% ; CARS 95.1% 🔥
  • Supporting ConvNext, Deit; 🔥 For ConvNext, CIFAR10: 98.8%; CIFAR100: 92.4%; CUB: 91.2% ; CARS 93.3% 🔥
  • Supporting Reliability Metric: AURC, see here for details
  • 95.62% top-1 acc on CIFAR10, 79.11% top-1 acc on CIFAR100 with resnet18
  • Support CIFAR10, CIFAR100, CUB and CARS
  • Support standard retrained weight on Inet1K in supervised and self-supervised fashion (MOCOV2)
  • Fixing the train-test resolution discrepancy
  • Tensorboard Visualization, EMA, ONLY simple data aug

tb

3. Dependencies 📚

The model can be trained on a single GPU with more than 12 GB of memory.

  • Install PyTorch adapted to your CUDA version via Conda:

    conda install pytorch torchvision pytorch-cuda=11.8 -c pytorch -c nvidia

  • Install TensorBoard:

    conda install matplotlib tensorboard

4. Data 🔢

Dataset Nb CLS No. Training No. Test
CIFAR10 10 50,000 10,000
CIFAR100 100 50,000 10,000
CUB 200 5,994 5,794
CARS 196 8,144 8,041

One can directly go to ./data/, launch download.sh to download CUB and CARS dataset.

5. Baseline Results 📝

  • CUB, CARS are trained and tested with 224 * 224
  • CIFAR10, CIFAR100 are trained and tested with 32 * 32
LR EMA Arch CUB CARS CIFAR10 CIFAR100
0.05 - ResNet18 60.74 87.33 95.31 79.13
0.05 TRUE ResNet18 64.64 87.54 95.62 79.11
0.05 - ResNet50 57.16 88.87 - -
0.05 TRUE ResNet50 63.43 89.27 - -

EMA improves in most cases Reproducing the above exp with:

bash baseline.sh

6. Pretraining Results 📑

  • CUB, CARS are trained and tested with 224 * 224
  • Report result with EMA
LR Pretrained Arch CUB CARS
0.05 - ResNet18 64.64 87.54
0.005 Inet1K ResNet18 77.11 88.42
0.05 - ResNet50 63.43 89.27
0.005 Inet1K ResNet50 84.47 91.38
0.005 MocoV2 ResNet50 79.01 92.33

DEIT-B and ConvNext-B

Accuracy with DEIT-B and ConvNext-B Models (Resolution 384 * 384)

LR Pretrained EMA Arch CUB CARS CIFAR10 CIFAR100
0.002 Inet1K - DEIT-B-384 87.05 92.44 98.33 88.24
0.002 Inet1K True DEIT-B-384 87.42 92.87 98.83 90.47
0.002 Inet21K - DEIT-B-384 88.87 91.56 98.61 89.73
0.002 Inet21K True DEIT-B-384 89.61 92.17 98.99 92.31
0.002 Inet21K + 1K - ConvNext-B 91.09 93.35 98.73 92.04
0.002 Inet21K + 1K True ConvNext-B 91.22 93.32 98.87 92.40
0.002 Inet21K - ConvNext-B 91.49 93.14 98.73 91.76
0.002 Inet21K True ConvNext-B 91.58 93.17 98.84 92.14

AURC with DEIT-B and ConvNext-B Models (Resolution 384 * 384)

LR Pretrained EMA Arch CUB CARS CIFAR10 CIFAR100
0.002 Inet1K - DEIT-B-384 28.67 16.20 0.78 19.28
0.002 Inet1K True DEIT-B-384 27.05 14.44 0.45 13.59
0.002 Inet21K - DEIT-B-384 23.76 18.12 0.63 15.77
0.002 Inet21K True DEIT-B-384 23.02 15.14 0.50 9.72
0.002 Inet21K + 1K - ConvNext-B 19.28 12.92 0.42 10.26
0.002 Inet21K + 1K True ConvNext-B 18.79 12.75 0.39 9.14
0.002 Inet21K - ConvNext-B 19.78 12.61 0.59 11.08
0.002 Inet21K True ConvNext-B 19.41 12.30 0.48 9.78

DINOv2-B

Accuracy with DINOv2 Models (Resolution 378 * 378) + AdamW optimizer

LR Register Arch CUB CARS CIFAR10 CIFAR100
2e-5 - DINOv2-B-378 90.52 94.66 99.09 92.68
2e-5 True DINOv2-B-378 90.84 95.11 99.25 93.73

AURC with DINOv2 Models (Resolution 378 * 378) + AdamW optimizer

LR Register Arch CUB CARS CIFAR10 CIFAR100
2e-5 - DINOv2-B-378 20.65 9.96 0.37 8.83
2e-5 True DINOv2-B-378 18.88 9.23 0.30 7.42

Pretraining matters Reproducing the above exp with:

cd pretrained_weight/
bash download.sh
cd ..
bash pretraining.sh 
## for reproduction of dinov2 experiments
bash dinov2_b.sh

7. FixRes Results 📈

Stage 1 Stage 2 Test Size 224*224 Test Size 448*448
LR Epoch Train Size LR Epoch Train Size CUB CARS CUB CARS
0.005 500 224*224 - - - 84.47 91.38 73.86 91.58
0.005 500 448*448 - - - 78.56 79.23 85.71 91.88
0.005 400 224*224 0.001 100 448*448 82.59 87.46 86.07 92.18

  • Training and testing with the same resolution achieves more stable accuracy

  • Training on a small resolution then finetuning on larger resolution manage to achieve consistent gain

  • Check paperswithcode, these results 86.07 on CUB and 92.18 on CARS are not bad.

Reproducing the above exp with:

  cd pretrained_weight/
  bash download.sh
  cd ..
  bash fix_resolution.sh

8. References 📜

关注我们，其他感兴趣的内容