This repository contains the implementation code for paper:
On Learning Contrastive Representations for Learning With Noisy Labels
Li Yi, Sheng Liu, Qi She, A. Ian McLeod, Boyu Wang
IEEE / CVF Computer Vision and Pattern Recognition Conference (CVPR 2022)
[Paper]
Please refer to [run.sh] for running different experiments.
This repository implements our method with BYOL framework for better performance as indicated in our paper (Tab.5).
Here we show the accuracy of our method + CE on CIFAR-10 in 3 trails:
| 0% | SYM 20% | SYM 40% | SYM 60% | SYM 80% | SYM 90% | ASYM 40% | |
|---|---|---|---|---|---|---|---|
| Seed 2021 (Best/Last) | 94.40/94.31 | 93.19/93.19 | 92.14/91.91 | 89.18/89.08 | 88.00/87.99 | 84.58/84.49 | 89.23/89.23 |
| Seed 2022 (Best/Last) | 94.22/94.15 | 93.54/93.29 | 92.36/92.22 | 89.46/88.38 | 87.56/87.39 | 83.48/83.12 | 89.43/88.16 |
| Seed 2023 (Best/Last) | 94.36/94.24 | 93.22/93.09 | 92.16/92.08 | 88.87/88.61 | 87.10/86.87 | 84.32/84.10 | 89.71/89.60 |
| Average | 94.32/94.23 | 93.31/93.19 | 92.21/92.07 | 89.17/88.69 | 87.55/87.41 | 84.12/83.90 | 89.45/89.00 |
Here we show the accuracy of our method + CE on CIFAR-100 in 3 trails:
| 0% | SYM 20% | SYM 40% | SYM 60% | SYM 80% | ASYM 40% | |
|---|---|---|---|---|---|---|
| Seed 2021 (Best/Last) | 75.75/75.54 | 71.80/71.52 | 69.21/69.06 | 62.95/62.70 | 55.07/54.94 | 55.76/54.81 |
| Seed 2022 (Best/Last) | 76.03/75.93 | 71.75/71.50 | 68.09/67.97 | 62.80/62.33 | 56.04/55.93 | 54.57/54.10 |
| Seed 2023 (Best/Last) | 76.35/76.14 | 71.84/71.63 | 67.97/67.72 | 63.43/63.18 | 54.83/54.42 | 55.46/54.68 |
| Average | 76.04/75.87 | 71.79/71.55 | 68.42/68.25 | 63.06/62.73 | 55.31/55.09 | 54.93/54.49 |
Here we show the accuracy of our method + GCE on CIFAR-10 in 3 trails:
| SYM 20% | SYM 40% | SYM 60% | SYM 80% | SYM 90% | |
|---|---|---|---|---|---|
| Seed 2021 (Best/Last) | 94.40/94.36 | 93.85/93.76 | 93.09/92.96 | 91.81/91.63 | 89.29/88.92 |
| Seed 2022 (Best/Last) | 94.51/94.38 | 93.49/93.44 | 93.15/92.98 | 91.20/91.03 | 89.79/89.65 |
| Seed 2023 (Best/Last) | 94.34/94.25 | 93.70/93.36 | 93.20/93.14 | 92.00/91.81 | 90.44/90.34 |
| Average | 94.41/94.33 | 93.67/93.52 | 93.14/93.02 | 91.67/91.49 | 89.83/87.63 |
Here we show the accuracy of our method + GCE on CIFAR-100 in 3 trails:
| SYM 20% | SYM 40% | SYM 60% | SYM 80% | |
|---|---|---|---|---|
| Seed 2021 (Best/Last) | 74.26/73.90 | 72.65/72.37 | 70.66/70.29 | 64.59/64.38 |
| Seed 2022 (Best/Last) | 74.67/74.55 | 73.12/72.90 | 70.47/70.13 | 64.05/63.81 |
| Seed 2023 (Best/Last) | 74.66/74.41 | 72.85/72.71 | 70.97/70.70 | 63.47/63.29 |
| Average | 74.53/74.28 | 72.87/72.66 | 70.70/70.46 | 64.03/63.82 |
This code inherits some codes from SimSiam.