Adaptive Boosting for Domain Adaptation: Towards Robust Predictions in Scene Segmentation
The method is designed for domain adaptation segmentation (https://github.com/layumi/AdaBoost_Seg). It also can be applied to the semi-supervised setting on Cifar-10.
python -m experiments.cifar10_testHere we do not change the original weight averaging setting in MeanTeacher, but add the adaptive sampler (every 2 epochs update sampler once).
python -m experiments.ada_cifar10_alpha2We run all experiments 10 times for a fair comparison. We only have 3 GPU on one machine, so we re-run the code. It is slightly different with the paper.
| CIFAR-10 using 4000 labels | test error |
|---|---|
| CT-GAN [paper] | 9.98 ± 0.21 |
| Mean Teacher ResNet-26 | 6.28 ± 0.15 |
| Mean Teacher ResNet-26 (Our Re-implementation with 3 GPU) | 6.14 ± 0.24 |
| Adaboost Teacher ResNet-26 (Our Re-implementation with 3 GPU) | 6.05 ± 0.12 |
| All labels, state of the art [paper] | 2.86 |
This code is largely borrowed from Mean Teacher for a fair comparison. We really appreciate the author to make the implementation open-source!