A PyTorch implementation of AdaMatch: A Unified Approach to Semi-Supervised Learning and Domain Adaptation.
This implementation is heavily based off of Sayak Paul's excellent keras blog post. I also used Ang Yi Zhe's PyTorch implementation to fix some problems that my initial implementation had.
You can run this code simply by doing:
python run.py
Note that this implementation is to be used more as a starting point, and you'll have to dig a little deeper on the python code in order to change hyperparameters, transforms and the data that is being loaded.
Running the the code on a MNIST -> USPS adapation using a ResNet18 architecture:
- Accuracy on source dataset = 0.9822
- Accuracy on target dataset = 0.9561
Some differences:
- Network used: due to limited computing resources my implementation uses a ResNet18 architecture. This could be easily changed by importing other models on the
network.pyfile. - The paper uses CTAugment for strong augmentations. I implemented a pipeline similar to CTAugment using the availabe torchvision transforms.
- The original implementation uses a cosine decay for the learning rate scheduler. I tried using PyTorch's cosine-related learning rate schedulers and didn't manage to achieve good results, so I used a simple step learning rate scheduler.