DDC-transfer-learning
A simple implementation of Deep Domain Confusion: Maximizing for Domain Invariance which is inspired by transferlearning. The project contains Pytorch code for fine-tuning Alexnet as well as DDCnet implemented according to the original paper which adds an adaptation layer into the Alexnet. The office31 dataset used in the paper is also used in this implementation to test the performance of fine-tuning Alexnet and DDCnet with additional linear MMD loss.
Run the work
- Run command
python alextnet_finetune.pyto fine-tune a pretrained Alexnet on office31 dataset with full-training. - Run command
python DDC.pyto fine-tune a pretrained Alexnet on office31 dataset with full-training.
Experiment Results
Here we have to note that full-training protocol, which is taking all the samples from one domain as the source or target domain, and dowm-sample protocol, which is choosing 20 or 8 samples per category to use as the domain data, are quite different data preparation methods with different experiment results.
| Methods | Results (amazon to webcame) |
|---|---|
| fine-tuning Alexnet (full-training) in Pytorch | Around 51% |
| DDC ( pretrained Alexnet with adaptation layer and MMD loss) in Pytorch | Around 56% |
Future work
- Write data loader using down-sample protocol mentioned in the paper instead of using full-training protocol.
- Considering trying a tensorflow version to see if frameworks can have a difference on final experiment results.
Reference
Tzeng E, Hoffman J, Zhang N, et al. Deep domain confusion: Maximizing for domain invariance[J]. arXiv preprint arXiv:1412.3474, 2014.