Handling Class Imbalance In Domain Adaptation

This repository contains the codes for one of my projects on Domain Adaptation at Image Analysis and Computer Vision Lab.

Brief Description

In this repository, I have codes to run experiments on standard Semi-Supervised Domain Adaptation (SSDA) benchmarks such as DomainNet, Office-31 and Office-Home Datasets. There are also additional experiments for artificially imbalanced step and exponential imbalanced datasets.

Steps to run the code

Minimax Entropy

1. This repository is built on top of SSDA_MME [1].
2. Download (one of) the datasets using bash download_data.sh. Modify this file appropriately to download the desired dataset.
3. To run the entropy maximization setting for Domain Adaptation, run:

python main.py
--method MME # options: {ENT, S+T}
--dataset multi # options: {office, office_home}
--source real # source domain
--target sketch # target domain
--net resnet34 # network architecture
--attribute glove_50 # semantic initialization of prototypes
--dim 50 # dimensions of semantics
--loss CBFL # options: {CE, CBFL, FL}
--alpha 1 # CBFL parameter
--beta 0.999 # another cbfl parameter
--deep 1 # option to set classifier type
--patience 10 # early stopping patience
--mode train # train or inference mode
--save_check # save checkpoint

4. For a more detailed description of each of the parameters please see ./main.py

Target Reweighting

To run the second stage of training i.e., target reweighting, please run python main_stagetwo.py --pretrained /path/to/model_ckpt, by passing the saved checkpoint from stage 1 as the parameter.

Miscelleneous

To plot t-SNE plots of the feature representations please navigate to the ./tsne folder.

Dependencies

• pytorch
• scikit-learn
• tsne-cuda

References

1. Semi-Supervised Domain Adaptation vie Minimax Entropy. Kuniaki Saito, Donghyun Kim, Stan Sclaroff, Trevor Darrell, Kate Saenko. ICCV 2019.