Adaptive Entropy-Guided Universal Domain Adaptation (AEG-UDA)

Overview

Adaptive Entropy-Guided Universal Domain Adaptation (AEG-UDA) introduces a novel approach to address universal domain adaptation challenges, focusing on four specific scenarios:

1. Open-Set Domain Adaptation (ODA)
2. Open Partial Domain Adaptation (OPDA)
3. Closed Domain Adaptation (CDA)
4. Partial Domain Adaptation (PDA)

This framework is designed to compare our model, AEG-UDA, with other state-of-the-art models using:

1. A unique Dynamic Adaptive Threshold.

2. An Entropy-Guided Refinement Pseudo-Labeling Strategy.

3. A novel Dynamic Rejection Loss to handle domain overlaps and unknown classes effectively.

• AEG-UDA is tested against leading benchmarks in domain adaptation, leveraging its robust mechanisms to balance performance across diverse tasks.

---

Results

Results for the emulated DANCE baseline and our AEG-UDA approach can be found in the results folder. This folder contains:

1. Results for DANCE emulated runs categorized by adaptation scenario (ODA, OPDA, CDA, PDA).
2. Results for AEG-UDA, categorized similarly for direct comparisons.

Additionally, the Testing Other Adaptation Models folder includes implementations and testing for other models such as:

• SF-UDA
• DANN
• UAN

These models serve as benchmarks for evaluating AEG-UDA in diverse adaptation scenarios.

---

Dataset Preparation

To begin, download the Office-31 dataset, which is required for all experiments. The dataset can be obtained from the following link:
Office-31 Dataset.

Prepare the dataset in the following directory structure:

data/ ├── amazon/Images/ ├── dslr/Images ├── webcam/Images

---

Once downloaded:

1. Place the zipped dataset in your Google Drive.
2. Ensure it is accessible during training by mounting your Google Drive in Colab.

---

How to Run

1. Open the corresponding file in Google Colab. This can be:

   • DANCE emulated script.
   • AEG-UDA script.

2. Mount your Google Drive to access the dataset:

   from google.colab import drive
   drive.mount('/content/drive')

3. Connect to a GPU and Run the Cells Sequentially

--

1. ODA : !sh script/run_office_obda.sh 0 /content/DANCE/configs/office-train-config_ODA.yaml
2. OPDA : !sh script/run_office_opda.sh 0 /content/DANCE/configs/office-train-config_OPDA.yaml
3. CDA : !sh script/run_office_cls.sh 0 /content/DANCE/configs/office-train-config_CDA.yaml
4. PDA : !sh script/run_office_cls.sh 0 /content/DANCE/configs/office-train-config_PDA.yaml