https://github.com/davidemodolo/homm-domainadaptation

Project for "Deep Learning" course at UNITN - Unsupervised Domain Adaptation on Adaptiope Dataset
https://github.com/davidemodolo/homm-domainadaptation

adaptiope computer-vision deep-learning domain-adaptation resnet-34 unsupervised-domain-adaptation

Last synced: 10 months ago
JSON representation

Project for "Deep Learning" course at UNITN - Unsupervised Domain Adaptation on Adaptiope Dataset

• Host: GitHub
• URL: https://github.com/davidemodolo/homm-domainadaptation
• Owner: davidemodolo
• Created: 2025-04-28T08:41:54.000Z (about 1 year ago)
• Default Branch: main
• Last Pushed: 2025-04-28T08:42:08.000Z (about 1 year ago)
• Last Synced: 2025-08-22T02:59:47.005Z (11 months ago)
• Topics: adaptiope, computer-vision, deep-learning, domain-adaptation, resnet-34, unsupervised-domain-adaptation
• Language: Jupyter Notebook
• Homepage:
• Size: 7.43 MB
• Stars: 1
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# HoMM Domain Adaptation

This project implements an unsupervised domain adaptation model based on the paper "[HoMM: Higher-order Moment Matching for Unsupervised Domain Adaptation](https://arxiv.org/pdf/1912.11976.pdf)". The method extends standard discrepancy-based losses (MMD, CORAL) by matching higher-order moments in the feature space.

## Overview

- **Architecture:**  

  The model uses a ResNet34 backbone with a custom adapted layer (using tanh activation instead of relu) to extract features. An added classification layer produces the final predictions. Two loss components are used:

  - **Domain discrepancy loss (HoMM loss):**  

    This loss matches higher-order statistics between source and target domains. Several versions are implemented, including direct 3rd-order, grouped 4th-order, and an arbitrary-order variant via random sampling.

  - **Discriminative clustering loss:**  

    This loss enforces that pseudo-labeled target samples become closer to their respective class centers. The centers are updated with a moving average.

- **Data:**  

  A custom `SubsetImageFolder` class is used to load only a subset of classes from the dataset. Two domains are considered, for example, `product_images` and `real_life`.

- **Training:**  

  Training is performed with two modes:

  - A full UDA training step that combines the classification, HoMM, and clustering losses.

  - A baseline training step that uses cross-entropy loss only.

  Hyperparameters such as batch size, learning rate, HoMM order, and lambda factors can be tuned. For instance, the script uses a batch size of 128, a learning rate of 0.001, and lambda values to weight the loss contributions.

## Dependencies

- Python with PyTorch and torchvision

- tqdm

- Google Colab (for drive mounting if running in Colab)

- matplotlib (for plotting training curves)

## How to Run

1. **Preparation:**

   - Download the dataset (e.g., `Adaptiope.zip`) and unzip it.

   - Adjust the `img_root` and subset names as needed.

2. **Running the Experiment:**

   - To run the full UDA model, execute the notebook cells (or run the following via a main function):

     ```python

     main()

     ```

   - To run the baseline model, call:

     ```python

     main(baseline=True)

     ```

   - To reverse the domains, use:

     ```python

     main(reverse=True)

     ```

   - Both reverse and baseline options can be combined:

     ```python

     main(reverse=True, baseline=True)

     ```

3. **Monitoring:**

   - The training progress is printed and accuracy and loss plots are generated at the end of training.

## Hyperparameters

Some key hyperparameter settings in the main function:

- `batch_size=128`

- `homm_order=4`

- `num_samples=350000`

- `lambda_d=100` (discrepancy loss weight)

- `lambda_dc=0.1` (clustering loss weight)

- `alpha=0.7` for center updates

- `threshold=0.65` for selecting target samples for clustering