The classic DANN paper, ported to PyTorch 1.0

What it does

This repo is a straightforward PyTorch reimplementation of the Unsupervised Domain Adaptation by Backpropagation paper—better known as DANN. It trains a feature extractor, label predictor, and domain classifier in an adversarial loop to nudge a model learned on MNIST toward the visually noisier MNIST-M target domain. The code follows the paper’s original architecture.

The interesting bit

The value here is archival fidelity, not novelty: you get a minimal, readable reference implementation of a foundational domain-adaptation architecture. If you want to see how adversarial domain alignment works without wading through a full framework, this is a compact sandbox.

Key highlights

• Faithful PyTorch 1.0 implementation of the original DANN paper
• Demonstrates the standard MNIST → MNIST-M synthetic-to-real shift
• README includes an architecture diagram showing the network structure
• Author maintains a separate Python 3 / Docker fork for modern environments

Caveats

• Requires manual dataset download from Baidu Drive or Google Drive before training
• Locked to PyTorch 1.0 and Python 2.7, both well past end-of-life
• README omits accuracy numbers, training curves, or comparisons to published results

Verdict

Worth a look if you need a concise, working reference for the original DANN architecture. Look elsewhere if you want a modern, batteries-included toolkit or state-of-the-art domain-adaptation methods.