📘 Project: Domain Origin Detection in CINIC-10

Can a neural network distinguish whether an image in CINIC-10 comes from CIFAR-10 or ImageNet — even when class labels are the same?

This project explores a meta-learning task using the CINIC-10 dataset. We train a binary classifier to detect the origin domain of an image (CIFAR vs. ImageNet), despite both domains sharing the same 10 class labels (e.g., airplane, dog, ship).

🔍 Goals

• Create a binary domain label (CIFAR or ImageNet) for each image.
• Train a simple CNN or MLP to predict the source domain.
• Compare with baseline methods like random guessing or pixel statistics.
• Visualize image embeddings using t-SNE to explore learned patterns.

📊 Dataset

• CINIC-10: A 270,000-image dataset mixing CIFAR-10 and downsampled ImageNet images.
• All images are 32×32 pixels and belong to one of 10 shared classes.

🧠 Methods

• Basic CNN or shallow MLP as classifiers.
• Use of pretrained features (e.g. ResNet18) for comparison.
• Optional: ROC AUC, accuracy, t-SNE visualization.

📁 Structure

data/         # CINIC-10 images and labels
notebooks/    # Jupyter notebooks for exploration and training
models/       # Trained model files
results/      # Evaluation metrics and plots

Note that we did not use results but used a build directory inside the notebooks directory to store some images.

✅ Before You Start

Before running the code, make sure to download the CINIC-10 dataset, set up the environment and run the python script to generate Labels for the Dataset.

1. Navigate to the data/ directory.

2. Follow the instructions in loading_the_dataset.md to download and extract the dataset.

3. Install the required Python packages using the provided environment file. We recommend using Mamba (a faster alternative to conda):

   mamba env create -f environment_2025.yml

4. Activate the environment before working on the project:

   mamba activate ml

5. Navigate to generate_domain_labels.py and follow the instructions provided in the comments in lines 6-8.

6. Once the environment is active and the dataset is ready, you can:

   • Run the notebooks in notebooks/
   • Load trained models from models/
   • View visualizations and metrics in results/

Everything should work out of the box if the steps above were followed correctly.

---

GPU Support (Optional)

"You are welcome to add and correct instructions for GPU Support"

To use GPU acceleration with TensorFlow, a compatible NVIDIA GPU and driver must be installed.

How to Check for GPU Availability

1. Run:

nvidia-smi

If you see:

zsh: command not found: nvidia-smi

your system does not have the NVIDIA driver installed.

2. Check for NVIDIA hardware:

lspci | grep -i nvidia

If you see output like:

01:00.0 VGA compatible controller: NVIDIA Corporation ...

you do have an NVIDIA GPU and just need the driver.

---

🐧 NVIDIA Driver Installation on Fedora 42

I did not get it to work but maybe someone can add proper instructions

---

📎 Official source: NVIDIA CUDA Downloads for Fedora

---

👥 Team

• Akram Aki, Bachelor Semester 6
• Muhammad Hassan, Master
• ML Seminar, TU Dortmund, Germany