Mindsight Challenge – Dress vs Sneaker Classifier 👗👟

This project implements a complete end-to-end machine learning pipeline for binary image classification using FashionMNIST.
The model distinguishes Dress (label 3) from Sneaker (label 7) using a custom CNN, an inference API, and evaluation metrics.

---

🚀 Project Structure

mindsight-challenge/
│── src/
│   │── train.py
│   │── model.py
│   │── infer.py
│   │── utils.py
│
│── saved_models/
│   └── fashion_mnist_cnn.pth
│
│── dataset_samples_for_inference/
│   ├── dress_0.png
│   ├── dress_1.png
│   ├── sneaker_0.png
│   ├── sneaker_1.png
│
│── 1_data_analysis.py
│── 2_prepare_inference_samples.py
│── 3_compute_metrics.py
│── app.py
│── requirements.txt
│── Dockerfile
│── README.md

---

📊 1. Dataset Analysis

Run:

python3 1_data_analysis.py

This script:

• Downloads FashionMNIST
• Prints dataset size
• Shows class distribution
• Visualizes sample images (train + test)
• Confirms that:
  • Dress = label 3
  • Sneaker = label 7

---

🎨 2. Training the Model (Binary Classifier)

Run:

python3 src/train.py

Training pipeline:

• Filters only classes 3 (Dress) and 7 (Sneaker)
• Remaps labels:
  • Dress → 0
  • Sneaker → 1
• Trains a lightweight CNN
• Achieves > 99% accuracy
• Saves trained model in:

saved_models/fashion_mnist_cnn.pth

---

🧪 3. Preparing Sample Images for Inference

Run:

python3 2_prepare_inference_samples.py

This script extracts real test samples from FashionMNIST and stores them inside:

dataset_samples_for_inference/
    dress_0.png
    dress_1.png
    sneaker_0.png
    sneaker_1.png

You can upload these directly into Swagger UI.

---

📈 4. Computing Metrics (Accuracy, F1-score)

Run:

python3 3_compute_metrics.py

This script evaluates the trained model using the filtered test set
(Dress vs Sneaker only) and prints:

• Accuracy
• Precision
• Recall
• F1-score
• Confusion Matrix

A full report is also saved as:

metrics_report_binary.txt

---

🌐 5. REST API Inference (FastAPI + Swagger UI)

Start the API:

uvicorn app:app --reload

Then open:

👉 http://127.0.0.1:8000/docs

You can upload any image (PNG/JPG).
The API will:

• Convert to grayscale
• Resize to 28×28
• Normalize
• Predict "Dress" or "Sneaker"

🔍 Example — Dress Prediction

👟 Example — Sneaker Prediction

---

💻 6. Terminal Inference

python3 src/infer.py dataset_samples_for_inference/sneaker_0.png

---

🐳 7. Docker Support

Build:

docker build -t mindsight-api .

Run:

docker run -p 8000:8000 mindsight-api

---

📦 Requirements

fastapi
uvicorn
torch
torchvision
pillow
numpy
python-multipart
scikit-learn
matplotlib

Install all with:

pip install -r requirements.txt

---

✨ Author

Sofia Brunori — Mindsight Challenge Submission