🌸 Grad-CAM Visualization Demo (TF-Flowers CNN)

🔍 Overview

This project shows how to use Grad-CAM (Gradient-weighted Class Activation Mapping) to visualise where a CNN looks when it makes a prediction. The demo trains a small CNN on the TensorFlow Flowers dataset (5 flower classes) and overlays Grad-CAM heatmaps on top of the input images.

Grad-CAM helps explain why the model classified an image the way it did — a key tool in Explainable AI (XAI).

🎓 Originally built for the CS42 Explainable AI unit. Designed to give a hands-on intuition for what convolutional layers actually attend to.

---

📦 Features

• Loads and preprocesses the TF-Flowers dataset (5 classes: daisy, dandelion, roses, sunflowers, tulips)
• Builds a CNN for image classification
• Computes Grad-CAM heatmaps for any prediction
• Visualises original image + heatmap overlay side-by-side
• Self-contained Jupyter / Colab notebook

---

🌼 Dataset — TensorFlow Flowers

Property	Details
Source	tensorflow_datasets → tf_flowers
Classes	daisy, dandelion, roses, sunflowers, tulips
Samples	~3,670
Task	Multi-class image classification

---

🖼 Sample Output

Sample of the dataset:

Grad-CAM heatmap output:

---

🧰 Requirements

Install the dependencies:

pip install -r requirements.txt

Pinned versions:

• tensorflow>=2.12
• tensorflow-datasets
• numpy
• matplotlib
• scikit-learn

---

⚙️ How to Run

Option 1 — Jupyter (local)

jupyter notebook Gradcam_tf_Flowers.ipynb

Option 2 — Google Colab

1. Open Colab → File → Upload notebook
2. Pick Gradcam_tf_Flowers.ipynb
3. Run all cells (GPU recommended: Runtime → Change runtime type → GPU)

---

📁 Project Structure

Grad-CAM-Visualization-Demo-TF-Flowers-CNN-/
├── Gradcam_tf_Flowers.ipynb   # Main notebook (training + Grad-CAM)
├── Output.png                 # Saved Grad-CAM example
├── Sample images.png          # Dataset preview
├── requirements.txt           # Python dependencies
├── LICENSE
└── README.md

---

🧠 How Grad-CAM Works (in 30 seconds)

1. Pick a target class (e.g., the predicted class).
2. Compute the gradient of the target class score with respect to the activations of the last convolutional layer.
3. Average those gradients spatially → these are the importance weights for each feature map.
4. Multiply each feature map by its weight, sum them all → you get a 2-D heatmap.
5. Upsample the heatmap to the input image size and overlay it.

The brighter the region, the more it influenced the prediction.

---

🎓 Learning Objectives

1. Understand what a CNN's last convolutional layer "looks at"
2. Compute gradients with tf.GradientTape
3. Implement and apply Grad-CAM end-to-end
4. Critically evaluate model decisions (XAI mindset)

---

🔗 Related

• Original Grad-CAM paper — Selvaraju et al., 2016: "Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization"
• Keras Grad-CAM tutorial: https://keras.io/examples/vision/grad_cam/

---

📄 License

MIT — Credits to CS42.org.