1) Project Title

Mini Inception Network — Custom CNN Architecture in TensorFlow/Keras

---

2) Problem Statement and Goal of Project

This project implements a compact version of the Inception architecture in TensorFlow/Keras for educational purposes. The goal is to:

• Understand and implement parallel convolutional branches.
• Learn how to design multi-path CNN modules.
• Train a miniaturized Inception-style model on CIFAR-10.

---

3) Solution Approach

The notebook builds a Mini Inception Network from scratch:

1. Inception Module:

   • Parallel paths with:

     • 1x1 convolutions
     • 3x3 convolutions
     • 5x5 convolutions
     • 3x3 max pooling followed by 1x1 convolution

   • Concatenation of feature maps along the channel axis.

2. Model Architecture:

   • Input layer for 32×32×3 images.
   • One or more Mini Inception modules.
   • Global Average Pooling → Dense softmax output.

3. Training:

   • Dataset: CIFAR-10 (via tf.keras.datasets).
   • Normalization by dividing pixel values by 255.0.
   • One-hot encoded labels with tf.keras.utils.to_categorical.
   • Loss: categorical_crossentropy
   • Optimizer: adam

4. Evaluation:

   • Model compiled and trained for a limited number of epochs for demonstration.

---

4) Technologies & Libraries

• Python
• TensorFlow / Keras
• NumPy
• Matplotlib (for plotting training history)

---

5) Description about Dataset

• CIFAR-10:

  • 60,000 images (32×32 RGB), 10 classes.
  • Training set: 50,000 images.
  • Test set: 10,000 images.
  • Classes: airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck.

• Preprocessing in code:

  • Normalization to [0, 1]
  • One-hot encoding of labels.

---

6) Installation & Execution Guide

Prerequisites

• Python 3.x
• TensorFlow

Install (pip)

pip install tensorflow numpy matplotlib

Run

1. Open Mini Inception Network.ipynb in Jupyter/VS Code.

2. Run all cells sequentially.

3. The notebook will:

   • Load and preprocess CIFAR-10.
   • Build the Mini Inception model.
   • Train and evaluate the model.
   • Optionally plot training/validation accuracy and loss.

---

7) Key Results / Performance

• Not provided — This notebook focuses on demonstrating architecture design and implementation rather than achieving state-of-the-art accuracy.

---

8) Screenshots / Sample Output

• Not provided in repository — Running the notebook will generate model summary and training history plots.

---

9) Additional Learnings / Reflections

• This project reinforces the concept of parallel convolutional branches from Google’s Inception networks.
• Demonstrates how to implement modular blocks that can be reused in deeper architectures.
• Shows how even simplified versions of complex networks can be valuable for learning and prototyping.
• The limited training setup is intentional to prioritize architecture demonstration over exhaustive training.

---

👤 Author

Mehran Asgari Email: [email protected] GitHub: https://github.com/imehranasgari

---

📄 License

This project is licensed under the Apache 2.0 License – see the LICENSE file for details.

---

💡 Some interactive outputs (e.g., plots, widgets) may not display correctly on GitHub. If so, please view this notebook via nbviewer.org for full rendering.