Image Classifier with TensorFlow (Transfer Learning)

End-to-end image classification project using TensorFlow and MobileNetV2.
This project demonstrates a complete deep learning workflow: data ingestion, preprocessing, baseline modeling, transfer learning, evaluation, model export, and inference.

---

Project Overview

Objective:
Build an accurate image classifier using a pretrained convolutional neural network and evaluate its performance on unseen data.

Key Techniques:

• Convolutional Neural Networks (CNNs)
• Transfer Learning with MobileNetV2 (ImageNet weights)
• tf.data pipelines for efficient input processing
• Model evaluation with accuracy, confusion matrix, and classification report
• Exporting models for reuse and inference

Dataset:
CIFAR-10 (10 image classes, loaded via tensorflow_datasets)

---

Repository Structure

image-classifier-tensorflow/
│
├── notebooks/
│   └── 01_image_classifier_end_to_end.ipynb
│
├── src/
│   └── inference.py
│
├── models/
│   ├── keras/                # exported .keras models (ignored by git)
│   └── class_names.json
│
├── reports/
│   ├── figures/              # confusion matrix, training curves, prediction grids
│   └── metrics/              # classification report + metrics JSON
│
├── requirements.txt
├── .gitignore
└── README.md

---

Workflow Summary

Data Loading

• CIFAR-10 loaded using tensorflow_datasets
• Train / validation / test splits created

Preprocessing

• Resize images to model input size
• Normalize pixel values to [0, 1]
• Build efficient tf.data pipelines

Baseline Model

• Lightweight CNN trained from scratch
• Used as a performance reference

Transfer Learning

• MobileNetV2 pretrained on ImageNet
• Backbone frozen, custom classification head trained
• Callbacks for early stopping and learning-rate scheduling

Evaluation

• Accuracy on validation and test sets
• Confusion matrix
• Precision / recall / F1-score per class

Export & Inference

• Model exported in .keras format
• CLI inference script for predictions on new images

---

Results (CIFAR-10)

Transfer Learning – MobileNetV2 (Frozen Backbone)

• Validation Accuracy: high 80s / low 90s (varies by run)
• Test Accuracy: comparable to validation

Saved Artifacts

• Confusion Matrix:
  reports/figures/confusion_matrix_mobilenetv2_frozen.png

• Training Curves:

  • reports/figures/mobilenetv2_frozen_accuracy.png
  • reports/figures/mobilenetv2_frozen_loss.png

• Classification Report:
  reports/metrics/classification_report_mobilenetv2_frozen.txt

---

How to Run

1. Set up environment

python -m venv .venv
.venv\Scripts\activate   # Windows
pip install -r requirements.txt

2. Run the notebook (training + evaluation)

Open and run:

notebooks/01_image_classifier_end_to_end.ipynb

This notebook:

• trains the baseline and transfer learning models

• evaluates performance

• exports the trained model

• saves all figures and metrics

---

Run Inference (CLI)

Use the exported .keras model to run predictions from the command line.

python src/inference.py \
  --model models/keras/<your_model>.keras \
  --num 9 \
  --save-grid reports/figures/pred_grid.png

What this does

• loads the trained model

• runs inference on CIFAR-10 test images

• prints predictions + confidence scores

• saves a grid of predicted images (optional)

Technologies Used

• Python

• TensorFlow / Keras

• TensorFlow Datasets

• NumPy

• Matplotlib

• scikit-learn

Next Improvements

• Fine-tune upper MobileNetV2 layers for additional accuracy gains

• Add custom dataset support for real-world images

• Package inference as a lightweight API or web demo

• Experiment with alternative pretrained backbones (EfficientNet, ResNet)

Author

Juante Wilson Aspiring Applied Data Scientist / Machine Learning Engineer Focused on end-to-end ML projects and AWS-centric workflows