Skip to content

Latest commit

 

History

History
178 lines (122 loc) · 6.5 KB

README.md

File metadata and controls

178 lines (122 loc) · 6.5 KB
Raw

MNIST CNN Model: Training, Quantization, and Evaluation

This project demonstrates how to train a Convolutional Neural Network (CNN) on the MNIST dataset, quantize the trained model to INT8 using TensorFlow Lite, and evaluate the performance of the quantized model.

Project Overview

The project consists of two Python scripts:

  • model.py: Handles the creation, training, and quantization of a CNN model.
  • evaluate.py: Evaluates the performance of the quantized INT8 model.

1. model.py

This script performs the following steps:

  • Model Definition: Defines a CNN architecture tailored for MNIST digit classification.
  • Data Preparation: Loads and preprocesses the MNIST dataset, normalizing the pixel values to the range [0, 1].
  • Model Training: Trains the CNN model on the MNIST training dataset for 5 epochs.
  • Model Evaluation: Evaluates the model on the test dataset and reports the accuracy.
  • Model Saving: Saves the trained model in full precision (FP32) format as cnn_fp32_model.
  • Model Quantization: Converts the FP32 model to an INT8 model using TensorFlow Lite, incorporating a representative dataset to optimize the quantization process while maintaining input and output in FP32.
  • Quantized Model Saving: Saves the quantized model as cnn_int8_model.tflite.

2. evaluate.py

This script is responsible for:

  • Model Loading: Loads the previously saved INT8 TensorFlow Lite model (cnn_int8_model.tflite).
  • Dataset Preparation: Loads and preprocesses the MNIST test dataset.
  • Model Evaluation: Runs inference on the test dataset using the quantized model and calculates the accuracy.
  • Results Reporting: Prints the accuracy of the INT8 model, allowing comparison with the original FP32 model.

Installation and Setup

1. Environment Setup

Before running the scripts, ensure your environment is properly configured.

  • Step 1: Clone the Repository
git clone https://github.com/OpenGenus/int8-training.git
cd int8-training
  • Step 2: Set Up the Environment (Install dependencies)
pip install NumPy
pip install TensorFlow

2. Running the Scripts

To train and quantize the model, run:

python3 model.py

This script will:

  • Train a CNN on the MNIST dataset.
  • Save the trained FP32 model as cnn_fp32_model.
  • Quantize the model to INT8 and save it as cnn_int8_model.tflite.

To evaluate the performance of the quantized model, run:

python3 evaluate.py

This script will:

  • Load the cnn_int8_model.tflite file.
  • Run inference on the MNIST test dataset.
  • Print the accuracy of the quantized model.

Project Structure

├── model.py         # Script for training and quantizing the CNN model
├── evaluate.py      # Script for evaluating the quantized model
└──  README.md        # Project documentation (this file)

Acknowledgements

This project leverages the TensorFlow and Keras libraries for deep learning and model quantization. The MNIST dataset is a well-known dataset of handwritten digits, widely used for training and evaluating image processing systems.

Authors

Vidhi Srivastava

Show Your Support ⭐️⭐️

If you find this project helpful or interesting, please consider giving it a star!

MNIST CNN Model: Training, Quantization, and Evaluation

This project demonstrates how to train a Convolutional Neural Network (CNN) on the MNIST dataset, quantize the trained model to INT8 using TensorFlow Lite, and evaluate the performance of the quantized model.

Project Overview

The project consists of two Python scripts:

  • model.py: Handles the creation, training, and quantization of a CNN model.
  • evaluate.py: Evaluates the performance of the quantized INT8 model.

1. model.py

This script performs the following steps:

  • Model Definition: Defines a CNN architecture tailored for MNIST digit classification.
  • Data Preparation: Loads and preprocesses the MNIST dataset, normalizing the pixel values to the range [0, 1].
  • Model Training: Trains the CNN model on the MNIST training dataset for 5 epochs.
  • Model Evaluation: Evaluates the model on the test dataset and reports the accuracy.
  • Model Saving: Saves the trained model in full precision (FP32) format as cnn_fp32_model.
  • Model Quantization: Converts the FP32 model to an INT8 model using TensorFlow Lite, incorporating a representative dataset to optimize the quantization process while maintaining input and output in FP32.
  • Quantized Model Saving: Saves the quantized model as cnn_int8_model.tflite.

2. evaluate.py

This script is responsible for:

  • Model Loading: Loads the previously saved INT8 TensorFlow Lite model (cnn_int8_model.tflite).
  • Dataset Preparation: Loads and preprocesses the MNIST test dataset.
  • Model Evaluation: Runs inference on the test dataset using the quantized model and calculates the accuracy.
  • Results Reporting: Prints the accuracy of the INT8 model, allowing comparison with the original FP32 model.

Installation and Setup

1. Environment Setup

Before running the scripts, ensure your environment is properly configured.

  • Step 1: Clone the Repository
git clone https://github.com/OpenGenus/int8-training.git
cd int8-training
  • Step 2: Set Up the Environment (Install dependencies)
pip install NumPy
pip install TensorFlow

2. Running the Scripts

To train and quantize the model, run:

python3 model.py

This script will:

  • Train a CNN on the MNIST dataset.
  • Save the trained FP32 model as cnn_fp32_model.
  • Quantize the model to INT8 and save it as cnn_int8_model.tflite.

To evaluate the performance of the quantized model, run:

python3 evaluate.py

This script will:

  • Load the cnn_int8_model.tflite file.
  • Run inference on the MNIST test dataset.
  • Print the accuracy of the quantized model.

Project Structure

├── model.py         # Script for training and quantizing the CNN model
├── evaluate.py      # Script for evaluating the quantized model
└──  README.md        # Project documentation (this file)

Acknowledgements

This project leverages the TensorFlow and Keras libraries for deep learning and model quantization. The MNIST dataset is a well-known dataset of handwritten digits, widely used for training and evaluating image processing systems.

Authors

Vidhi Srivastava

Show Your Support ⭐️⭐️

If you find this project helpful or interesting, please consider giving it a star!