rzarno / classify-dogs-tensorflow-aws-ground-truth

Dog images classificator build using tensorflow library and AWS Ground Truth service for dataset preparation.

Dog image classification example

Overview

This repository contain notebook with full analysis and machine learning process for dog image classification. We create model that is able to classify if dog is present on image.

sample images:

Run

install requirements: pip install -r requirements.txt Open jupyter notebook "dog_image_classification.ipynb" and run all cells.

Adjust

If you want to create model for different image dataset, replace images in data_image/images folder. Then adjust "data_images/output.manifest.json" file which contain all train set images paths and their labels.

AWS Ground truth

Dataset for training is prepared based on AWS Academy activity in course "AWS Data Engineering". This notebook is continuation af machine learning process started in activity "Processing Data for ML Activity: Labeling with SageMaker Ground Truth". In course activity AWS SageMaker Ground Trouth service is used to label images containing dog, cats and other animals and objects. Notebook presents how to train image classifier using AWS labeling task output.

more information on AWS SageMaker Ground Truth: https://aws.amazon.com/sagemaker/groundtruth/

more information on AWS Academy Data Engineering course: https://aws.amazon.com/training/awsacademy/

Used libraries:

tensorflow and kreas - for creating model convolutional neural network

matplotlip - data visualisation

opencv - image preparation and augmentation (creating larger training set by modifying existing images, this technique also help with model generalization)

numpy and pandas - for data preparation and analysis

Neural network architecture design

Used CNN architecture introduced by Nvidia team and described in this article: https://developer.nvidia.com/blog/deep-learning-self-driving-cars/