Machine Learning Nanodegree

Deep Learning

Project: Image Classifier for Flowers using Tensorflow

---

Introduction

This project was part of my 'Intro to Machine Learning Nanodegree' provided by Udacity. The code and text in this project is a combination of my own work and that of Udacity.

The goal of this project was to create a tensorflow DNN model that can accurately make predictions on flower images by correctly classifying them. The dataset used to train the network is from the oxford_flowers102 available on tensorflow_datasets, which contains 102 flow categories commonly occuring in the United Kingdom.

This project is divided into two parts; training of the model in the notebook, and converting of the trained model into a command line application using predict.py.

1. The notebook contains:

• Importing of the training, test and validation sets
• Training a keras model using the training and validation sets
• Evaluating the model using the test set
• Saving of the trained keras model

2. predict.py works by:

• Accepting two positional arguments; the image_path of the image wanting to make inference on, and the saved_model from the notebook.
• It also accepts two optional arguments; top_k representing top K labels with the highest probabilities, and category_names, a json file that maps the numerical labels to flower names.
• It then returns a table with the top K labels and their probabilities. If top_k was not specified it would only return the label with the highest probability. Using category_names would return the flower names instead of the numerical labels.