{"id":2306,"date":"2019-02-19T18:44:19","date_gmt":"2019-02-19T18:44:19","guid":{"rendered":"https:\/\/www.danielpradilla.info\/blog\/?p=2306"},"modified":"2019-02-22T08:14:35","modified_gmt":"2019-02-22T08:14:35","slug":"classifying-convolutional-neural-network-keras","status":"publish","type":"post","link":"https:\/\/www.danielpradilla.info\/blog\/classifying-convolutional-neural-network-keras\/","title":{"rendered":"Classifying fruits with a Convolutional Neural Network in Keras"},"content":{"rendered":"

I followed a tutorial on Convolutional Neural Networks that left many questions unanswered. Soon I realized that the actual process of architecting a Neural Network and setting the parameters seemed to be much more experimental than I thought. It took a while to find explanations that a rookie like me could understand. Most of the posts I found gloss over the difficult details, like what NN configuration to choose, or which library to choose, or why would you use ReLU in an image-recognition NN<\/a>  and focused on easier and more consistent tasks, like setting up a Python environment.<\/p>\n

 <\/p>\n

TL;DR<\/h3>\n

Keras<\/a> allows you to build a neural network in about 10 minutes. You spend the remaining 20 hours training, testing, and tweaking.<\/p>\n

If you wish to learn how a Convolutional Neural Network is used to classify images, this is a pretty good video<\/a>.<\/p>\n

The ultimate guide to convolutional neural networks<\/a> honors its name.<\/p>\n

This example illustrates how a CNN recognizes numbers: Basic Convnet for MNIST<\/a><\/p>\n

Get the code at https:\/\/github.com\/danielpradilla\/python-keras-fruits<\/a>.<\/p>\n

Jump to the predictions<\/a>.<\/p>\n

 <\/p>\n

A simple problem<\/h3>\n

In order to learn the basics, I started a simple \u2013um, “simple”\u2013 project to build a deep neural network to correctly classify a set of 17 kinds of fruits from the Kaggle Fruits 360 dataset<\/a>.<\/p>\n

Keras<\/a> provides an easy interface to create and train Neural Networks, hiding most of the tedious details and allowing you to focus on the NN structure. It interfaces with the most popular NN frameworks: Tensorflow, CNTK, and Theano. This answered one of my first questions, “Keras vs Tensorflow, which one should you learn and when to use each one?<\/a>” \u2013short answer: learn Keras until you need Tensorflow.<\/p>\n

Keras allows you to go from idea to working NN in about 10 minutes. Which is impressive. The Keras workflow looks like this:<\/p>\n