Transfer learnings of the Inception model

Introduction

This tutorial aims at retraining a model to recognise images of flowers. We will retrain the last layer of Google's Inception-v3 model. This will be done using the science of transfer learning.

Below you will find some notes on some of the topics mentioned above.

TensorFlow

TensorFlow is an open source library for numerical computation, specializing in machine learning applications.

Transfer learning

Transfer learning allows us to leverage the knowledge gained while solving one problem and applying it to a different but related problem.

An introduction to Inception-v3

Inception v3 is the 2015 iteration of Google's Inception architecture for image recognition. Inception-v3 is trained for the ImageNet Large Visual Recognition Challenge using the data from 2012. This is a standard task in computer vision, where models try to classify entire images into 1000 different classes, like "Zebra", "Dalmatian", and "Dishwasher".

Why use transfer learning?

Before I answer the question, I would like to talk about what a pre-trained model is: Simply put, a pre-trained model is a model created by someone else, trying to solve a similar problem.

For example, if we want to recognise the image of a flower we can spend years at developing a model for it or we can leverage the knowledge of the Inception model to identify those images.

As, mentioned above Inception is a pre-trained image classifier; it would serve our purpose if we mould it to our needs.

Note: This tutorial assumes that, one is akin to the concepts of python programming, and that one knows the basics of Linux operating system.

Building our model

We will be using transfer learning, to retrain the inception network to tell apart a few different classes of flowers

This repository contains the following two scripts:

retrain.py
classify.py

Step 1: Installation

Let us first install TensorFlow on our system. Run the following command in terminal to do so.

$sudo pip install tensorflow

To check if TensorFlow has been successfully installed, run the following code as a python script.

import tensorflow as tf
hello = tf.constant('Hello, TensorFlow!')
sess = tf.Session() # this should print some warnings here, no worries
print(sess.run(hello))

If all went well, your output will be:

Hello, TensorFlow!

Step 2: Downloading dataset

We will be using a flower dataset which is readily available online. Let us create a working directory.

Do the following on your shell:

$mkdir transfer_learning && cd transfer_learning

Now let us download the flower data set.

Run the following command:

$curl -O https://goo.gl/srnxJm

once the download is complete, run this command:

$tar -xzf flower_photos.tgz

Reduce the size of the data

Optional, but it would be useful people who have a basic system. This step speeds up the training.

Since all the image file names start with digits, the following command will reduce the number of images by 70%

$ls flower_photos/roses | wc -l
$rm flower_photos/*/[3-9]*
$ls flower_photos/roses | wc -l

Similarly reduce the number of images of other flowers types too.

Step 3: Retraining the inception model

Now is the time to fetch the one of the script from my repository. Run the following command in the PWD:

$curl -O https://raw.githubusercontent.com/ujwalp15/transfer_learning/master/retrain.py

We can visualise and monitor the training process in the background by using TensorBoard. Before training, start TensorBoard by running the following command:

$tensorboard --logdir training_summaries &

Now lets start retraining the inception model to start prediction different types of flowers, run the following command in the terminal:

$python retrain.py \ # Runs the script
--bottleneck_dir=bottlenecks \ # Bottleneck directory
--how_many_training_steps=500 \ # Training steps
--model_dir=inception \ # Path to inception model
--summaries_dir=training_summaries/basic \ # Path: save summaries
--output_graph=retrained_graph.pb \ # Path: Retrained graph
--output_labels=retrained_labels.txt \ # Path: Retrained labels
--image_dir=flower_photos # Path: Dataset directory

You can google each of them, I will explain the training process in some days...

Step 4: Using the newly Trained model

Our model has been trained and the retrained graph can be found in the working directory: a file named "retrained_labels.pb".

The relevant training labels can also be found in the working directory: a file named "retrained_labels.txt".

We will now create a python script to classify images.

We will need to use the second script form my repository.

$curl -O https://raw.githubusercontent.com/ujwalp15/transfer_learning/master/classify.py

Lets us run the script,

$python classify.py target_image

where, target_image is the path to any image of flower you would like to label.

kingardor / transfer_learning