Training a InceptionV3 Image Classification Model

This Microsoft Azure ML sample will take you through the steps of training an image classification model based on the InceptionV3 Architecture on the Azure ML platform.

What is InceptionV3?

Source: Rethinking the Inception Architecture for Computer Vision The primary focus of this architecture is to utilize fewer computational resources in order to develop models quicker. InceptionV3 is a convolutional neural network based off the InceptionV1 Architecture, InceptionV3 improves upon V1's approach by...

1. Factorizing convolutions into multiple smaller ones.
2. Replacing convolutions with smaller ones. E.g. a 5x5 filter is replaced with two 3x3. A 5x5 filter would have 25 parameters whilst the two 3x3 filters would have 18 parameters total (2 x 3 x 3 = 18).
3. Asymmetric convolutions. E.g. a 3x3 convolution is replaced by a 1x3 and 3x1 convolutions.
4. Auxiliary classifiers that propagate loss incurred between the convolutional layers.
5. Grid size reduction

Getting Started

Some instructions about how to copy the model, load data, train, and validate on Jupyter Notebooks. WIP

Prerequisites

pip install azureml-sdk

Read more detailed instructions on how to set up your environment using Azure Notebook service, your own Jupyter notebook server, or Docker.