progressive-resizing

This repository contains the code for building a convolutional neural network machine learning classifier in three parts. It is the companion repo for the article "Boost your CNN performance with progressive resizing in Keras".

The three components are:

• A first model that works on 48x48 images.
• A second model that works on 96x96 images.
• A final model that works on 192x192 images.

The resultant model is a "three layer cake": each larger-scale model subsumes the previous smaller-scale model layers and weights in its architecture.

This approach is meant to demonstrate a workflow and technique for building neural networks known as "progressive resizing". Progressive resizing has been used to good effect by Jeremy Howard, who used to achieve a top 10% finish in the Planet Kaggle Competition, and he uses it throughout his fast.ai course "Practical Deep Learning for Coders".

The data

This project uses the "Open Fruit" dataset, a dataset of fruit images taken from Google's Open Images Dataset.

You can download the data for yourself from the source data package on Quilt T4:

$ pip install -U t4
$ python -c "import t4; t4.Package.install('s3://quilt-example', 'quilt/open_fruit', './')"

Alternatively, to build this data from source again (warning: this takes a long time!):

$ git clone https://github.com/quiltdata/open-images.git
$ cd open-images/
$ conda env create -f environment.yml
$ source activate quilt-open-images-dev
$ pip install -e ./src/openimager/
$ python -c "import openimager; openimager.download(['Apple', 'Grape', 'Orange', 'Pomegranate', 'Banana', 'Grapefruit', 'Peach', 'Strawberry', 'Cantaloupe', 'Lemon', 'Pear', 'Tomato', 'Common fig', 'Mango', 'Pineapple', 'Watermelon'])"

The run the code in the build-dataset.ipynb notebook to generate the final cropped images.

To learn more about using Google Open Images for building new datasets, see "How to classify photos in 600 classes using nine million Open Images".

The models

The build-models.ipynb notebook is the notebook where I work through the model definitions. The actual model builds were executed using the fahr remote training CLI, which I have been building side-by-side with this project. To model resources are in the various resnet* folders in this repository.

Note that you can execute setup.sh to install the required resources (a recent version of keras-preprocessing, fahr, etc.).