04_Dogs_and_cats_challenge.md

Challenge - Dogs and Cats

Guidelines

This challenge is a famous one about dogs and cats classification.

See Kaggle dataset dogs and cats here if you want to download the full dataset. For simplicity, we here provide a subsample of 1200 dogs and 1200 cats images in dogs_cats.zip.

Here you will use a very useful tensorflow tool: the image data generator. You can have a look at the documentation here. It will deal for you with the labels as well as the image resizing automatically.

Data Preparation

We propose to use only a subset of the training dataset for computing resources reasons: 1000 cats + 1000 dogs images in train, 200 cats + 200 dogs in test.

First split by hand your dataset like the following:

- data/
--- train/
----- cats/
------- cats1.jpg
------- cats2.jpg
        ...
------ dogs/
------- dogs1.jpg
------- dogs2.jpg
        ...
--- test/
----- cats/
------- cats1001.jpg
------- cats1002.jpg
        ...
----- dogs/
------- dogs1001.jpg
------- dogs1002.jpg
        ...

Data exploration

Feel free to open and display some images.

Model building

Build a Convolutional Neural Network.

Model training

Use the ImageDataGenerator to train your model on the data.

Performances estimation

Estimate the performances of your model with the accuracy metrics. Is it good enough?

Then try to improve by all means you can leverage: architecture, hyperparameters, transfer learning, data augmentation...