Here is the code of my first Convolutional Neural Network for my first Kaggle competition.
In order to clearly see the code and the outputs, I strongly recommand you to go directly on my notebook on Kaggle at this address. You will also find the dataset.
The goal of this notebook is to classify photos of animals (32x32 pixel images) with a CNN without using deeper neural networks as Resnet.
The given dataset is composed of a training set of 7200 pictures and test set of 1800 pictures(80%-20%).
Due to the bad quality of the images, I choose to resize them to 64x64 because it allowed to have more features and therefore more information at the expense of the calculation time.
In order to taste assembling, I chose to defined 5 different models, each more or less complex. After using data augmentation on the training set, I train each model and create an ensemble model which take the five first and compute the average of the probabilites for each class for a given picture.
I succeeded in obtaining 68% accuracy on the test set which is not very satisfying but is promising for the upcoming changes.
For a complete overview of my appoach of the problem, please see the report here (in french only).
Upcoming..