This project was completed by Hugo Hellainger, Louis Bertolotti and Nicolas Maisonneuve-Bonteil during the "Deep Learning" course of the X-HEC Joint Degree "Data Science for Business".
The goal of this project was to find the CNN trained on the CIFAR-10 dataset with the best balance of accuracy and computation time. Starting with a naïve model, we enhanced it and then compared it to a pre-trained model (InceptionResNetV2) which we adapted to our image input size.
We reached a final accuracy of 0.74 with a reasonable computation time.
The Keras and Tensorflow libraries were used to create the models. We also decided to make full use of the Tensorboard API, which collects logs on the model and helps to analyse parameters such as the number of epochs or the model accuracy.
The notebook contains a custom class called "UniversalHPOptimizer", which runs a GridSearch with given parameters and automatically returns the best model. It leverages the Tensorboard-compatible HParams data format, allowings callbacks to return information to Tensorboard.