TD2-DeepLearning

Here you can find my work for the TD2 - Deep Learning. This TD is dealing with Convolutional Neuron Networks (CNN) for image classification. This repository is organised as the following structure :

• In the "TD2 Deep Learning Instruction.ipynb" file, you can find the instructions of the TD.
• In the "exercice 1 and 2.ipynb", "exercice 3.ipynb", and "exercice 4.ipynb" files you can find the answer of the related exercice.
• You can find images for the training or testing different models in the folders "data" and "hymenoptera_data".
• Finally, in the README you can find the answers to specific questions of the TD.

Exercice 1 - Overfitting of the first CNN model

We are wondering if overfitting occurs with the first CNN model (2 convolutional layers, 1 pooling layer, 3 linear layers) and with n_epoch = 30.

As we can see, the training loss is decreasing over the epochs, and validation loss is decreasing until n_epoch = 16. If the training loss is decreasing and the validation loss is increasing at the same time, it means that the model is overfitting on the training set. We need to recover the model at the best validation loss, so we will keep n_epoch = 16.