practical_work_04.md

Practical work 4

CIFAR-10

Create a convolutional neural network to complete the CIFAR-10 task. The neural network have to be:

• Modular (using multiple nn.Module for different parts)
• Use dropout to prevent overfitting
• Use data augmentation to improve performances
• Trained with a scheduled decrease of its learning rate

Transfer learning

Once you are done with the neural network of the previous section, you can go on and apply the Transfer learning for Computer Vision Tutorial of the PyTorch documentation. Be careful to understand every step of the tutorial.

To get the dataset in Google Colab, you can paste and run this code and it will download and extract the data required for this model.

!wget https://download.pytorch.org/tutorial/hymenoptera_data.zip
!mkdir data
!unzip -q hymenoptera_data.zip -d data

The model used as a base in this tutorial is a deep residual neural network. You can complete the tutorial without knowing how it works but getting an intuition of the concept of skip connection will be useful for following courses.

This figure and the description of the model architecture are available at here: Deep Residual Learning for Image Recognition.