xysteve00 / cifar10_classification

CIFAR10 with "Custom" Loss and PyTorch Dataset Classes

The goal of this work is to provide examples of how to:

1. Use the PyTorch Dataset class in PyTorch for CIFAR10, instead of simply importing it in one line. Hopefully this will help someone bridge the gap towards implementing more complex datasets.

2. Create a custom loss class which can be extended for more advanced loss functions which are not included in the standard PyTorch arsenal.

Additionally, we have included a hard-coded VGG16 for anyone that is just beginning and needs that level of readability.

At the bottom the Data.py, you will find code which allows you to visualize the images of CIFAR10. This is the only place we have used OpenCV, so if you do not wish to visualize any images, you do not need to "import cv2".

With the parameters set as is, using a VGG16, the model acheieves 92.96% test accuracy.

The project dependencies can be found in requirements.txt .