In this Kaggle competition, the mission is to classify images of dogs and cats. The evaluation metric is the log loss function:
where
- n is the number of images.
- ŷi is the predicted probability of the image being a dog.
- yi is 1 if the image is a dog, 0 if cat.
My goal in this notebook was not to achieve a high accuracy (as I could do that simply by using a pretrained model from PyTorch), but to experiment with CNNs and building a deep neural network in PyTorch, as well as to work with a database not preloaded by PyTorch (unlike my work with MNIST).
Using this model I achieved 87% accuracy and 0.26 loss on the validation set, and 0.32 loss on the test set.
The model's network mainly consists of 3 convolution layers and 2 fully connnected layers, but for more details please find the notebook cats-vs-dogs-redux.ipynb in the root directory (refresh if it doesn't load), in which you'll find a complete walkthrough of this work, or head over to my Kaggle profile where you'll find the same notebook and see some of my other projects.
Notes: After implementing this network, I have read the AlexNet paper and implemented it, and I learned a few things that in hindsight I would change in my network:
- Aspect ratio should be maintained when resizing the images (i.e. resize such that the shorter edge =
X, and then crop the center). - Larger (more broad) fully connected layers at the end (would take a longer time to train, but seems to be worthy).
- Data Augmentation: Being very beneficial and easy to implement in PyTorch, I should have definitely used this tool for the sake of generalization.
- Data normalization: Should've normalized the data by the images mean and std.