Model that given an image of a dog, it will identify an estimate of the canine’s breed. If a human image is supplied, the code will identify the resembling dog breed.
Two datasets are provided dog-dataset(8351 images) and human-dataset (13233 images).
The project mainly consists of two parts, creating a CNN from scratch and transfer learning.
12% accuracy after 30 epochs
82% accuracy after only 5 epochs! VGG16 was used and only the very last classification layer was trained, the reason behind it is because VGG16 is already able to detect dog features and our training data was very small.
- Use a different network for transfer learning, resnet for example.
- Tweak VGG16 parameters (weight initialization, dropout, ).