Some updates along with the official publication of the following paper:
Deep Learning on Small Datasets without Pre-Training using Cosine Loss.
Björn Barz and Joachim Denzler.
IEEE Winter Conference on Applications of Computer Vision (WACV), 2020.
New features
- Support for random rotations and color distortions as data augmentation.
- Added ResNet-101 and ResNet-152 architectures supported in
keras-applications >= 1.0.7. - Interface to MIT67Scenes dataset.
- CLI argument
--top_k_accfor computing top-k accuracy.
Bug fixes
- Force L2-normalization of approximate low-dimensional class embeddings.
- Improved compatibility of
CifarGeneratorwith different pickle formats