The model was trained for the task of orientation classification. The model was trained on CIFAR-100 dataset which contains 60000 images covering 600 classes of 32x32 RGB images.

Data was split to 50000 train samples and 10000 test samples.

The training of the model on this dataset using Adam optimizer resulted in 100% validation accuracy.

This model can be considered as a "toy" model as CIFAR-100 image size is a huge disadvantage since everyday images are usually much larger. Models should be trained on better datasets with better resources.

The bottleneck of this model is that images of a class to be classified should be added to the training loop if we need to correctly classify their orientation. Although the model was not tested on classes different that those used in the dataset, the model should encode what it means to be "upright" and "upside-down". I think variants of the few-shots learning methods may be adapted for this type of problems; however, I do not know of the specifics of such approach.