Implement a simple recurrent model which can make use of "rotation" and "affine transformation" layers to iteratively preprocess the MNIST inputs into something more suitable for recognition.