Training a model (Encoder and a Classifier) with a standard decoder capable of compressing image and using the latent representation for classification
In this model we took the weight of the decoder in High Fidelity Generative Image Compression and we trained both the Encoder and the Classifier. with a loss function that combines both the restoration loss and the classificatino loss. We used CIFAR-10 as our dataset.
The global architecture can be shown in the following figure:
