Autoencoding

Autoencoding models encode an input image into a compresed embedding, in such a way that this embedding contains information relevant for reconstructing the original input through the decoder part of the model.
The model outputs 2 tensors: an embedding (which could be used by downstream tasks) and a reconstruction of the input.
If noise is added to the inputs during training (but not the targets), the model can learn to de-noise images.
This is a self-supervised task.

Dataset format

Datasets follow this structure:

endpoint_url/bucket
├── prefix/images/
└── prefix/metadata.yaml

Dataset images are placed directly inside images/ (subdirectories are ignored).
The metadata file looks like this:

task: autoencoding