Glossary of terms used in this Documentation and in the Geospatial AI area#

Encoder#

The neural network used to map between the inputs and the intermdiary stage (usually referred as embedding or sometimes as latent space) of the forward step. The encoder is also frequently called backbone and, for finetuning tasks, it is usually the part of the model which is not updated/trained.

Decoder#

The neural network employed to map between the intermediary stage (embedding/latent space) and the target output. For finetuning tasks, the decoder is the most essential part, since it is trained to map the embedding produced by a previoulsy trained encoder to a new task.

Head#

A network, usually very small when compared to the encoder and decoder, which is used as final step to adapt the decoder output to a specific task, for example, by applying a determined activation to it.

Neck#

Necks are operations placed between the encoder and the decoder stages aimed at adjusting possible discrepancies, as incompatible shapes, or applying some specific transform, as a normalization required for the task being executed.

Factory#

A Factory is a class which organizes the instantiation of a complete model, as a backbone-neck-decoder-head architecture. A class is intended to receive lists and dictionaries containing the required arguments used to build the model and returns a new instance already ready to be used.