Feature request: allow users to freeze model weights when training a classifier #272
Description
At the moment, the freeze=False option in the _initialize method of the StaticModelForClassification class is hardcoded, allowing the embedding model's weights to be updated during training. It would be beneficial for certain use cases to freeze the weights instead and only train the classifier's MLP head.
For example, embeddings of the documents to be classified could be pre-computed using the original embedding model elsewhere in a production system (e.g. for other purposes too, not only for classification). The classification part of the system would then only use the trained MLP head head based on such pre-computed embeddings, making that part extremely light-weight and efficient.