finetuner.tailor.paddle.projection_head module#

class finetuner.tailor.paddle.projection_head.ProjectionHead(in_features, output_dim=128, num_layers=2)[source]#

Bases: paddle.nn.Layer

Projection head used internally for self-supervised training. It is (by default) a simple 3-layer MLP to be attached on top of embedding model only for training purpose. After training, it should be cut-out from the embedding model.