Closed marued closed 3 years ago
It's possible to same checkpoints using hubs where model signatures were added to have access to the different projection heads. By default loading from a hub will yield no projection heads on top of the last pooling layer. Otherwise, it's possible to use 'full-projection' or 'projection-head-1' to use different projection heads just like it's specified in SimCLR v2.
example:
module = hub.Module(hub_path)
module(x1, signature='projection-head-1')
Figuring out the model outputs and module hubs to be able to choose which linear head layer we want to use.