NNCLR

NNCLR#

class stable_pretraining.methods.NNCLR(encoder_name: str | Module = 'vit_small_patch16_224', projector_dims: Sequence[int] = (2048, 256), predictor_hidden_dim: int = 4096, queue_length: int = 16384, temperature: float = 0.1, low_resolution: bool = False, pretrained: bool = False)[source]#

Bases: Module

NNCLR: SimCLR with a nearest-neighbour queue.

Parameters:

  • encoder_name – timm model name or pre-built nn.Module.

  • projector_dims – (hidden, output) for the projector (default (2048, 256)).

  • predictor_hidden_dim – Predictor hidden dim (default 4096).

  • queue_length – Number of past projections to keep for the NN lookup (default 16384).

  • temperature – NT-Xent temperature (default 0.1).

  • low_resolution – Adapt first conv for low-res input.

  • pretrained – Load pretrained timm weights.

forward(view1: Tensor, view2: Tensor | None = None) → NNCLROutput[source]#

Same as torch.nn.Module.forward().

Parameters:

  • *args – Whatever you decide to pass into the forward method.

  • **kwargs – Keyword arguments are also possible.

Returns:

Your model’s output