Returns: bool: whether a checkpoint exists in the target directory.

Inference interface for the model for PIL image Args: pil_img (PIL.Image.Image): input PIL image pad_input (b

ViT weight initialization, original timm impl (for reproducibility)

ViT weight initialization, original timm impl (for reproducibility)

Inverse attractor: dc = dx / (1 + alpha*dx^gamma), where dx = a - c, a = attractor point, c = bin center, dc = shift in bin center This is the def

Distributed synchronization of the metric Warning: does not synchronize the deque!

Expects a batch of images and depth as input batch["image"].shape : batch_size, c, h, w batch["depth"].shape : batch_size, 1,

Expects a batch of images and depth as input batch["image"].shape : batch_size, c, h, w batch["depth"].shape : batch_size, 1,

Unfreezes all the parameters of all the ModuleDicts children that are exclusively used for the given bin configuration

Close to ViT-giant, with embed-dim 1536 and 24 heads => embed-dim per head 64

Close to ViT-giant, with embed-dim 1536 and 24 heads => embed-dim per head 64