Linear classifier (1 or 4 layers) on top of a DINOv2 ViT-S/14 backbone (optionally) pretrained on the LVD-142M dataset and trained on ImageNet-1k

Enable distributed mode Args: set_cuda_current_device: If True, call torch.cuda.set_device() to set the current PyTorch CUDA

Fix random seeds.

Compute the results on all values of `self.nb_knn` neighbors from the full `self.max_k`

Args: student_output (BxD): backbone output of student

Cross-entropy between softmax outputs of the teacher and student networks. student_patch_tokens: (B, N, D) tensor teacher_pat

Cross-entropy between softmax outputs of the teacher and student networks.

Return camera intrinsics matrix K Returns: K (torch.Tensor): Camera intrinsics matrix, of shape (3, 3)

Returns: str: The latest checkpoint file in target directory.

Returns: The rank of the current process within the local (per-machine) process group.

Returns: The size of the per-machine process group, i.e. the number of processes per machine.

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

ViT weight initialization, original timm impl (for reproducibility)

This function takes the prediction and label of a single image, returns intersection and union areas for each class To compute over many

r"""Normalizes RGB image values from :math:`[0, 255]` range to :math:`[0, 1]` range. Args: rgb (torch.Tensor or numpy.ndarray): RGB image

r"""Normalizes pixel coordinates, so that each dimension (x, y) is now in the range [-1, 1]. x coordinates get mapped from [0, height-1] to [

This function takes the prediction and label of a single image, returns pixel-wise accuracy To compute over many images do: for i = range

r"""Converts a quaternion to an axis angle. Args: quat (torch.Tensor): Quaternion (qx, qy, qz, qw) (shape: :math:`* \times 4`

r"""Converts a quaternion to a rotation matrix. The quaternion should be in (x, y, z, w) format. Args: quaternion (torch.Tensor): Qua

r"""Given two 3D homogeneous transforms `trans_01` and `trans_02` in the global frame '0', this function returns a relative transform `trans_1

r"""Scales the intrinsics appropriately for resized frames where :math:`h_\text{ratio} = h_\text{new} / h_\text{old}` and :math:`w_\text{ratio} =

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

Convert the input to a scalar

r"""Applies a rotation to a tensor containing point normals. Args: normals (torch.Tensor): Normal vectors (shape: numpoints x 3)

r"""Applies a rigid-body transformation to a pointcloud. Args: pointcloud (torch.Tensor): Pointcloud to be transformed

r"""Unnormalizes pixel coordinates from the range [-1, 1], [-1, 1] to [0, `height`-1] and [0, `width`-1] for x and y respectively. Args: