Perform Differentiable Optimal Transport in Log-space for stability Adapted from SuperGlue (https://github.com/magicleap/SuperGluePre

Parse the raw outputs (losses) of the network. Args: losses (dict): Raw output of the network, which usually contain

make all ReLU inplace

xavier initialize all params

This function disables printing when not in the main process

Sample negative samples.

Sample positive samples.

Args: Same as in :meth:`__init__()`. Returns: fig (matplotlib.pyplot.figure): top level container for all th

Sinkhorn Normalization in Log-space as matrix scaling problem. Regularization strength is set to 1 to avoid manual checking for numer

Alternative to optimal transport :param attn: raw attention weight, [N,H,W,W] :return: updated attention weight, [N,H,W+1,W+

Transform inputs for decoder. Args: inputs (list[Tensor]): List of multi-level img features. Returns: Tensor:

Transform inputs for decoder. Args: inputs (list[Tensor]): List of multi-level img features. Returns: Tensor:

Transform inputs for decoder. Args: inputs (list[Tensor]): List of multi-level img features. Returns: Tensor:

Upsample the raw prediction to full resolution :param x: input data :param disp_pred: predicted disp at low res :par

Add prefix for dict. Args: inputs (dict): The input dict with str keys. prefix (str): The prefix to add. Returns: d

Add prefix for dict. Args: inputs (dict): The input dict with str keys. prefix (str): The prefix to add. Returns: d

compute weighted sum of loss :param loss_dict: dictionary of losses

compute weighted sum of loss :param loss_dict: dictionary of losses

Assign boxes to either a ground truth boxes or a negative boxes.

author: Qian Zhou This function is to support both SDPA and FlashAttention based on input dtype. if input dtype is fp32, it will use SDPA

Placeholder for augmentation test.

Test with augmentations. Only rescale=True is supported.

This function computes an assignment between the targets and the predictions of the network. The targets include null (=0) disparity. We do a 1-

Wrapper for grid_sample, uses pixel coordinates

Args: mask_preds (Tensor): Mask prediction in shape (N1, H, W). gt_masks (Tensor): Ground truth in shape (N2, H, W)

Build Assigner.

@refimg_fea: left image feature @targetimg_fea: right image feature

Builder of box sampler.