Computes and stores the average and current value

Combine several transformation in a serial manner

Decoupled Motion-Appearance Network

rescale the size of image and masks

stack adjacent frames into input tensors

Swin Transformer backbone. A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows` - https:/

convert value type to float

convert to torch.Tensor

sum additive noise

ResNet backbone with frozen BatchNorm.

A basic Swin Transformer layer for one stage. Args: dim (int): Number of feature channels depth (int): Depths of this stage.

A basic Swin Transformer layer for one stage. Args: dim (int): Number of feature channels depth (int): Depths of this stage.

Multilayer perceptron.

Multilayer perceptron.

Image to Patch Embedding Args: patch_size (int): Patch token size. Default: 4. in_chans (int): Number of input image channels. De

Video to Patch Embedding. Args: patch_size (int): Patch token size. Default: (2,4,4). in_chans (int): Number of input video chan

randomly modify the contrast of each frame

Randomly horizontally flip the video volume

Swin Transformer backbone. A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows` - https:/

Swin Transformer Block. Args: dim (int): Number of input channels. num_heads (int): Number of attention heads. window_siz

Swin Transformer Block. Args: dim (int): Number of input channels. num_heads (int): Number of attention heads. window_si

transpose the image and mask

A wrapper which allows using Video-Swin Transformer as a temporal encoder for MTTR. Check out video-swin's original paper at: https://arxiv.o

Window based multi-head self attention (W-MSA) module with relative position bias. It supports both of shifted and non-shifted window. Args:

Window based multi-head self attention (W-MSA) module with relative position bias. It supports both of shifted and non-shifted window. Args:

Skip calculating remaining time for the first few times. Args: skip_times (int): The number of times to skip. Defaults to 0.

Function wrapper for ``track_progress_rich``

save the loss: total(tensor), part1 and part2 can be 0

Very simple multi-layer perceptron (also called FFN)

Patch Merging Layer Args: dim (int): Number of input channels. norm_layer (nn.Module, optional): Normalization layer. Default:

Patch Merging Layer Args: dim (int): Number of input channels. norm_layer (nn.Module, optional): Normalization layer. Default: n

This is a more standard version of the position embedding, very similar to the one used by the Attention is all you need paper, generalized t

Affine Transformation to each frame

eval training output