Computes and stores the average and current value

An attention layer that allows for downscaling the size of the embedding after projection to queries, keys, and values.

An attention layer that allows for downscaling the size of the embedding after projection to queries, keys, and values.

A structure for storing masks and their related data in batched format. Implements basic filtering and concatenation.

Computes and stores the current, avg, and max of peak Mem usage per iteration

Image to Patch Embedding.

A list of video frames to be load asynchronously without blocking session start.

A list of video frames to be load asynchronously without blocking session start.

Multi-head Attention block with relative position embeddings.

Multi-head Attention block with relative position embeddings.

This class represents a batch of videos with associated annotations and metadata. Attributes: img_batch: A [TxBxCxHxW] tensor contain

This class represents metadata about a batch of videos. Attributes: unique_objects_identifier: A tensor of shape Bx3 containing uniqu

Transformer blocks with support of window attention

Transformer blocks with support of window attention

Only decodes segments that are actually used.

A logger class that can interface with multiple loggers. It now supports tensorboard only for simplicity, but you can extend it with your own log

Image to Patch Embedding.

Positional encoding using random spatial frequencies.

Positional encoding using random spatial frequencies.

SAM2VideoTrainer is a PyTorch module for training a video segmentation model using SAM2. Attributes: device (torch.device): The devic

A callable which is passed to submitit to launch the jobs.

A simple logger for TensorBoard.

Refers to an image/video and all its annotations

r"""ConvNeXt Block. There are two equivalent implementations: (1) DwConv -> LayerNorm (channels_first) -> 1x1 Conv -> GELU -> 1x1 Conv; all in (N,

r"""ConvNeXt Block. There are two equivalent implementations: (1) DwConv -> LayerNorm (channels_first) -> 1x1 Conv -> GELU -> 1x1 Conv; all in (N,

Removes the keys from the given model state_dict that match the key_pattern. Args: key_pattern: Patterns used to select the keys in

Attention with rotary position encoding.

Attention with rotary position encoding.

The predictor class to handle user interactions and manage inference states.

The predictor class to handle user interactions and manage inference states.

Optimized for the VOS setting

Optimized for the VOS setting

VOS Sampler for evaluation: sampling all the frames and all the objects in a video

A modified variant of Feature Pyramid Network (FPN) neck (we remove output conv and also do bicubic interpolation similar to ViT pos embe

Gather tensors from all workers with support for backward propagation: This implementation does not cut the gradients as torch.distributed.al

Gradient clipping utils that works for DDP

Reference: https://arxiv.org/abs/2306.00989

Dataset where the annotation in the format of SA-V json files

A structure for storing masks and their related data in batched format. Implements basic filtering and concatenation.

Progressively downsample a mask by total_stride, each time by stride. Note that LayerNorm is applied per *token*, like in ViT. With each

Progressively downsample a mask by total_stride, each time by stride. Note that LayerNorm is applied per *token*, like in ViT. With each