Tokenize strings and create time tokens.

Creates input dataset from (tf)records files for train/eval.

Creates a coverage based pileup dataset from a filepath. Args: filenames: The data directory/pattern containing data files. with_label: whe

Creates a tf.SequenceExample TFRecord.

Gets the train state from the input ckpt. Args: ckpt_path: Path to the ckpt to be loaded.. config: The configuration used to train the mode

Crop the image + bbox (+ mask) to region. WARNING! Only use during train. In eval mode the original_size would need to be updated somehow. Arg

Crop the image + bbox (+ mask) to region. WARNING! Only use during train. In eval mode the original_size would need to be updated somehow. Arg

The same as detr crop, with additional keys (e.g., object captioning.).

Crops and resizes the images in the given sequence of images. Args: frames: A tensor of dimension [timesteps, input_h, input_w, channels].

Crops and resizes the images in the given sequence of images. Args: frames: A tensor of dimension [timesteps, input_h, input_w, channels].

Cuts out portions of the given image by sampling random bounding boxes. Args: image: A single image, must be in 0-1 range. rng: Seed for sa

Randomly cuts out portions of the given image. Noise is sampled at `patch_size` times lower resolution and upsampled to form a checkerboard patte

Returns generators for COCO object detection 2017 train & validation set. Args: batch_size: int; Determines the train batch size. eval_batc

Gives a scaling factor based on scheduling with a decay every n-steps. Args: step: int; Current step. steps_per_decay: int; How often to de

Augment an image into a video.

Convert custom tfrecord into tfds builder format.

Convert yxyx [0, 1] normalized boxes to xyxy unnormalized format.

Decode boxes. Args: encoded_boxes: a tensor whose last dimension is 4 representing the coordinates of encoded boxes in ymin, xmin, ymax,

Convert yxyx [0, 1] normalized boxes to xyxy unnormalized format.

Convert yxyx [0, 1] normalized boxes to xyxy unnormalized format.

Convert points to detection boxes.

Given an instance and raw labels, creates <inputs, label> pair. Decoding includes. 1. Converting images from uint8 [0, 255] to [0, 1.] float32.

Convert custom tfrecord into training pipeline builder format.

Convert custom tfrecord into training pipeline builder format.

Convert custom tfrecord into tfds builder format.

Convert sharded file names into a list.

Convert sharded file names into a list.

Convert sharded file names into a list.

Convert sharded file names into a list.

Convert sharded file names into a list.

Convert sharded file names into a list.

Convert sharded file names into a list.

Convert sharded file names into a list.

Convert custom tfrecord into tfds builder format.

Default config for the flax model that is built in `build_flax_model`. This function in particular serves the testing functions and supposed to

Default config for the flax model that is built in `build_flax_model`. This function in particular serves the testing functions and supposed to

Default config for the flax model that is built in `build_flax_model`. This function in particular serves the testing functions and supposed to

Performs deformable attention sampling calculation. Args: values: [bs, len_v, nembed]-array of values. sampling_locations: [bs, nlevels, np

Takes a deterministic crop of input images. Args: images: `Tensor` of shape shape [t, h, w, c] size: Integer ; size of height and width to

Takes a deterministic crop of input images. Args: images: `Tensor` of shape shape [t, h, w, c] size: Integer ; size of height and width to

Set to zero all elements of a 2D array that are WITHIN radius of a point. Args: array: 2D array of non-zero elements center: [x,y] ([column

Applies RandomErasing to video. Args: video (tf.Tensor): Of shape [temporal, height, width, 3] representing a video. Returns:

Generates cropped_image using one of the bboxes randomly distorted. See `tf.image.sample_distorted_bounding_box` for more documentation. Args:

Check if the line segments p1q1 and p2q2 intersect.

Computes dot-product attention given query, key, and value. This is the core function for applying attention based on https://arxiv.org/abs/1706.

Returns the solution of max_x y^T x s.t. ||x||_2 <= 1.

Standard ViT method of embedding input patches.

Embed 3D input patches into tokens.

Embed 3D input patches into tokens.

Prepares the input for the screen encoder.

Prepares the input for the screen encoder.

Applies Transformer encoder-branch on the inputs.

Encodes positions using random features-based encoder.

Encodes text.

Encodes video. We use a separate encoder to encode each modality and the output is obtained by concatenating encoded tokens in the channel di

Eval the model and write the summary.

Evaluate the model and write the summary.

Evaluate caption switch.

Calculates the HOTA metrics for one sequence. Args: data: dict with keys: 'num_tracker_dets': int 'num_gt_dets': int

Runs evaluation code.

Runs evaluation code.

Evaluate an UnLoc model. This function runs a pretrained model on the test split of the specified dataset, and then evaluates the model. Args:

Runs evaluation code.

Perform validation and log results, possibly including mAP.