Run evaluation.

Compute the mean NLP metrics over the given IOU thresholds. Args: ground_truths_filtered: Filtered list of ground truth captions for each th

Compute the mean P/R between the predicted and ground truth segments. Args: predicted_segments: A numpy array of shape [K x 2] containing the p

Compute both the P/R and NLP metrics for the given predictions. Args: predicted_segments: A numpy arrays, of shape [K x 2] containing the p

Gives a scaling factor based on scheduling with an exponential decay. Args: step: int; Current step. decay_steps: int; Number of steps to d

Extracts the label for each bbox, and sorts the results by score. More specifically, after extracting each bbox's label, the bboxes and score

Extract 2hop graphs for all pages in each file. Args: kg_graph: the global KG (i.e. WikiData Knowledge Graph), stored as dict. i: index of

Use multiprocessing to parallize graph extraction. Args: kg_graph: the global KG (i.e. WikiData Knowledge Graph), stored as dict. n_pool: n

Extracts tubelets from videos for use as regression targets. Args: x: Input tensor of shape [batch, time, height, width, channels] tubelet_

Compute F1 score, but any prediction != 0 or 1 is counted as incorrect. Args: target: Numpy array of targets, either 0 or 1 (binary label space

Sample boxes in [x, y, x, y] format in un-normalized image coordinates.

Computes bidirectional (noncausal) normalized FAVOR+ attention. Computes FAVOR+ linear attention from Performers,based on: "Rethinking Attention

Remove shapes that are not visible in the image. Args: config: configuration object (hyper-parameters for shape generation) imagedict: shap

Flattens the first two dimensions of a non-scalar array.

Flattens the first two dimensions of a non-scalar array.

Flattens the first two dimensions of a non-scalar array.

Formats predictions to COCO annotation format. Args: scores: [num_images, num_instances] array of confidence scores. labels: [num_images, n

Tokenize raw string with tokenizer.

Generate captions using augmented features from tracking. Args: object_features: (batch_size, num_objs, res, res, D) matched_text: (b

Forward global video captioning. Args: features: list of arrays: FPN features. video_caption_tokens: 1 x N x max_caption_length. Only

Forward tracking head. The images from the batch are from the same video. Args: last_proposals: (batch_size, num_objs, 4) object

Fuses video and text tokens. Args: video_tokens: A 3D float tensor of shape (batch_size, sequence_length, channels) representing th

Flatten leading two dims, e.g. to get global batch after all_gather.

Gathers the top-k beam slices given by score_or_log_prob array. Args: nested: pytree of arrays or scalars (the latter ignored). score_or_lo

Gathers the top-k beam slices given by score_or_log_prob array. Args: nested: pytree of arrays or scalars (the latter ignored). score_or_lo

Gathers the top-k beam slices given by score_or_log_prob array. Args: nested: pytree of arrays or scalars (the latter ignored). score_or_lo

Gathers the top-k beam slices given by score_or_log_prob array. Args: nested: pytree of arrays or scalars (the latter ignored). score_or_lo

Blurs the given image with separable convolution. Args: image: Tensor of shape [height, width, channels] and dtype float to blur. kernel_s

Constructs kernel features for fast generalized attention. Args: data: input for which features are computes projection_matrix: matrix used

Automatically generate masks for all objects. This function is from the original SamAutomaticMaskGenerator at https://github.com/facebookrese

Generate a random COCO example.

Generate one sample for parity task. Args: batch_size: Determines the batch size. seq_len: Determines the sequence length of parity vector.

Generate one sample for parity task. Args: batch_size: Determines the batch size. seq_len: Determines the sequence length of parity vector.

Generate a sequence example.

Maps output of model to alpha, omega, vertex.

Computes average batch size.

Computes average batch size.

Find smoothed patches of the image along with wedge colors. Args: input_features: Input features with shape [N, C, H, W] global_featu

Converts 2d inputs to 1d for axial attention.

Converts 1d inputs back to 2d after axial attention.

Compute the boundary consistency loss. Args: dists: Array of shape [N, 2, R, R, H', W'] with samples of the two distance functions fo

Returns a [num_classes, 3] array of colors for the model output labels.

Returns a list with the class names of the model output labels.

Computes class proportions in case of label imbalance.

Returns a [num_classes] array of pixel frequency proportions.

Returns the COCO Caption eval source.

Returns the COCO Caption eval source.

Returns the COCO Caption eval source.

Returns the COCO Caption eval source.

Returns the COCO Caption eval source.

Returns the COCO Caption eval source.

Returns the COCO Caption train source.

Returns the COCO Caption train source.

Returns the COCO Caption train source.

Returns the COCO Caption train source.

Returns the RefCOCO/RefCOCOg/RefCOCO+/ eval source.

Returns the RefCOCO/RefCOCOg/RefCOCO+/ train source.

Returns the RefCOCO/RefCOCOg/RefCOCO+/ eval source.

Returns the RefCOCO/RefCOCOg/RefCOCO+/ train source.

Returns the RefCOCO/RefCOCOg/RefCOCO+/ eval source.

Returns the RefCOCO/RefCOCOg/RefCOCO+/ eval source.

Returns the RefCOCO/RefCOCOg/RefCOCO+/ train source.

Returns the RefCOCO/RefCOCOg/RefCOCO+/ train source.

Returns the RefCOCO/RefCOCOg/RefCOCO+/ train source.

Returns the RefCOCO/RefCOCOg/RefCOCO+/ train source.

Compute the spatial color consistency loss. Args: dists: Array of shape [N, 2, R, R, H', W'] with samples of the two distance functio

Returns the ViT experiment configuration for ImageNet. This file is a copy of config/imagenet_augreg_matvit_config.py.

Computes the confusion matrix that is necessary for global mIoU.

Computes confusion matrix from predictions. Args: labels: [n_batch] or [n_batch, n_classes] array. In the latter case, labels are assumed

Labels the inner products compute from `compute_inners`. Given an output from `_compute_inners` and text_mask, returns: -labels: indicates where

Return a (sub)split adapted to a given host.

Returns a tf.data.Dataset and class names.

Returns a generator for the audiovisual dataset.

Returns dataset given config.

Returns dataset config.

Returns dataset config.

Returns dataset config.

Data provider.

Creates dataset from config.

Get default adversarial config.

Gets default configs for wit_internal (en) dataset.

Gets default configs for wit_internal (en) dataset.

Gets default configs for CC12M dataset.

Gets default configs for argus dataset.

Gets default configs for wit_internal (en) dataset.

Gets default configs for wit_internal (en) dataset.