Adjusts logits and label to have the same shape. Args: logits: Output of model in shape [batch, batch * max_num_captions, num_frames, 1].

Adjusts prediction to have the same shape with ground truth. Regression loss is only computed at positive frames. Args: displacements: Predi

Applies dropout on attention weights. This always applies the dropout. There is no `deterministic` parameter. Arguments: attn_weights: Atten

Find Bounding box overlap. Args: boxes: first set of bounding boxes gt_boxes: second set of boxes to compute IOU Returns: iou: Inter

Find Bounding box overlap. Args: boxes: first set of bounding boxes gt_boxes: second set of boxes to compute IOU Returns: iou: Inter

This function is to compute the weights for box loss.

Build a tf.data.Dataset pipeline using clu.deterministic_data. The pipeline has the following steps: 0. All used datasets should have text label

Builds TAL head with shared decoders.

Builds TAL head with separate decoders.

This function is to compute the weights for classification.

Computes a cover for the bipartite graph. We compute a cover using the construction provided at https://en.wikipedia.org/wiki/K%C5%91nig%27s_theo

Calculate sub-metric values which only depend on other values.

Computes (x,y) linear regression accuracy on (x_test, y_test). Args: x: An [n_examples, n_classes] matrix of feature representations. This will

Finds an augmenting path given an assignment and an adjacency matrix. The augmenting path search starts from the unassigned workers, then goes on

Makes old checkpoints forward-compatible.

Fix layer norm numbering of old checkpoints.

Fix resblock naming of old checkpoints.

Flattens a dictionary, keeping empty leaves.

Forward box head and get losses or post-processed predictions. Args: features: A sorted list of (B, Hi, Wi, C) features. proposals: a

Yields examples.

Calculate absolute positional embeddings. If needed, resize embeddings and remove cls_token dimension for the original embeddings. Args

Calculate absolute positional embeddings. If needed, resize embeddings and remove cls_token dimension for the original embeddings. Args

Returns an attention module.

Generates camera noise.

Combines the 1d index & their interpolation weights to do einsum. Args: grid: (T, output_size * sampling_ratio), 1d grid for each box.

Manually create a single example that has known results.

Prepares input data from one prompt. Args: batch: A batch of input data. num_classes: Number of classes. prompt_index: The index of pro

Returns method_fn function corresponding to method str.

Generates perlin noise.

Generates a random offset for sampling the sequence.

Compute representation for the whole dataset.

Compute representation for the whole dataset.

Compute representation for the whole dataset.

Gets result lists for the task and dataset.

Gets tensor shape as a list, allowing mixing static and dynamic shapes.

Returns class_splits and split_names.

Return self-attention variant specific list of attn args.

Handle legacy format. To remove this function, make sure all checkpoints that are to be loaded are in the non-legacy format: * checkpoint['pa

Checks which elements of string tensor 'labels' are forbidden.

Helper to check if sched itself or (fn, base_lr) of sched are None.

Performs maximum bipartite matching using augmented paths. Args: adj_matrix: A bool [b, n, m] array, where each element of the inner matrix

Normalizes predicted distances to the start and end times.

Return the function that corresponds to `name` and update `level` param.

Partitions parameters into frozen and learned parameter trees. Args: params: Pytree of model parameters. frozen_mapping: Dictionary mapping

Partitions parameters into frozen and learned parameter trees. Args: params: Pytree of model parameters. frozen_mapping: Dictionary mapping

PE encoding.

Adds boxes to the provided Bokeh plot.

Adds the predicted probabilities to prediction dict.

Replaces values in model dictionary with restored ones from checkpoint.

Resizes the image. Args: image: Tensor; Input image. image_size: int; Image size. Returns: Resized image.

Samples a bounding box in the [cx, cy, w, h] in [0, 1] range format.

Randomly samples a crop region (bounding box) for random cropping. The region is represented as two tensors, `begin` and `size`, which can be

Computes `frozen_mask` and updates `masks` and `scheds`.

Validates dataset_configs and normalizes it to a common format. Common and train/eval-mode-specific configs are merged into one. TFDS data dirs a

Multiplies the stack of H masks M (shape [L, L] each) by the inp. tensor. We denote by L the length of the input sequence and by H the numbe

Creates new beam dimension in non-scalar array and tiles into it.