Return the value for key if it exists otherwise the default.

Return the value for key if it exists otherwise the default.

Return the value for key if it exists otherwise the default.

Update with the given per-step metrics.

Return the value for key if it exists otherwise the default.

Return the value for key if it exists otherwise the default.

Get a dataset from the registry by its name. Args: name: Dataset name. Returns: Dataset builder function that accepts dataset-sp

Loads backbone parameters for this model from a checkpoint.

Returns the number of examples in the TFRecord files.

A chronometer tick.

Add a dataset to the registry, i.e. register a dataset. Args: name: Dataset name (must be unique). builder_fn: Function to be called

Step to learning rate function.

Provide some extra info that's only known later in the program.

Write to the given writer and training step. After writing, the state gets reset. Args: writer: The summary will be written with this

Clip gradients stored as a pytree of arrays to maximum norm `max_norm`.

Vision task with val and test splits.

Update with the given per-step metrics.

Applies given weights of the inputs in the minibatch to outputs. Note that weights can be per example (i.e. of shape `[batch,]`) or per pixel/tok

Update with the given per-step metrics.

Update per-frame localization predictions. Args: predictions: dict of image_id ('{video_id}_{frame_id}') to dict {'boxes': [[l, t,

Evaluate on given predictions.

Returns a callable metric function for the model. Args: split: The split for which we calculate the metrics. It should be one of th

Fold patches of an image using function fn. Args: unfolded: Array of shape [N, C, R, R, H', W'] containing all unfolded patches.

Proprocess images. Normalize pixels for non-padded pixels.

Compute METEOR scores.

Create a fake features dictionary. Args: n: int; number of boxes/objects. h: int; height of fake image. w: int; width of fake image.

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

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

Returns a callable metric function for the model. Args: split: The split for which we calculate the metrics. It should be one of th

Returns a callable metric function for the model. Args: split: The split for which we calculate the metrics. It should be one of th

Initialize the positional embeddings.

Replaces values in model dictionary with restored ones from checkpoint.

Brings text labels into a standard form.

Applies Transformer decoder-branch on encoded-input and target.

Applies Transformer encoder-branch on the inputs.

Returns a callable metric function for the model. Args: split: The split for which we calculate the metrics. It should be one of th

Assert whether two feature dicts are equal.

Construct and apply spatial mask to the inputs. Args: inputs: Input tensor with dimensions [batch, spatial_dim, dim] to which the mask sh

TruncatedNormal(0.02) initializer.

Clears the annotations collected in this object. It is important to call this method either at the end or at the beginning of a new evaluatio

Return the height, width of image as float32 tf.Tensors.

Returns a callable metric function for the model. Args: split: The split for which we calculate the metrics. It should be one of th

Converts a float image to uint8. Also un-whitens the image, if needed.

Initialize input embedding.

Call `writer` with a descriptive string and the results.

Resets the record of config field accesses.

Compute summary over all `datasets` that comes from config.

Scale the 2D image using the autocontrast rule.

Transfers array (replicated on multiple hosts) to a single host. Args: array: Replicated array of shape [num_hosts, num_devices, local_ba

Extract patches from an image. Args: im: Array of shape [N, C, H, W] to be unfolded into patches patchsize: Size of extracted patches

Combine a batch of padded instance labels into a single-example label.

Maps `ds` using the preprocess_fn and a deterministic RNG per example. Args: ds: Dataset containing Python dictionary with the features. The 'r

Returns generators for pretraining dataset or OVEN sets. Args: batch_size: int; Determines the train batch size. eval_batch_size: int; Dete

Forward decoder. Args: text_tokens: (batch_size, max_caption_length) encoded: (batch_size, num_tot_tokens, dim) encoder_mask: (

Computes dot-product attention given query, key, and value. This is the core function for applying fast approximate dot-product attention. It

Embeds image and returns image embeddings and boxes. Args: image: Single uint8 Numpy image of any size. Will be converted to float

Gathers the beam slices indexed by beam_indices into new beam array. Args: nested: pytree of arrays or scalars (the latter ignored). beam_i

Compute distance functions and piecewise-constant patches. Args: params: Array of shape [N, 5, H', W'] holding N field of junctions

Test get_label_map.