Creates a list of evaluators based on `config`.

grid_size: int of the grid height and width return: pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/

Returns a dataset with unrolled "caption" for every example.

Makes central crop of a given size. Args: crop_size: either an integer H, where H is both the height and width of the central crop, or a

Returns class name for `dataset_name` from `source`.

Returns config for mscoco retrieval zero-shot. Args: log_steps: How often the evaluators should be run. pp_img: Pre-processing string

Makes a random crop of a given size. Args: color_jitter: . Returns: A function, that applies color jitter.

Makes a random crop of a given size. Args: color_jitter: . Returns: A function, that applies color jitter.

Concatenates elements along some axis.

The base configuration.

Builds the base configuration.

Copies value of `inkey` into `outkey`.

Returns `ScaleByScheduleState.count` from `opt_state` as an integer.

Decode an encoded image string, see tf.io.decode_image.

Decode jpeg string and make inception-style image crop. Inception-style crop is a random image crop (its size and aspect ratio are random) that w

Returns config for discriminative_classifier of specified datasets.

Drops the given keys.

Flattens the keys of data with separator '/'.

Makes inception-style image crop. Inception-style crop is a random image crop (its size and aspect ratio are random) that was used for training I

Keeps only the given keys.

Map string to number.

One-hot encodes the input. Args: depth: Length of the one-hot vector (how many classes). key: Key of the data to be one-hot encoded. ke

Pads tensor to specified `shape`.

Convert i1k label numbers to strings, using CLIP's class names.

Extracts tokens with tensorflow_text.BertTokenizer. Args: vocab_path: Path to a file containing the vocabulry for the WordPiece tokenizer

Convert i1k label numbers to strings, using CLIP's class names.

Extracts tokens with tensorflow_text.BertTokenizer. Add custom args for coca training. Args: vocab_path: Path to a file containing the vocabu

Transform an input string into the preprocessing function. The minilanguage is as follows: fn1|fn2(arg, arg2,...)|... And describes the suc

Returns prompt templates.

Converts the specified feature to ragged tensor.

Creates a function that applies RandAugment. RandAugment is from the paper https://arxiv.org/abs/1909.13719, Args: num_layers: Integer, the

Makes a random crop of a given size. Args: crop_size: either an integer H, where H is both the height and width of the random crop, or a

Flips an image horizontally with probability 50%.

Resizes the smaller side to `smaller_size` keeping aspect ratio. Args: smaller_size: an integer, that represents a new size of the smaller side

Transforms a [in_min,in_max] image to [vmin,vmax] range. Input ranges in_min/in_max can be equal-size lists to rescale the invidudal channels ind

VGG-style preprocessing, subtracts mean and divides by stddev. This preprocessing is very common for ImageNet pre-trained models since VGG, and t

Runs the image-to-text retrieval eval from the distance matrix. Args: dist_matrix: Distance matrix between text and image embeddings (shape

Inverts the image pixels.

returns list of pretrained models Returns a tuple (model_name, pretrain_tag) by default or 'name:tag' if as_str == True

Loads both towers, `init_files` is now a dict with `img` and `txt` keys.

Load init from checkpoint, both old model and this one. +Hi-res posemb.

Loads pre-trained weights for the model's components. This function handles loading weights for the image encoder and text decoder from separ

Loads pre-trained model weights from a checkpoint file. Args: init_params: A PyTree of randomly initialized model parameters. ini

Load init from checkpoint, both old model and this one. +Hi-res posemb.

Loads a big_vision checkpoint saved by `save_checkpoint_ts`.