Computes quantization range to ensure that zero is represented if covered. Gymnastics with nudged zero point is to ensure that real zero maps to an

Reads the requirements file and returns the packages

Generates dictionaries that map tensor names to the loaded saved model resource id, allowing for matching of initializer outputs to inference inp

Create a tf.function that creates and initializes all the resources used by the model. For more information on resources, please see the TensorFlow

Gets the weights entries manifest JSON for a group. Args: group: A list of weight entries. Returns: An list of manifest entries (dicts) t

Prepopulate required fields for all data foramts. Args: h5file: Valid h5file object. Returns: A dictionary with common fields sets, shar

Quantizes the weights in the entry, returning a new entry. The weights are quantized by linearly re-scaling the values between the minimum and ma

Serializes a numeric numpy array into bytes. Args: data: A numeric numpy array. Returns: bytes of the array to be serialized on disk.

Serializes a numpy array of dtype `string` into bytes. Each string value is preceded by 4 bytes which denote a 32-bit unsigned integer in little

Shards the concatenated bytes for a group to disk. Args: write_dir: The directory to write the files to. group_index: The index for the gro

Stacks the bytes for a weight group into a flat byte array. Args: group: A list of weight entries. Returns: A type: (group_bytes, total_b

Standardize input and output formats. Args: input_format: Input format as a string. output_format: Output format as a string. Returns:

* Add a loader element under the tab page. * * @param {string} tabId

* Helper function to check the dtype and shape compatibility of a feed value.

* Assembles a list of same-shaped numbers, number arrays, or Tensors * into a single new Tensor where axis 0 is the batch dimension.

* Binds the ~ (tilde) key to toggle the visor. * * This is called by default when the visor is initially created. * * @doc {heading: 'Viso

* Perform byte-pair merge for each word in the inputs.

* Process unseen tokens and add to cache.

* Perform one step of byte-pair merge.

* Broadcasting batch normalization for use in training (not inference). * * The input is normalized to zero mean and unit variance along the * `red

* Group data into histogram buckets. * * @param data A `Tensor` of any shape. Must be castable to `float32` * @param bucketCount Optional pos

* Calls node-gyp for Node.js Tensorflow binding after lib is downloaded.

* * @param inputShape decoderSequenceShape or * [decoderSequenceShape, encoderSequenceShape]

* Builds multi-head dot-product attention computations. * * This function builds attributes necessary for `computeAttention` to * customize a

* Builds einsum equations for the attention computation. * * Query, key, value inputs after projection are expected to have the shape as: * `(bs, <

* The outputIndex represents the index in the output tensor * where the first element of a particular dimension would be written. * If it is -1,

* Calls the layer and returns extra information like the paddingMask used to * pack the sequence, the label data, and the sample weights used.

Center crops an image