encode_ternary_array¶

keras_mml.utils.array.encoding.encode_ternary_array(x)[source]¶

Encodes a ternary array into a more space efficient format.

For a given ternary matrix, which consists of only the elements 0, 1, and -1, we convert the individual elements into bit sequences. Specifically,

the most common element becomes 0;
the second most common element becomes 10; and
the least common element becomes 11.

We then convert these bit sequences into bytes. The first three to four bits describes which elements were converted to 0 and 10 respectively.

0 will be represented by 0;
1 will be represented by 10; and
-1 will be represented by 11.

This function will prepend this information in front of the bit sequences, and then return them as bytes.

Parameters:: x (ndarray) – Ternary array to encode.
Returns:: Tuple[Tuple[int, ...], bytes] – A tuple. The first element is the shape of the array. The second element is the encoded representation of the array.

Examples

>>> x = np.array([1, -1, 1, 0, -1, 1])
>>> shape, encoded = encode_ternary_array(x)
>>> shape
(6,)
>>> encoded
b'\xb4\xe0'

>>> x = np.array([[0, 1, -1], [-1, 0, 0]])
>>> shape, encoded = encode_ternary_array(x)
>>> shape
(2, 3)
>>> encoded
b'n\x80'

>>> x = np.array([[[0, 1, -1], [-1, 1, 0]], [[1, 0, 0], [0, -1, 1]]])
>>> shape, encoded = encode_ternary_array(x)
>>> shape
(2, 2, 3)
>>> encoded
b'K\xe48'

>>> x = np.array([[[0, 1, -1], [-1, 1, 0]], [[1, -1, 0], [0, -1, 1]]])
>>> shape, encoded = encode_ternary_array(x)
>>> shape
(2, 2, 3)
>>> encoded
b'K\xe5\x9c'