tf.keras.quantizers.pack_int2

Pack an int2 tensor into an int8 tensor with 4 values per byte.

tf.keras.quantizers.pack_int2(
    arr, axis=0, dtype='int8'
)

The input values must already be int8/uint8 representing the desired int2 values (signed range [-2, 1], or unsigned range [0, 3]). Packing is performed along the specified axis (default is 0). Four consecutive values along the packing axis are stored in a single output byte, from the least significant 2-bit field to the most significant one.

This mirrors the design of pack_int4 (padding + original-length trim), but achieves a 4x rather than 2x storage reduction, which is what makes 2-bit storage worthwhile relative to a plain uint8 tensor.

arr An int8 or uint8 tensor containing int2 values in the range [-2, 1] (signed) or [0, 3] (unsigned).
axis The axis along which to pack the tensor. Defaults to 0.
dtype The data type of the input and packed tensor. Can be "int8" or "uint8". Defaults to "int8".

tuple A tuple (packed, packed_shape, orig_len) where packed is the packed tensor with four int2 values per byte, packed_shape is the shape of the packed tensor, and orig_len is the original (unpacked) length along axis prior to any padding that was inserted to reach a multiple of four.

Example:

import numpy as np from keras.quantizers import pack_int2 codes = np.array([[-2, 1, 0, -1, 1]], dtype="int8") packed, packed_shape, orig_len = pack_int2(codes, axis=1) np.asarray(packed) # four 2-bit fields per byte; axis 1 padded to 8 array([[-58, 1]], dtype=int8) packed_shape, orig_len ((1, 2), 5)