Dequantize signed-power embedding codes.

The decode sign(x) * (|x| / scale)^power: per-value, nonlinear and sign-preserving. Geo-embedding products store float vectors as Int8 codes this way, spending more codes near zero where most values sit; with the usual scale = 127.5 the code range [-127, 127] maps to about [-1, 1]. The signed_power method of the Embeddings Zarr Convention's geoemb:quantization metadata names the two parameters. dequantize_aef() and dequantize_olmoearth() are this decode with power = 2.

Usage

dequantize_signed_power(x, power = 2, scale = 127.5)

Arguments

x

Int8 codes (traced array or plain numeric).

power

Exponent of the decode (2 for AEF and OlmoEarth).

scale

Code magnitude that decodes to 1 (127.5 for both).

Value

The dequantized values, same shape as x.

Details

Written in the g_* vocabulary, so applying it with lazy_map() after lazy_dataset() fuses the decode onto the read, on the device rather than as a separate decode pass. A nodata code (-128 in both products) must reach the decode as NaN: lazy_dataset() does that whenever the source declares it (a GeoTIFF nodata value, a Zarr fill value), and NaN passes through the decode unchanged.

See also

dequantize_aef(), dequantize_olmoearth()