Crates.io | numquant |
lib.rs | numquant |
version | 0.2.0 |
source | src |
created_at | 2022-04-11 09:38:23.522984 |
updated_at | 2022-04-11 21:25:34.053924 |
description | Quantize numbers to a smaller range to save bandwidth or memory data types and back again. |
homepage | |
repository | https://github.com/vilcans/numquant |
max_upload_size | |
id | 565550 |
size | 27,548 |
Quantize numbers to a smaller range to save bandwidth or memory.
The input floating point value is expected within a given range. Values outside this range will be clamped. The input value will then be quantized into a given integer range.
For example, given the allowed range -1000.0 to 1000.0, and the quantized range 0 to 255 (to fit in a byte), the value -1000.0 would be quantized to 0, and 1000.0 would be quantized to 255, and values in-between are linearly interpolated between 0 and 255.
This example uses the type Quantized<U8<0, 1000>>
that converts any floating point number between 0.0 and 1000.0 to a byte (which has the range 0 to 255). Some precision is lost, but an approximate value can be brought back.
let original = 500.0;
// Quantize the value into a byte.
// Quantization supports inputs between 0 and 1000.
let quantized = Quantized::<U8<0, 1000>>::from_f64(original);
// Convert it back to an f64
let dequantized = quantized.to_f64();
// The conversion isn't lossless, but the dequantized value is close to the original:
approx::assert_abs_diff_eq!(original, dequantized, epsilon = U8::<0, 1000>::max_error());