Leo Quantization

Implement quantization operations in Leo, including $+ * /$, on quantized values.

Description

Use quantized data in the zkML program in Leo and perform operations on quantized data. Choose fixed FP32 quantization, which is 32-bit floating-point quantization.

Directory Structure

• quantization: Python FP32 quantization implement
• quantize: a Leo library that implements Float arithmetic operations

Key Points

FP32

FP32 uses 32 bits to represent a number: 1 bit for the sign, 8 bits for the exponent, and the remaining 23 bits for the significant figures. Although it provides high precision, the disadvantage of FP32 is its high computation and memory usage.

Leo Language

Support u8, u16, u32, u64, u128, i8, i16, i32, i64, i128 numeric types, "Leo Quantization" convert Float values in ML/AI to integer values supported by Leo.

Usage

# code clone
git clone https://github.com/storswiftlabs/leo-quantization.git
# Python
python -m unittest quantization/test_float.py
# Leo run
cd quantize
# Execute quantized arithmetic
leo run float_add
leo run float_sub
leo run float_mul
leo run float_div
# Change inputs data
vi inputs/quantize.in

References

[1] Maxime Labonne. Introduction to Weight Quantization, https://towardsdatascience.com/introduction-to-weight-quantization-2494701b9c0c, 2023.

[2] LIU, Xuanming, et al. Evaluate and Guard the Wisdom of Crowds: Zero Knowledge Proofs for Crowdsourcing Truth Inference. arXiv preprint arXiv:2308.00985, 2023.

[3] Fixed-point arithmetic, https://en.wikipedia.org/wiki/Fixed-point_arithmetic#Operations

[4] zkTI, https://github.com/Blockchain-Research-Center/zkTI.