quantize_.py

from onnxruntime.quantization import quantize_dynamic, QuantType
from onnxruntime.quantization.preprocess import quant_pre_process
import argparse
import os

# 1. Clean the graph first
def main() : 
    parser = argparse.ArgumentParser(description="ONNX Quantization")
    parser.add_argument('model_path',type=str)
    parser.add_argument('output_path',type=str)
    parser.add_argument('--per_channel', action='store_true', help='Use per-channel quantization')
    parser.add_argument('--quint8', action='store_true', help='Use QInt8 (signed); default is QUInt8') 

    #python3 quantize_.py  model_path =""

    args = parser.parse_args()
    temp_file = "processed.onnx"
    # Skip symbolic shape inference (often fails on detection/vision models with dynamic shapes)
    quant_pre_process(args.model_path, temp_file, skip_symbolic_shape=True)

    # 2. Quantize with SIGNED integers (better for OpenVINO)
    quantize_dynamic(
        model_input=temp_file,
        model_output=args.output_path,
        weight_type=QuantType.QInt8 if args.quint8 else QuantType.QUInt8,
        per_channel= args.per_channel,
        
    )

    os.remove(temp_file)

if __name__ == "__main__" :
     main()