quantification-MobileNetV5-V4 / scripts /tflite_v4s_scale_int4.py
Lohan11's picture
Initial release: technical documentation, quantization scripts, and NPU benchmarks (part 2)
2248c41 verified
Raw History Blame Contribute Delete
1.74 kB
import tensorflow as tf
import numpy as np
import ai_edge_quantizer
import argparse
import os
def quantize_v4s_scale_int4(model_path, calib_path, output_path):
"""
Quantization script for MobileNetV4-Small (Scale Metrology) - W4A8 (INT4 Weights).
Optimized for 1024x1024 resolution.
"""
print(f"Loading calibration data: {calib_path}")
# Load NCHW and transpose to NHWC
calib_data = np.load(calib_path).astype(np.float32).transpose(0, 2, 3, 1)
quantizer = ai_edge_quantizer.TFLiteQuantizer(
model_path=model_path,
calibration_data=[calib_data]
)
compression_config = ai_edge_quantizer.CompressionConfig(
weight_bits=4,
activation_bits=8,
asymmetric=True
)
print("Exporting MobileNetV4-Small Scale to W4A8 (INT4) TFLite...")
quantizer.export(
output_path,
compression_config=compression_config,
input_type=ai_edge_quantizer.DType.INT8,
output_type=ai_edge_quantizer.DType.INT8
)
print(f"Successfully saved W4A8 model to {output_path}")
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="Lithotheque V4S Scale INT4 Quantizer")
parser.add_argument("--model", type=str, required=True, help="Path to FP32 SavedModel")
parser.add_argument("--calib", type=str, default="calib_echelle.npy", help="Path to calibration .npy")
parser.add_argument("--out", type=str, default="echelle_v4s_int4.tflite", help="Output path")
args = parser.parse_args()
if os.path.exists(args.model):
quantize_v4s_scale_int4(args.model, args.calib, args.out)
else:
print(f"Error: {args.model} not found.")