quantize_fp16.py

import os

import onnx
from onnxconverter_common import float16

# Configuration
ONNX_MODEL_PATH = "model/efficientnet-b0.onnx"
FP16_MODEL_PATH = "model/efficientnet-b0_fp16.onnx"

print("Starting conversion to Float16 (FP16)...")

# Load ONNX model
model = onnx.load(ONNX_MODEL_PATH)

# Convert model to Float16
model_fp16 = float16.convert_float_to_float16(model)

# Save the FP16 model
onnx.save(model_fp16, FP16_MODEL_PATH)

print(f"FP16 conversion complete -> Size: {os.path.getsize(FP16_MODEL_PATH) / 1e6:.2f} MB")