MiniCPM-V-4.7-1B-RKNN3-RK1828 / convert /export_vision_rknn.py
jj90's picture
Add model card, deploy guide, host + conversion scripts
fd11f30 verified
Raw History Blame Contribute Delete
1.67 kB
"""MiniCPM-V 4.7 1B vision ONNX -> RKNN (w4a16 / normal / group32 / 8 cores)。
参数照抄官方 zoo examples/MiniCPM_V_4/python/vision/export_rknn.py:
uint8 NHWC 输入 + mean/std=0.5*255(RKNN 在输入端做归一化),
ONNX 里只保留已归一化 fp32 NCHW 的计算图。
"""
import os
from rknn.api import RKNN
ONNX = os.environ.get("MINICPM_VISION_ONNX", "MiniCPM-V-4.7-1B-vision.onnx")
OUT = os.environ.get("MINICPM_VISION_RKNN", "MiniCPM-V-4.7-1B-vision.rknn")
def main():
rknn = RKNN(verbose=True)
print("--> config model", flush=True)
rknn.config(
target_platform="rk1820",
quantized_dtype="w4a16",
quantized_algorithm="normal",
quantized_method="group32",
core_num=8,
mean_values=[[0.5 * 255, 0.5 * 255, 0.5 * 255]],
std_values=[[0.5 * 255, 0.5 * 255, 0.5 * 255]],
input_attrs={"pixel": {"dtype": "uint8", "layout": "NHWC"}},
)
print("done", flush=True)
print("--> Loading model", flush=True)
ret = rknn.load_onnx(model=ONNX)
if ret != 0:
print("Load model failed!", ret, flush=True)
return ret
print("done", flush=True)
print("--> Building model", flush=True)
ret = rknn.build(do_quantization=True)
if ret != 0:
print("Build model failed!", ret, flush=True)
return ret
print("done", flush=True)
print("--> Export RKNN model", flush=True)
ret = rknn.export_rknn(OUT)
if ret != 0:
print("Export rknn failed!", ret, flush=True)
return ret
print("done", flush=True)
rknn.release()
return 0
if __name__ == "__main__":
raise SystemExit(main())