转rknn后输出和原onnx输出不一样的问题

模型为https://github.com/commaai/openp ... els/supercombo.onnx，在PC端转为rknn（因为板端无法使用1.7.6的rknn_toolkit转rknn），在板端运行后，两者输出差别非常大
onnx到rknn的转换代码
    rknn = RKNN(verbose=True)
    print('--> Config model')
    #不使用config的均值和标准差，对于多行的dataset,，需指定batch_size=1
    rknn.config(target_platform='rk3399pro', optimization_level=3, output_optimize=1,batch_size=1)
     # Load model
    print('--> Loading model')
    ret = rknn.load_onnx(model=ONNX_MODEL)
    if ret != 0:
        print('Load model failed!')
        exit(ret)
    print('done')
    # Build model
    print('--> Building model')
    ret = rknn.build(do_quantization=True, dataset='./dataset_npy_hand255.txt')
    if ret != 0:
        print('Build model failed!')
        exit(ret)
    print('done')
    # Export rknn model
    print('--> Export rknn model')
    ret = rknn.export_rknn(RKNN_MODEL)

rknn的运行代码
    rknn = RKNN(verbose=True)
    ret = rknn.load_rknn(path=RKNN_MODEL)
    # Init runtime environment
    print('--> Init runtime environment')
    ret = rknn.init_runtime(device_id='81ab609370303eac')
    if ret != 0:
        print('Init runtime environment failed!')
        exit(ret)
    print('done')
    # Set inputs
    in1 = np.load('straight1/straight1.npy')
    in2 = in1
    in3 = np.load('datas/desire_fp32.npy')
    in4 = np.load('datas/traffic_convention_fp32.npy')
    in5 = np.load('datas/nav_features_fp32.npy')
    in6 = np.load('datas/features_buffer_fp32.npy')
    # Inference
    print('--> Running model')
    print("模型路径",RKNN_MODEL)
    output = rknn.inference(inputs=[in1, in2, in3, in4, in5, in6])
    output[0] = output[0].reshape((1, -1))
    print('inference result: ', output)
    output=np.array(output).reshape(6120,)
    np.savetxt("./straight1/rk3399pro_straight1.txt",output)

测试的图片及模型数据
链接: https://pan.baidu.com/s/1b-iBe0E7YfkBdexxXTH-6Q?pwd=evji 提取码: evji