在pc上使用rknn-toolkit 1.3.2 调用rknn.inference(inputs=[img])需要5分钟

[复制链接] · 发表于 2020-6-19 14:33:36

root@843ead96d6af:/shares/PyTorch-YOLOv3-train# python rknn_yolo.py
--> config model
done
--> Loading model
/usr/local/lib/python3.5/dist-packages/onnx_tf/common/__init__.py:87: UserWarning: FrontendHandler.get_outputs_names is deprecated. It will be removed in future release.. Use node.outputs instead.
warnings.warn(message)
./yoloV3.pt ********************
WARNING: Token 'COMMENT' defined, but not used
WARNING: There is 1 unused token
done
--> Building model
W The target_platform is not set in config, using default target platform rk1808.
done
--> Export RKNN model
done
--> Init runtime environment
W [set_chip_platform_env:187]evaluate model on RK1808
done
--> Running model
W [rknn_inputs_set:1288] warning: inputs[0] expected input len is 602112, but actual len is 502500!
done
--> Begin evaluate model performance
W When performing performance evaluation, inputs can be set to None to use fake inputs.
W [set_chip_platform_env:187]evaluate model on RK1808
W [set_chip_platform_env:187]evaluate model on RK1808

在开始评估之后，需要大约5分钟才有结果。

只看该作者 · 发表于 2020-6-19 14:35:27

    model = Darknet(opt.model_def, img_size=opt.img_size)

    model.load_state_dict(torch.load("./checkpoints/custom_yolov3_ckpt_900.pth", map_location=torch.device('cpu')))

    model.eval()

    trace_model = torch.jit.trace(model, torch.Tensor(1,3,448,448))

    trace_model.save('./yoloV3.pt')

    np.set_printoptions(formatter={'float': '{: 0.3f}'.format})

    model_path = './yoloV3.pt'

    input_size_list = [[3, 448, 448]]



    # Create RKNN object

    rknn = RKNN()

    print('--> config model')

    rknn.config(channel_mean_value='0. 0. 0. 255.', reorder_channel='0 1 2')

    print('done')



    # Load pytorch model

    print('--> Loading model')

    ret = rknn.load_pytorch(model=model_path, input_size_list=input_size_list)

    if ret != 0:

        print('Load pytorch model failed!')

        exit(ret)

    print('done')



    # Build model

    print('--> Building model')

    # 不量化，结果基本一致

    ret = rknn.build(do_quantization=False, dataset='./valid.txt')

    if ret != 0:

        print('Build pytorch failed!')

        exit(ret)

    print('done')



    # Export rknn model

    print('--> Export RKNN model')

    ret = rknn.export_rknn('./yoloV3.rknn')

    if ret != 0:

        print('Export yoloV3.rknn failed!')

        exit(ret)

    print('done')



    ret = rknn.load_rknn('./yoloV3.rknn')



    # Set inputs

    img = cv2.imread('./data/custom/images/train.jpg')

    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

    # 归一化

    img = np.array(img, dtype=np.float32)

    img = np.array((img-img.min()) / img.ptp())

    # print("img: ", img)



    # init runtime environment

    print('--> Init runtime environment')

    ret = rknn.init_runtime()

    if ret != 0:

        print('Init runtime environment failed')

        exit(ret)

    print('done')



    # Inference

    print('--> Running model')

    outputs = rknn.inference(inputs=[img])

    print('done')



    # # perf

    print('--> Begin evaluate model performance')

    perf_results = rknn.eval_perf(inputs=[img])

    print('done')

    rknn.release()
复制代码

只看该作者 · 发表于 2020-6-19 23:28:48

1. 先保证drv和api是匹配的，否则运行不出结果，5分钟可能是错误超时时间。
2. 需要量化后再来看，不量化的时间是没意义的。
3. yolov3看我们教学贴都有，自己对比下和我们有什么区别。