使用Caffe模型+DPU进行实时人脸检测：Vitis-AI3.0的非官方支持流程

def detect():    resolution = 4    threshold = 0.9    nms_threshold = 0.3       g = xir.Graph.deserialize("../densebox320/densebox320.xmodel")    subgraphs = get_child_subgraph_dpu(g)    assert len(subgraphs) == 1  # only one DPU kernel       dpu_runner = vart.Runner.create_runner(subgraphs[0], "run")    camera = cv2.VideoCapture(0)    width = 640    height = 360    camera.set(cv2.CAP_PROP_FRAME_WIDTH,width)    camera.set(cv2.CAP_PROP_FRAME_HEIGHT,height)    while True:        (grabbed, image_ori) = camera.read()        imagePad = padProcess(image_ori)        image = cv2.resize(imagePad,(32*int(320/32), 32*int(320/32)), interpolation = cv2.INTER_CUBIC)        szs = (float(imagePad.shape[0])/float(image.shape[0]), float(imagePad.shape[1])/float(image.shape[1]))        sz = image.shape        image = image.astype(np.float)        image = image - 128        # image = np.transpose(image, (2, 0, 1))        output = rundensebox(image, dpu_runner)        # generate result        prob = output[1][0, ..., 1]        bb = output[0][0, ...]        bb = np.transpose(bb, (2, 0, 1))        gy = np.arange(0, sz[0], resolution)        gx = np.arange(0, sz[1], resolution)        [x, y] = np.meshgrid(gx, gy)        bb[0, :, :] = bb[0, :, :] + x        bb[0, :, :] = bb[0, :, :] * szs[1]        bb[1, :, :] = bb[1, :, :] + y        bb[1, :, :] = bb[1, :, :] * szs[0]        bb[2, :, :] = bb[2, :, :] + x        bb[2, :, :] = bb[2, :, :] * szs[1]        bb[3, :, :] = bb[3, :, :] + y        bb[3, :, :] = bb[3, :, :] * szs[0]        bb = np.reshape(bb, (4, -1)).T        prob = np.reshape(prob, (-1, 1))        bb = bb[prob.ravel() > threshold, :]        prob = prob[prob.ravel() > threshold, :]        # nms        rects = np.hstack((bb, prob))        keep = nms(rects, nms_threshold)        rects = rects[keep, :]        # write result to file        for rect in rects:            cv2.rectangle(image_ori,(int(rect[0]),int(rect[1])),(int(rect[2]),int(rect[3])),(0,255,0),3)        cv2.imshow("img", image_ori)    camera.release()    cv2.destroyAllWindows()

def rundensebox(img, dpu_runner):    """get tensor"""    inputTensors = dpu_runner.get_input_tensors()    outputTensors = dpu_runner.get_output_tensors()     input_ndim = tuple(inputTensors[0].dims)    # print(input_ndim)    input_fixpos = inputTensors[0].get_attr("fix_point")    input_scale = 2**input_fixpos    output_ndim_0 = tuple(outputTensors[0].dims)     # print(output_ndim_0)    output_ndim_1 = tuple(outputTensors[1].dims)    # print(output_ndim_1)     output_fixpos_0 = outputTensors[0].get_attr("fix_point")    output_scale_0 = 1 / (2**output_fixpos_0)    output_fixpos_1 = outputTensors[1].get_attr("fix_point")    output_scale_1 = 1 / (2**output_fixpos_1)    imgquant = img * input_scale    runSize = input_ndim[0]    """prepare batch input/output """    inputData = [np.empty(input_ndim, dtype=np.int8, order="C")]    outputData = [np.empty(output_ndim_0, dtype=np.int8, order="C"),                np.empty(output_ndim_1, dtype=np.int8, order="C")]       """init input image to input buffer """    imageRun = inputData[0]    imageRun[0, ...] = imgquant.reshape(input_ndim[1:])       """run with batch """    job_id = dpu_runner.execute_async(inputData, outputData)    dpu_runner.wait(job_id)       result = [outputData[0] * output_scale_0,              outputData[1] * output_scale_1]    return result