高分辨率图像的目标检测推理在cpu上花费了大量的时间

Question

我已经在pascal数据集上训练了ML模型，图像大小为224，但是当在新图像(一些是高分辨率的，一些是比pascal图像略高的)上推断时，我在pil2tensor()中得到了错误

@app.route('/analyze', methods=['POST'])
async def analyze(request):
    data = await request.form()
    img_bytes = await (data['file'].read())
    img = open_image(BytesIO(img_bytes))

    t_img= PIL.Image.open(BytesIO(img_bytes)).convert('RGB')
    t_img = pil2tensor(t_img, np.float32)
    t_img = t_img.div_(255)
    with torch.no_grad():
        # test_output = learn.model.eval()(t_img.unsqueeze_(0).cuda())
        test_output = learn.model.eval()(t_img.unsqueeze_(0))

对于较小的图像尺寸(如来自google的一些低分辨率图像)，ML模型能够在几秒钟内做出正确的推断，但对于稍高分辨率的图像，它需要大约20-40分钟！

Answer 1

修复了这个问题，下面是正确的代码：

@app.route('/analyze', methods=['POST'])
async def analyze(request):
    data = await request.form()
    img_bytes = await (data['file'].read())
    img = open_image(BytesIO(img_bytes))
    localtime = _utc_to_local(datetime.utcnow())

    current_dir = os.path.dirname(__file__)
    media_path = os.path.join(current_dir, "media")
    media_path_original = os.path.join(media_path, "original")
    media_path_processed = os.path.join(media_path, "processed")

    img_path = os.path.join(media_path_original, localtime+".jpg")
    img.save(img_path)
    verify_image(Path(img_path), idx=0, delete=False, max_size=600, dest=Path(media_path_processed))
    processed_img_path = os.path.join(media_path_processed, localtime+".jpg")
    processed_img = open_image(processed_img_path) if os.path.exists(processed_img_path) else img
    processed_img.refresh()
    with torch.no_grad():
        test_output = learn.model.eval()(processed_img.data.unsqueeze(0))
        predictions = show_preds(processed_img, test_output, 0, detect_thresh=0.4, classes=t_classes)
    img_height, img_width = processed_img.size
    return JSONResponse({'predictions': predictions, 'img_height': img_height, 'img_width': img_width})