Component recognize goes slow on colab

Question

Component recognize goes slow on colab

upupzealot opened this issue 4 years ago · comments

Component recognize goes slow on colab, with yolo-v5 model it suppose to be finish in less than an hour, but in my case it cost more than 10mins for only one epoch.
For detail, my pipline.json looks like

{
  "plugins": {
    "dataCollect": {
      "package": "@pipcook/plugins-object-detection-pascalvoc-data-collect",
      "params": {
        "url": "https://zhijiansha.oss-cn-hangzhou.aliyuncs.com/deep-learning/output.zip"
      }
    },
    "dataAccess": {
      "package": "@pipcook/plugins-coco-data-access"
    },
    "modelDefine": {
      "package": "@pipcook/plugins-pytorch-yolov5-model-define"
    },
    "modelTrain": {
      "package": "@pipcook/plugins-pytorch-yolov5-model-train",
      "params": {
        "epochs": 300
      }
    },
    "modelEvaluate": {
      "package": "@pipcook/plugins-pytorch-yolov5-model-evaluate"
    }
  }
}

Yorkie Makoto · Answer 1 · Sun Dec 20 2020 04:06:37 GMT+0800 (China Standard Time)

@FeelyChau @rickycao-qy it seems that we are just using a single core to train.

Yorkie Makoto · Answer 2 · Sun Dec 20 2020 05:43:06 GMT+0800 (China Standard Time)

And the given dataset structure is different from example's:

@rickycao-qy It seems the first one is valid but not working, see https://github.com/alibaba/pipcook/blob/main/docs/tutorials/component-object-detection.md#data-preparation.

upupzealot · Answer 3 · Mon Dec 21 2020 10:15:31 GMT+0800 (China Standard Time)

I write a program to generate these pics, both VOC and COCO format are supported, so tell me if I can help for format reason : )

upupzealot · Answer 4 · Mon Dec 21 2020 10:17:09 GMT+0800 (China Standard Time)

An other reason I can image is the size of the picture, in my collection, the width of pic is always 1000px.
Have no idea if that matters a lot