{
  "id": 308943,
  "title": "Just for fun: ScaledYOLOv4 for COTS",
  "url": "/competitions/tensorflow-great-barrier-reef/discussion/308943",
  "author_name": "Alex Wong",
  "post_date": "2022-02-21T05:04:47.001000",
  "votes": 3,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Just announcing I created a set of training / inference notebooks to use ScaledYOLOv4 for COTS detection. Just for fun.</p>\n<p>Training: <a href=\"https://www.kaggle.com/alexchwong/scaledyolov4-for-cots-training\" target=\"_blank\">https://www.kaggle.com/alexchwong/scaledyolov4-for-cots-training</a><br>\nInference: <a href=\"https://www.kaggle.com/alexchwong/scaledyolov4-for-cots-inference\" target=\"_blank\">https://www.kaggle.com/alexchwong/scaledyolov4-for-cots-inference</a></p>\n<p>A simple P5-based model trained at 1280p resolution for 6 epochs at default settings achieves 0.484 public LB and 0.565 private LB. I'm sure that heavier training approaches using the larger versions of this model will achieve higher scores [WIP].</p>\n<p>I'm surprised this has not been used for the competition. I suspect this is in part because WongKinYiu's ScaledYOLOv4 repo is not very well maintained. Installation is finicky and there are many game-breaking bugs. I have implemented a few tweaks to make ScaledYOLOv4 functional for Kaggle's current version of torch (1.9.1).</p>\n<p>Installation notebook to download version from my github fork, and pre-trained models (attach as dataset to your notebooks): <a href=\"https://www.kaggle.com/alexchwong/scaledyolov4-installation\" target=\"_blank\">https://www.kaggle.com/alexchwong/scaledyolov4-installation</a></p>\n<p>Did anyone use YOLOv4 and/or related versions?</p>",
  "messages": [
    {
      "id": 1699296,
      "postDate": "2022-02-21T05:04:47Z",
      "content": "<p>Just announcing I created a set of training / inference notebooks to use ScaledYOLOv4 for COTS detection. Just for fun.</p>\n<p>Training: <a href=\"https://www.kaggle.com/alexchwong/scaledyolov4-for-cots-training\" target=\"_blank\">https://www.kaggle.com/alexchwong/scaledyolov4-for-cots-training</a><br>\nInference: <a href=\"https://www.kaggle.com/alexchwong/scaledyolov4-for-cots-inference\" target=\"_blank\">https://www.kaggle.com/alexchwong/scaledyolov4-for-cots-inference</a></p>\n<p>A simple P5-based model trained at 1280p resolution for 6 epochs at default settings achieves 0.484 public LB and 0.565 private LB. I'm sure that heavier training approaches using the larger versions of this model will achieve higher scores [WIP].</p>\n<p>I'm surprised this has not been used for the competition. I suspect this is in part because WongKinYiu's ScaledYOLOv4 repo is not very well maintained. Installation is finicky and there are many game-breaking bugs. I have implemented a few tweaks to make ScaledYOLOv4 functional for Kaggle's current version of torch (1.9.1).</p>\n<p>Installation notebook to download version from my github fork, and pre-trained models (attach as dataset to your notebooks): <a href=\"https://www.kaggle.com/alexchwong/scaledyolov4-installation\" target=\"_blank\">https://www.kaggle.com/alexchwong/scaledyolov4-installation</a></p>\n<p>Did anyone use YOLOv4 and/or related versions?</p>",
      "rawMarkdown": "Just announcing I created a set of training / inference notebooks to use ScaledYOLOv4 for COTS detection. Just for fun.\n\nTraining: https://www.kaggle.com/alexchwong/scaledyolov4-for-cots-training\nInference: https://www.kaggle.com/alexchwong/scaledyolov4-for-cots-inference\n\nA simple P5-based model trained at 1280p resolution for 6 epochs at default settings achieves 0.484 public LB and 0.565 private LB. I'm sure that heavier training approaches using the larger versions of this model will achieve higher scores [WIP].\n\nI'm surprised this has not been used for the competition. I suspect this is in part because WongKinYiu's ScaledYOLOv4 repo is not very well maintained. Installation is finicky and there are many game-breaking bugs. I have implemented a few tweaks to make ScaledYOLOv4 functional for Kaggle's current version of torch (1.9.1).\n\nInstallation notebook to download version from my github fork, and pre-trained models (attach as dataset to your notebooks): https://www.kaggle.com/alexchwong/scaledyolov4-installation\n\nDid anyone use YOLOv4 and/or related versions?",
      "votes": 3
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1699296": "Just announcing I created a set of training / inference notebooks to use ScaledYOLOv4 for COTS detection. Just for fun.\n\nTraining: https://www.kaggle.com/alexchwong/scaledyolov4-for-cots-training\nInference: https://www.kaggle.com/alexchwong/scaledyolov4-for-cots-inference\n\nA simple P5-based model trained at 1280p resolution for 6 epochs at default settings achieves 0.484 public LB and 0.565 private LB. I'm sure that heavier training approaches using the larger versions of this model will achieve higher scores [WIP].\n\nI'm surprised this has not been used for the competition. I suspect this is in part because WongKinYiu's ScaledYOLOv4 repo is not very well maintained. Installation is finicky and there are many game-breaking bugs. I have implemented a few tweaks to make ScaledYOLOv4 functional for Kaggle's current version of torch (1.9.1).\n\nInstallation notebook to download version from my github fork, and pre-trained models (attach as dataset to your notebooks): https://www.kaggle.com/alexchwong/scaledyolov4-installation\n\nDid anyone use YOLOv4 and/or related versions?"
  }
}