{
  "id": 70613,
  "title": "Any good score with yolov3?",
  "url": "/competitions/rsna-pneumonia-detection-challenge/discussion/70613",
  "author_name": "",
  "post_date": "2018-11-05T21:35:57.568194300Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Could you share your solution?</p>",
  "messages": [
    {
      "id": "415892",
      "postDate": "11/05/2018 21:35:57",
      "content": "<p>Could you share your solution?</p>",
      "rawMarkdown": "Could you share your solution?",
      "votes": null
    },
    {
      "id": "417622",
      "postDate": "11/08/2018 14:34:08",
      "content": "<p>I would be interested in this as well - we were trying to implement a inceptionresnetv2 classifier pipeline to a YOLOv3 segmenter but ran out of time on the classifier so just submitted the best YOLOv3 model which we expected would underperform.  It did, yielding a middling 0.14611.  We experimented training the YOLO model with one class (opacity), two classes (opacity &amp; normal) and all three classes.  The two class YOLO performed best, and the three class yolo worst (0.04)  Best results on YOLO were obtained with agumentation on the original images with mild rotation and scaling.</p>\n\n<p>Hyperparameter selection on our YOLOv3 was:\nsize: 704x704\nsaturation - 1.0\n5411 max_batches (yielded 0.201 on stage 1 LB) - note this is opposite of advice to train for &gt;10,000 steps\nthreshold = 0.18</p>\n\n<p>Obviously, we overfit to stage 1 data - ensembling YOLO (planned, but again didn't get to it) would have helped.  First kaggle competition - chased the LB - lesson learned!</p>",
      "rawMarkdown": "I would be interested in this as well - we were trying to implement a inceptionresnetv2 classifier pipeline to a YOLOv3 segmenter but ran out of time on the classifier so just submitted the best YOLOv3 model which we expected would underperform.  It did, yielding a middling 0.14611.  We experimented training the YOLO model with one class (opacity), two classes (opacity &amp; normal) and all three classes.  The two class YOLO performed best, and the three class yolo worst (0.04)  Best results on YOLO were obtained with agumentation on the original images with mild rotation and scaling.\n\n\nHyperparameter selection on our YOLOv3 was:\nsize: 704x704\nsaturation - 1.0\n5411 max_batches (yielded 0.201 on stage 1 LB) - note this is opposite of advice to train for &gt;10,000 steps\nthreshold = 0.18\n\nObviously, we overfit to stage 1 data - ensembling YOLO (planned, but again didn't get to it) would have helped.  First kaggle competition - chased the LB - lesson learned!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 417622,
      "author_name": "drsxr1",
      "author_url": "",
      "post_date": "11/08/2018 14:34:08",
      "content": "<p>I would be interested in this as well - we were trying to implement a inceptionresnetv2 classifier pipeline to a YOLOv3 segmenter but ran out of time on the classifier so just submitted the best YOLOv3 model which we expected would underperform.  It did, yielding a middling 0.14611.  We experimented training the YOLO model with one class (opacity), two classes (opacity &amp; normal) and all three classes.  The two class YOLO performed best, and the three class yolo worst (0.04)  Best results on YOLO were obtained with agumentation on the original images with mild rotation and scaling.</p>\n\n<p>Hyperparameter selection on our YOLOv3 was:\nsize: 704x704\nsaturation - 1.0\n5411 max_batches (yielded 0.201 on stage 1 LB) - note this is opposite of advice to train for &gt;10,000 steps\nthreshold = 0.18</p>\n\n<p>Obviously, we overfit to stage 1 data - ensembling YOLO (planned, but again didn't get to it) would have helped.  First kaggle competition - chased the LB - lesson learned!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "415892": "Could you share your solution?",
    "417622": "I would be interested in this as well - we were trying to implement a inceptionresnetv2 classifier pipeline to a YOLOv3 segmenter but ran out of time on the classifier so just submitted the best YOLOv3 model which we expected would underperform.  It did, yielding a middling 0.14611.  We experimented training the YOLO model with one class (opacity), two classes (opacity &amp; normal) and all three classes.  The two class YOLO performed best, and the three class yolo worst (0.04)  Best results on YOLO were obtained with agumentation on the original images with mild rotation and scaling.\n\n\nHyperparameter selection on our YOLOv3 was:\nsize: 704x704\nsaturation - 1.0\n5411 max_batches (yielded 0.201 on stage 1 LB) - note this is opposite of advice to train for &gt;10,000 steps\nthreshold = 0.18\n\nObviously, we overfit to stage 1 data - ensembling YOLO (planned, but again didn't get to it) would have helped.  First kaggle competition - chased the LB - lesson learned!"
  },
  "source": "meta"
}