{
  "id": 163502,
  "title": "Some Noobie Doubts about Object Detection",
  "url": "/competitions/global-wheat-detection/discussion/163502",
  "author_name": "",
  "post_date": "2020-07-02T09:45:33.017546400Z",
  "votes": 6,
  "comment_count": 6,
  "views": 0,
  "content": "<p>As a noob in Object Detection I am learning by trail and error , along with theory to understand how things work here. Here are some questions of which I didn't find a good answer to after googling -\n* What does Adam perform poorly with FasterRCNN ? and my SGD and momentum work better?\n* I have tried to understand yolo code , but it really grows up on me and I feel lost as I find it too complex, any links ,resources will be highly valuable\n* What would be a good epoc\nAnswers to above will be helpful for me , I will also add more questions as I dig in deeper , thanks to the community in advance for help</p>",
  "messages": [
    {
      "id": "912198",
      "postDate": "07/02/2020 09:45:33",
      "content": "<p>As a noob in Object Detection I am learning by trail and error , along with theory to understand how things work here. Here are some questions of which I didn't find a good answer to after googling -\n* What does Adam perform poorly with FasterRCNN ? and my SGD and momentum work better?\n* I have tried to understand yolo code , but it really grows up on me and I feel lost as I find it too complex, any links ,resources will be highly valuable\n* What would be a good epoc\nAnswers to above will be helpful for me , I will also add more questions as I dig in deeper , thanks to the community in advance for help</p>",
      "rawMarkdown": "As a noob in Object Detection I am learning by trail and error , along with theory to understand how things work here. Here are some questions of which I didn't find a good answer to after googling -\n* What does Adam perform poorly with FasterRCNN ? and my SGD and momentum work better?\n* I have tried to understand yolo code , but it really grows up on me and I feel lost as I find it too complex, any links ,resources will be highly valuable\n* What would be a good epoc\nAnswers to above will be helpful for me , I will also add more questions as I dig in deeper , thanks to the community in advance for help",
      "votes": null
    },
    {
      "id": "912962",
      "postDate": "07/02/2020 21:16:25",
      "content": "<p>I actually had the same problem using Adam optimizer with FasterRCNN. I think using a lower learning rate like 0.00005 helps a little bit.</p>",
      "rawMarkdown": "I actually had the same problem using Adam optimizer with FasterRCNN. I think using a lower learning rate like 0.00005 helps a little bit.",
      "votes": null
    },
    {
      "id": "913531",
      "postDate": "07/03/2020 09:05:36",
      "content": "<p>There are some really nice notebooks to start with :)</p>",
      "rawMarkdown": "There are some really nice notebooks to start with :)",
      "votes": null
    },
    {
      "id": "913722",
      "postDate": "07/03/2020 11:24:46",
      "content": "<p>On the yolo question, if you have some time there is no better resource than Andrew Ngs Convolutional Neural Networks course: <a href=\"https://www.coursera.org/learn/convolutional-neural-networks?specialization=deep-learning\">https://www.coursera.org/learn/convolutional-neural-networks?specialization=deep-learning</a>. Week 3 is the part about object detection which includes a really good look at yolo.</p>\n\n<p>Otherwise Joseph Redmons tech reports are reasonably quick reads and go into the detail of how yolo works. I personally found these useful though they do make some assumptions about your knowledge and leaps in logic in places. Here's the link to version 3 which I think is the easiest read: <a href=\"https://pjreddie.com/media/files/papers/YOLOv3.pdf\">https://pjreddie.com/media/files/papers/YOLOv3.pdf</a>. </p>",
      "rawMarkdown": "On the yolo question, if you have some time there is no better resource than Andrew Ngs Convolutional Neural Networks course: https://www.coursera.org/learn/convolutional-neural-networks?specialization=deep-learning. Week 3 is the part about object detection which includes a really good look at yolo.\n\nOtherwise Joseph Redmons tech reports are reasonably quick reads and go into the detail of how yolo works. I personally found these useful though they do make some assumptions about your knowledge and leaps in logic in places. Here's the link to version 3 which I think is the easiest read: https://pjreddie.com/media/files/papers/YOLOv3.pdf.",
      "votes": null
    },
    {
      "id": "914999",
      "postDate": "07/04/2020 12:10:27",
      "content": "<p>Yeah I did start with the same haha , just wanted to know why SGD works better with FRCNN</p>",
      "rawMarkdown": "Yeah I did start with the same haha , just wanted to know why SGD works better with FRCNN",
      "votes": null
    },
    {
      "id": "915000",
      "postDate": "07/04/2020 12:10:55",
      "content": "<p>Yeah , adam requires low lr otherwise gradients will start to blow</p>",
      "rawMarkdown": "Yeah , adam requires low lr otherwise gradients will start to blow",
      "votes": null
    },
    {
      "id": "915001",
      "postDate": "07/04/2020 12:11:37",
      "content": "<p>Thanks <a href=\"/mattbast\">@mattbast</a> Really appreciate it</p>",
      "rawMarkdown": "Thanks @mattbast Really appreciate it",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 912962,
      "author_name": "jonykarki",
      "author_url": "",
      "post_date": "07/02/2020 21:16:25",
      "content": "<p>I actually had the same problem using Adam optimizer with FasterRCNN. I think using a lower learning rate like 0.00005 helps a little bit.</p>",
      "votes": null,
      "replies": [
        {
          "id": 915000,
          "author_name": "tanulsingh077",
          "author_url": "",
          "post_date": "07/04/2020 12:10:55",
          "content": "<p>Yeah , adam requires low lr otherwise gradients will start to blow</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 913531,
      "author_name": "bendvd",
      "author_url": "",
      "post_date": "07/03/2020 09:05:36",
      "content": "<p>There are some really nice notebooks to start with :)</p>",
      "votes": null,
      "replies": [
        {
          "id": 914999,
          "author_name": "tanulsingh077",
          "author_url": "",
          "post_date": "07/04/2020 12:10:27",
          "content": "<p>Yeah I did start with the same haha , just wanted to know why SGD works better with FRCNN</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 913722,
      "author_name": "mattbast",
      "author_url": "",
      "post_date": "07/03/2020 11:24:46",
      "content": "<p>On the yolo question, if you have some time there is no better resource than Andrew Ngs Convolutional Neural Networks course: <a href=\"https://www.coursera.org/learn/convolutional-neural-networks?specialization=deep-learning\">https://www.coursera.org/learn/convolutional-neural-networks?specialization=deep-learning</a>. Week 3 is the part about object detection which includes a really good look at yolo.</p>\n\n<p>Otherwise Joseph Redmons tech reports are reasonably quick reads and go into the detail of how yolo works. I personally found these useful though they do make some assumptions about your knowledge and leaps in logic in places. Here's the link to version 3 which I think is the easiest read: <a href=\"https://pjreddie.com/media/files/papers/YOLOv3.pdf\">https://pjreddie.com/media/files/papers/YOLOv3.pdf</a>. </p>",
      "votes": null,
      "replies": [
        {
          "id": 915001,
          "author_name": "tanulsingh077",
          "author_url": "",
          "post_date": "07/04/2020 12:11:37",
          "content": "<p>Thanks <a href=\"/mattbast\">@mattbast</a> Really appreciate it</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "912198": "As a noob in Object Detection I am learning by trail and error , along with theory to understand how things work here. Here are some questions of which I didn't find a good answer to after googling -\n* What does Adam perform poorly with FasterRCNN ? and my SGD and momentum work better?\n* I have tried to understand yolo code , but it really grows up on me and I feel lost as I find it too complex, any links ,resources will be highly valuable\n* What would be a good epoc\nAnswers to above will be helpful for me , I will also add more questions as I dig in deeper , thanks to the community in advance for help",
    "912962": "I actually had the same problem using Adam optimizer with FasterRCNN. I think using a lower learning rate like 0.00005 helps a little bit.",
    "913531": "There are some really nice notebooks to start with :)",
    "913722": "On the yolo question, if you have some time there is no better resource than Andrew Ngs Convolutional Neural Networks course: https://www.coursera.org/learn/convolutional-neural-networks?specialization=deep-learning. Week 3 is the part about object detection which includes a really good look at yolo.\n\nOtherwise Joseph Redmons tech reports are reasonably quick reads and go into the detail of how yolo works. I personally found these useful though they do make some assumptions about your knowledge and leaps in logic in places. Here's the link to version 3 which I think is the easiest read: https://pjreddie.com/media/files/papers/YOLOv3.pdf.",
    "914999": "Yeah I did start with the same haha , just wanted to know why SGD works better with FRCNN",
    "915000": "Yeah , adam requires low lr otherwise gradients will start to blow",
    "915001": "Thanks @mattbast Really appreciate it"
  },
  "source": "meta"
}