{
  "id": 78091,
  "title": "Siamese inestable results",
  "url": "/competitions/humpback-whale-identification/discussion/78091",
  "author_name": "",
  "post_date": "2019-01-19T16:52:54.762652100Z",
  "votes": 2,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Hi, I'm a little impatient and during the Siamese training I like to checkpoints and submissions ... The fact is that by making minor changes on the Siamese the results change a lot and when analyzing how the Siamese works long epochs I see that over the first 100 epochs gets a LB 0.6 (without new whales in the submission), then drops to LB 0.55 and then try to climb again following the approach of martin ... It happens the same, you have solved or treated this?</p>\n\n<p>When I try to make anneal, the siamesa works even more unstable (perhaps because it is made in an anneal way by changing the K). But I think it is the K that makes the results partly unstable.</p>",
  "messages": [
    {
      "id": "458443",
      "postDate": "01/19/2019 16:52:54",
      "content": "<p>Hi, I'm a little impatient and during the Siamese training I like to checkpoints and submissions ... The fact is that by making minor changes on the Siamese the results change a lot and when analyzing how the Siamese works long epochs I see that over the first 100 epochs gets a LB 0.6 (without new whales in the submission), then drops to LB 0.55 and then try to climb again following the approach of martin ... It happens the same, you have solved or treated this?</p>\n\n<p>When I try to make anneal, the siamesa works even more unstable (perhaps because it is made in an anneal way by changing the K). But I think it is the K that makes the results partly unstable.</p>",
      "rawMarkdown": "Hi, I'm a little impatient and during the Siamese training I like to checkpoints and submissions ... The fact is that by making minor changes on the Siamese the results change a lot and when analyzing how the Siamese works long epochs I see that over the first 100 epochs gets a LB 0.6 (without new whales in the submission), then drops to LB 0.55 and then try to climb again following the approach of martin ... It happens the same, you have solved or treated this?\n\nWhen I try to make anneal, the siamesa works even more unstable (perhaps because it is made in an anneal way by changing the K). But I think it is the K that makes the results partly unstable.",
      "votes": null
    },
    {
      "id": "458706",
      "postDate": "01/20/2019 10:00:20",
      "content": "<p>Check out the original kernel from Martin. There he explains his train schedule and the issue with overfitting (and why his schedule can beat it).</p>\n\n<p>I am trying to use his kernel as a baseline, however all results I can achieve are worse than his pretrained model. </p>",
      "rawMarkdown": "Check out the original kernel from Martin. There he explains his train schedule and the issue with overfitting (and why his schedule can beat it).\n\nI am trying to use his kernel as a baseline, however all results I can achieve are worse than his pretrained model.",
      "votes": null
    },
    {
      "id": "459009",
      "postDate": "01/21/2019 02:08:26",
      "content": "<p>same issue for me, training is very unstable. By the way, did you use bounding box during training? It seems that adding bbox worse the results for me.</p>",
      "rawMarkdown": "same issue for me, training is very unstable. By the way, did you use bounding box during training? It seems that adding bbox worse the results for me.",
      "votes": null
    },
    {
      "id": "459147",
      "postDate": "01/21/2019 09:24:25",
      "content": "<p>Yep I am adding bbox and had not tried without... I will try!</p>",
      "rawMarkdown": "Yep I am adding bbox and had not tried without... I will try!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 458706,
      "author_name": "spsancti",
      "author_url": "",
      "post_date": "01/20/2019 10:00:20",
      "content": "<p>Check out the original kernel from Martin. There he explains his train schedule and the issue with overfitting (and why his schedule can beat it).</p>\n\n<p>I am trying to use his kernel as a baseline, however all results I can achieve are worse than his pretrained model. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 459009,
      "author_name": "zjucor",
      "author_url": "",
      "post_date": "01/21/2019 02:08:26",
      "content": "<p>same issue for me, training is very unstable. By the way, did you use bounding box during training? It seems that adding bbox worse the results for me.</p>",
      "votes": null,
      "replies": [
        {
          "id": 459147,
          "author_name": "maparla",
          "author_url": "",
          "post_date": "01/21/2019 09:24:25",
          "content": "<p>Yep I am adding bbox and had not tried without... I will try!</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "458443": "Hi, I'm a little impatient and during the Siamese training I like to checkpoints and submissions ... The fact is that by making minor changes on the Siamese the results change a lot and when analyzing how the Siamese works long epochs I see that over the first 100 epochs gets a LB 0.6 (without new whales in the submission), then drops to LB 0.55 and then try to climb again following the approach of martin ... It happens the same, you have solved or treated this?\n\nWhen I try to make anneal, the siamesa works even more unstable (perhaps because it is made in an anneal way by changing the K). But I think it is the K that makes the results partly unstable.",
    "458706": "Check out the original kernel from Martin. There he explains his train schedule and the issue with overfitting (and why his schedule can beat it).\n\nI am trying to use his kernel as a baseline, however all results I can achieve are worse than his pretrained model.",
    "459009": "same issue for me, training is very unstable. By the way, did you use bounding box during training? It seems that adding bbox worse the results for me.",
    "459147": "Yep I am adding bbox and had not tried without... I will try!"
  },
  "source": "meta"
}