{
  "id": 143785,
  "title": "Is there a budget option for semi-supervised",
  "url": "/competitions/semi-inat-2020/discussion/143785",
  "author_name": "",
  "post_date": "2020-04-16T08:53:34.519549300Z",
  "votes": 2,
  "comment_count": 6,
  "views": 0,
  "content": "<p>It seems that compute resources limit participation in this challenge. Some of the SOTA approaches based on contrastive learning like SimCLR uses 8000+ batchsize and 32 TPUs for ImageNet trining which is 5x the size of this dataset. Also MoCo takes 2.5 days to converge with 8 V100 instances. </p>\n\n<p>My question is what would be the best approach to semi-supervised learning given limited budget? </p>\n\n<p>PS: I hope the top guys wan't be silent forever.</p>",
  "messages": [
    {
      "id": "809509",
      "postDate": "04/16/2020 08:53:34",
      "content": "<p>It seems that compute resources limit participation in this challenge. Some of the SOTA approaches based on contrastive learning like SimCLR uses 8000+ batchsize and 32 TPUs for ImageNet trining which is 5x the size of this dataset. Also MoCo takes 2.5 days to converge with 8 V100 instances. </p>\n\n<p>My question is what would be the best approach to semi-supervised learning given limited budget? </p>\n\n<p>PS: I hope the top guys wan't be silent forever.</p>",
      "rawMarkdown": "It seems that compute resources limit participation in this challenge. Some of the SOTA approaches based on contrastive learning like SimCLR uses 8000+ batchsize and 32 TPUs for ImageNet trining which is 5x the size of this dataset. Also MoCo takes 2.5 days to converge with 8 V100 instances. \n\nMy question is what would be the best approach to semi-supervised learning given limited budget? \n\nPS: I hope the top guys wan't be silent forever.",
      "votes": null
    },
    {
      "id": "833428",
      "postDate": "05/04/2020 19:55:50",
      "content": "<p>Have you attempted simple pseudo-labeling? I have tried it but got poor validation scores, so curious if others have implemented this simple baseline as well.</p>\n\n<p>I guess with a supervised baseline of about 40-50% error rate, it is hard to use the model to properly pseudolabel the dataset.</p>",
      "rawMarkdown": "Have you attempted simple pseudo-labeling? I have tried it but got poor validation scores, so curious if others have implemented this simple baseline as well.\n\nI guess with a supervised baseline of about 40-50% error rate, it is hard to use the model to properly pseudolabel the dataset.",
      "votes": null
    },
    {
      "id": "833852",
      "postDate": "05/05/2020 05:22:08",
      "content": "<p>no. i have not. it's just supervision without tricks.  seresnext50 + efficientnetB3. Pseudo labeling is rather primitive technique to combat loaded and silent gray geese :) not worth the effort i am afraid. if they would take training time into account than we would have chance to contribute. like this hardly.</p>",
      "rawMarkdown": "no. i have not. it's just supervision without tricks.  seresnext50 + efficientnetB3. Pseudo labeling is rather primitive technique to combat loaded and silent gray geese :) not worth the effort i am afraid. if they would take training time into account than we would have chance to contribute. like this hardly.",
      "votes": null
    },
    {
      "id": "839014",
      "postDate": "05/09/2020 02:10:49",
      "content": "<p>Thanks for the suggestion! However, I think it's not good to change the rules in the middle of the competition. It is also hard to set a fixed budget since everyone has different equipment. We will re-evaluate the policy maybe for the next year.</p>",
      "rawMarkdown": "Thanks for the suggestion! However, I think it's not good to change the rules in the middle of the competition. It is also hard to set a fixed budget since everyone has different equipment. We will re-evaluate the policy maybe for the next year.",
      "votes": null
    },
    {
      "id": "839123",
      "postDate": "05/09/2020 05:21:01",
      "content": "<p>i am not saying you should change the rules. I am asking weather you know some budget option for semi-supervised. </p>\n\n<p>Now I do have a suggestion and that is to ask participants to provide training times. if that is not to much hustle. It would be just an orientation.</p>",
      "rawMarkdown": "i am not saying you should change the rules. I am asking weather you know some budget option for semi-supervised. \n\nNow I do have a suggestion and that is to ask participants to provide training times. if that is not to much hustle. It would be just an orientation.",
      "votes": null
    },
    {
      "id": "839125",
      "postDate": "05/09/2020 05:26:57",
      "content": "<p>Yes we will ask the winners to present their solutions, including the architecture and training details, at the workshop.</p>",
      "rawMarkdown": "Yes we will ask the winners to present their solutions, including the architecture and training details, at the workshop.",
      "votes": null
    },
    {
      "id": "839145",
      "postDate": "05/09/2020 05:41:14",
      "content": "<p>Please, share it on forum afterwards. Thanks </p>",
      "rawMarkdown": "Please, share it on forum afterwards. Thanks",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 833428,
      "author_name": "tanlikesmath",
      "author_url": "",
      "post_date": "05/04/2020 19:55:50",
      "content": "<p>Have you attempted simple pseudo-labeling? I have tried it but got poor validation scores, so curious if others have implemented this simple baseline as well.</p>\n\n<p>I guess with a supervised baseline of about 40-50% error rate, it is hard to use the model to properly pseudolabel the dataset.</p>",
      "votes": null,
      "replies": [
        {
          "id": 833852,
          "author_name": "valanm",
          "author_url": "",
          "post_date": "05/05/2020 05:22:08",
          "content": "<p>no. i have not. it's just supervision without tricks.  seresnext50 + efficientnetB3. Pseudo labeling is rather primitive technique to combat loaded and silent gray geese :) not worth the effort i am afraid. if they would take training time into account than we would have chance to contribute. like this hardly.</p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 839014,
      "author_name": "jcfredsu",
      "author_url": "",
      "post_date": "05/09/2020 02:10:49",
      "content": "<p>Thanks for the suggestion! However, I think it's not good to change the rules in the middle of the competition. It is also hard to set a fixed budget since everyone has different equipment. We will re-evaluate the policy maybe for the next year.</p>",
      "votes": null,
      "replies": [
        {
          "id": 839123,
          "author_name": "valanm",
          "author_url": "",
          "post_date": "05/09/2020 05:21:01",
          "content": "<p>i am not saying you should change the rules. I am asking weather you know some budget option for semi-supervised. </p>\n\n<p>Now I do have a suggestion and that is to ask participants to provide training times. if that is not to much hustle. It would be just an orientation.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 839125,
          "author_name": "jcfredsu",
          "author_url": "",
          "post_date": "05/09/2020 05:26:57",
          "content": "<p>Yes we will ask the winners to present their solutions, including the architecture and training details, at the workshop.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 839145,
          "author_name": "valanm",
          "author_url": "",
          "post_date": "05/09/2020 05:41:14",
          "content": "<p>Please, share it on forum afterwards. Thanks </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "809509": "It seems that compute resources limit participation in this challenge. Some of the SOTA approaches based on contrastive learning like SimCLR uses 8000+ batchsize and 32 TPUs for ImageNet trining which is 5x the size of this dataset. Also MoCo takes 2.5 days to converge with 8 V100 instances. \n\nMy question is what would be the best approach to semi-supervised learning given limited budget? \n\nPS: I hope the top guys wan't be silent forever.",
    "833428": "Have you attempted simple pseudo-labeling? I have tried it but got poor validation scores, so curious if others have implemented this simple baseline as well.\n\nI guess with a supervised baseline of about 40-50% error rate, it is hard to use the model to properly pseudolabel the dataset.",
    "833852": "no. i have not. it's just supervision without tricks.  seresnext50 + efficientnetB3. Pseudo labeling is rather primitive technique to combat loaded and silent gray geese :) not worth the effort i am afraid. if they would take training time into account than we would have chance to contribute. like this hardly.",
    "839014": "Thanks for the suggestion! However, I think it's not good to change the rules in the middle of the competition. It is also hard to set a fixed budget since everyone has different equipment. We will re-evaluate the policy maybe for the next year.",
    "839123": "i am not saying you should change the rules. I am asking weather you know some budget option for semi-supervised. \n\nNow I do have a suggestion and that is to ask participants to provide training times. if that is not to much hustle. It would be just an orientation.",
    "839125": "Yes we will ask the winners to present their solutions, including the architecture and training details, at the workshop.",
    "839145": "Please, share it on forum afterwards. Thanks"
  },
  "source": "meta"
}