{
  "id": 125819,
  "title": "If you are using mxresnet, here is small boost",
  "url": "/competitions/bengaliai-cv19/discussion/125819",
  "author_name": "",
  "post_date": "2020-01-13T22:02:33.043423800Z",
  "votes": 10,
  "comment_count": 3,
  "views": 0,
  "content": "<p>Try to use this version\n<a href=\"https://github.com/ducha-aiki/Ranger-Mish-ImageWoof-5/blob/master/mxresnet.py#L121\">https://github.com/ducha-aiki/Ranger-Mish-ImageWoof-5/blob/master/mxresnet.py#L121</a></p>\n\n<p>Tl;dr: replace avgpool2d with  MaxBlurPool2d from \"Making Convolutional Networks Shift-Invariant Again\" paper <a href=\"https://arxiv.org/abs/1904.11486\">https://arxiv.org/abs/1904.11486</a></p>\n\n<p>It improves my local CV (didn't submit here because I am not going to participate in this competition), as well ImagwWoof: <a href=\"https://forums.fast.ai/t/imagenette-imagewoof-leaderboards/45822/20?u=ducha-aiki\">https://forums.fast.ai/t/imagenette-imagewoof-leaderboards/45822/20?u=ducha-aiki</a></p>\n\n<p>Good luck to all</p>",
  "messages": [
    {
      "id": "718009",
      "postDate": "01/13/2020 22:02:33",
      "content": "<p>Try to use this version\n<a href=\"https://github.com/ducha-aiki/Ranger-Mish-ImageWoof-5/blob/master/mxresnet.py#L121\">https://github.com/ducha-aiki/Ranger-Mish-ImageWoof-5/blob/master/mxresnet.py#L121</a></p>\n\n<p>Tl;dr: replace avgpool2d with  MaxBlurPool2d from \"Making Convolutional Networks Shift-Invariant Again\" paper <a href=\"https://arxiv.org/abs/1904.11486\">https://arxiv.org/abs/1904.11486</a></p>\n\n<p>It improves my local CV (didn't submit here because I am not going to participate in this competition), as well ImagwWoof: <a href=\"https://forums.fast.ai/t/imagenette-imagewoof-leaderboards/45822/20?u=ducha-aiki\">https://forums.fast.ai/t/imagenette-imagewoof-leaderboards/45822/20?u=ducha-aiki</a></p>\n\n<p>Good luck to all</p>",
      "rawMarkdown": "Try to use this version\nhttps://github.com/ducha-aiki/Ranger-Mish-ImageWoof-5/blob/master/mxresnet.py#L121\n\nTl;dr: replace avgpool2d with  MaxBlurPool2d from \"Making Convolutional Networks Shift-Invariant Again\" paper https://arxiv.org/abs/1904.11486\n\nIt improves my local CV (didn't submit here because I am not going to participate in this competition), as well ImagwWoof: https://forums.fast.ai/t/imagenette-imagewoof-leaderboards/45822/20?u=ducha-aiki\n\nGood luck to all",
      "votes": null
    },
    {
      "id": "718013",
      "postDate": "01/13/2020 22:08:52",
      "content": "<p>Thank you for this!! To what I know <code>mxresnet</code> does not have any pretrained version, so do you think it could match other models with pretrained versions?</p>",
      "rawMarkdown": "Thank you for this!! To what I know `mxresnet` does not have any pretrained version, so do you think it could match other models with pretrained versions?",
      "votes": null
    },
    {
      "id": "718014",
      "postDate": "01/13/2020 22:10:16",
      "content": "<p>In my experiments, difference with resnet18 pretrained/not pretrained was small. Haven't tried too hard though</p>",
      "rawMarkdown": "In my experiments, difference with resnet18 pretrained/not pretrained was small. Haven't tried too hard though",
      "votes": null
    },
    {
      "id": "718234",
      "postDate": "01/14/2020 07:51:09",
      "content": "<p>I experimented with efficientnet-b0 and efficientnet-b4. Pretrained networks didn't make much difference for me :/. For b0, the score was worse for pretrained.   </p>",
      "rawMarkdown": "I experimented with efficientnet-b0 and efficientnet-b4. Pretrained networks didn't make much difference for me :/. For b0, the score was worse for pretrained.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 718013,
      "author_name": "bibek777",
      "author_url": "",
      "post_date": "01/13/2020 22:08:52",
      "content": "<p>Thank you for this!! To what I know <code>mxresnet</code> does not have any pretrained version, so do you think it could match other models with pretrained versions?</p>",
      "votes": null,
      "replies": [
        {
          "id": 718014,
          "author_name": "oldufo",
          "author_url": "",
          "post_date": "01/13/2020 22:10:16",
          "content": "<p>In my experiments, difference with resnet18 pretrained/not pretrained was small. Haven't tried too hard though</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 718234,
          "author_name": "timetraveller98",
          "author_url": "",
          "post_date": "01/14/2020 07:51:09",
          "content": "<p>I experimented with efficientnet-b0 and efficientnet-b4. Pretrained networks didn't make much difference for me :/. For b0, the score was worse for pretrained.   </p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "718009": "Try to use this version\nhttps://github.com/ducha-aiki/Ranger-Mish-ImageWoof-5/blob/master/mxresnet.py#L121\n\nTl;dr: replace avgpool2d with  MaxBlurPool2d from \"Making Convolutional Networks Shift-Invariant Again\" paper https://arxiv.org/abs/1904.11486\n\nIt improves my local CV (didn't submit here because I am not going to participate in this competition), as well ImagwWoof: https://forums.fast.ai/t/imagenette-imagewoof-leaderboards/45822/20?u=ducha-aiki\n\nGood luck to all",
    "718013": "Thank you for this!! To what I know `mxresnet` does not have any pretrained version, so do you think it could match other models with pretrained versions?",
    "718014": "In my experiments, difference with resnet18 pretrained/not pretrained was small. Haven't tried too hard though",
    "718234": "I experimented with efficientnet-b0 and efficientnet-b4. Pretrained networks didn't make much difference for me :/. For b0, the score was worse for pretrained."
  },
  "source": "meta"
}