{
  "id": 84727,
  "title": "Can't get above 0.85 LB",
  "url": "/competitions/histopathologic-cancer-detection/discussion/84727",
  "author_name": "",
  "post_date": "2019-03-19T09:11:11.012393500Z",
  "votes": null,
  "comment_count": 5,
  "views": 0,
  "content": "<p>Hi,</p>\n\n<p>I've been working on this compeition for a couple weeks now and can't seem to get above 0.85 LB. I'm using Keras and transfer learning with ResNet50. <a href=\"https://www.kaggle.com/jtmurkz/keras-resnet50-can-t-seem-to-get-above-0-85-lb\">This is my code:</a></p>\n\n<p>My most recent attempt I used TTA and cropped the images to the central 48 pixels (as the data only uses the central 32 pixels). I got some improvement but not as much as I was expecting. It's now gotten to the point where I'm really not sure what else I can add/change to improve my score. I would be grateful if anyone could point me in the right direction.</p>\n\n<p>Thank you!</p>",
  "messages": [
    {
      "id": "493937",
      "postDate": "03/19/2019 09:11:11",
      "content": "<p>Hi,</p>\n\n<p>I've been working on this compeition for a couple weeks now and can't seem to get above 0.85 LB. I'm using Keras and transfer learning with ResNet50. <a href=\"https://www.kaggle.com/jtmurkz/keras-resnet50-can-t-seem-to-get-above-0-85-lb\">This is my code:</a></p>\n\n<p>My most recent attempt I used TTA and cropped the images to the central 48 pixels (as the data only uses the central 32 pixels). I got some improvement but not as much as I was expecting. It's now gotten to the point where I'm really not sure what else I can add/change to improve my score. I would be grateful if anyone could point me in the right direction.</p>\n\n<p>Thank you!</p>",
      "rawMarkdown": "Hi,\n\nI've been working on this compeition for a couple weeks now and can't seem to get above 0.85 LB. I'm using Keras and transfer learning with ResNet50. [This is my code:](https://www.kaggle.com/jtmurkz/keras-resnet50-can-t-seem-to-get-above-0-85-lb)\n\nMy most recent attempt I used TTA and cropped the images to the central 48 pixels (as the data only uses the central 32 pixels). I got some improvement but not as much as I was expecting. It's now gotten to the point where I'm really not sure what else I can add/change to improve my score. I would be grateful if anyone could point me in the right direction.\n\nThank you!",
      "votes": null
    },
    {
      "id": "494208",
      "postDate": "03/19/2019 15:22:27",
      "content": "<p>Smaller resnets seem to have difficulty with these images, I don’t usually consider 50 to be small but here I think it is. Without changing anything else you should be able to improve your score by using the full 96x96 images. Also you can try changing to preresnet50 or cbam-resnet. Pre can just be dropped in, though it’s a bit tricky to find good parameters for cbam and it’s slower due to the attention layers. </p>\n\n<p>Another architecture that can give nice results without much tuning is airnet, but it’s very slow. Good luck!</p>",
      "rawMarkdown": "Smaller resnets seem to have difficulty with these images, I don’t usually consider 50 to be small but here I think it is. Without changing anything else you should be able to improve your score by using the full 96x96 images. Also you can try changing to preresnet50 or cbam-resnet. Pre can just be dropped in, though it’s a bit tricky to find good parameters for cbam and it’s slower due to the attention layers. \n\nAnother architecture that can give nice results without much tuning is airnet, but it’s very slow. Good luck!",
      "votes": null
    },
    {
      "id": "494245",
      "postDate": "03/19/2019 15:56:21",
      "content": "<p>Thank you for the reply, I'll try experimenting with those architectures and see how I get on. Also why should I use the full 96x96 image when it's only dependent on the central 32 pixels? Is it to do with zero padding?</p>",
      "rawMarkdown": "Thank you for the reply, I'll try experimenting with those architectures and see how I get on. Also why should I use the full 96x96 image when it's only dependent on the central 32 pixels? Is it to do with zero padding?",
      "votes": null
    },
    {
      "id": "494290",
      "postDate": "03/19/2019 16:45:28",
      "content": "<p>Its true the organizers claim the label is determined exclusively by the central 32px, but I think the way they phrased it in the description is not doing them or anyone else any favors. They say a single pixel of metastasis in the center will make the label positive, so are we trying to classify single pixels? </p>\n\n<p>Though the label may be determined by the center this does not imply there is no useful information in the rest of the image so discarding 89% of the pixels likely throws out too much. </p>",
      "rawMarkdown": "Its true the organizers claim the label is determined exclusively by the central 32px, but I think the way they phrased it in the description is not doing them or anyone else any favors. They say a single pixel of metastasis in the center will make the label positive, so are we trying to classify single pixels? \n\nThough the label may be determined by the center this does not imply there is no useful information in the rest of the image so discarding 89% of the pixels likely throws out too much.",
      "votes": null
    },
    {
      "id": "494364",
      "postDate": "03/19/2019 18:57:20",
      "content": "<p>I tried center cropping based on the same ideas and it definitely works worse. Resizing to 196x196 or 244x244 works much much better. You should try that. </p>\n\n<p>P.S. I tested this with resnet50 and se_resnet50. In both case resizing to 196x196 or 244x244 worked tremendously better. </p>",
      "rawMarkdown": "I tried center cropping based on the same ideas and it definitely works worse. Resizing to 196x196 or 244x244 works much much better. You should try that. \n\nP.S. I tested this with resnet50 and se_resnet50. In both case resizing to 196x196 or 244x244 worked tremendously better.",
      "votes": null
    },
    {
      "id": "494638",
      "postDate": "03/20/2019 04:12:18",
      "content": "<p>HI, Jake. I have tried central crop 64 or 48, but it seems not as good as using the full size image(96x96). And resizing it to 224x224 may better if u use an imagenet pretrained models. And you can try ensemble,  which is a very good way to improve your score (<a href=\"https://www.kaggle.com/c/histopathologic-cancer-detection/discussion/83313\">see this topic</a>).</p>",
      "rawMarkdown": "HI, Jake. I have tried central crop 64 or 48, but it seems not as good as using the full size image(96x96). And resizing it to 224x224 may better if u use an imagenet pretrained models. And you can try ensemble,  which is a very good way to improve your score ([see this topic](https://www.kaggle.com/c/histopathologic-cancer-detection/discussion/83313)).",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 494208,
      "author_name": "interneuron",
      "author_url": "",
      "post_date": "03/19/2019 15:22:27",
      "content": "<p>Smaller resnets seem to have difficulty with these images, I don’t usually consider 50 to be small but here I think it is. Without changing anything else you should be able to improve your score by using the full 96x96 images. Also you can try changing to preresnet50 or cbam-resnet. Pre can just be dropped in, though it’s a bit tricky to find good parameters for cbam and it’s slower due to the attention layers. </p>\n\n<p>Another architecture that can give nice results without much tuning is airnet, but it’s very slow. Good luck!</p>",
      "votes": null,
      "replies": [
        {
          "id": 494245,
          "author_name": "jtmurkz",
          "author_url": "",
          "post_date": "03/19/2019 15:56:21",
          "content": "<p>Thank you for the reply, I'll try experimenting with those architectures and see how I get on. Also why should I use the full 96x96 image when it's only dependent on the central 32 pixels? Is it to do with zero padding?</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 494290,
          "author_name": "interneuron",
          "author_url": "",
          "post_date": "03/19/2019 16:45:28",
          "content": "<p>Its true the organizers claim the label is determined exclusively by the central 32px, but I think the way they phrased it in the description is not doing them or anyone else any favors. They say a single pixel of metastasis in the center will make the label positive, so are we trying to classify single pixels? </p>\n\n<p>Though the label may be determined by the center this does not imply there is no useful information in the rest of the image so discarding 89% of the pixels likely throws out too much. </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 494364,
      "author_name": "ivanpan",
      "author_url": "",
      "post_date": "03/19/2019 18:57:20",
      "content": "<p>I tried center cropping based on the same ideas and it definitely works worse. Resizing to 196x196 or 244x244 works much much better. You should try that. </p>\n\n<p>P.S. I tested this with resnet50 and se_resnet50. In both case resizing to 196x196 or 244x244 worked tremendously better. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 494638,
      "author_name": "seefun",
      "author_url": "",
      "post_date": "03/20/2019 04:12:18",
      "content": "<p>HI, Jake. I have tried central crop 64 or 48, but it seems not as good as using the full size image(96x96). And resizing it to 224x224 may better if u use an imagenet pretrained models. And you can try ensemble,  which is a very good way to improve your score (<a href=\"https://www.kaggle.com/c/histopathologic-cancer-detection/discussion/83313\">see this topic</a>).</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "493937": "Hi,\n\nI've been working on this compeition for a couple weeks now and can't seem to get above 0.85 LB. I'm using Keras and transfer learning with ResNet50. [This is my code:](https://www.kaggle.com/jtmurkz/keras-resnet50-can-t-seem-to-get-above-0-85-lb)\n\nMy most recent attempt I used TTA and cropped the images to the central 48 pixels (as the data only uses the central 32 pixels). I got some improvement but not as much as I was expecting. It's now gotten to the point where I'm really not sure what else I can add/change to improve my score. I would be grateful if anyone could point me in the right direction.\n\nThank you!",
    "494208": "Smaller resnets seem to have difficulty with these images, I don’t usually consider 50 to be small but here I think it is. Without changing anything else you should be able to improve your score by using the full 96x96 images. Also you can try changing to preresnet50 or cbam-resnet. Pre can just be dropped in, though it’s a bit tricky to find good parameters for cbam and it’s slower due to the attention layers. \n\nAnother architecture that can give nice results without much tuning is airnet, but it’s very slow. Good luck!",
    "494245": "Thank you for the reply, I'll try experimenting with those architectures and see how I get on. Also why should I use the full 96x96 image when it's only dependent on the central 32 pixels? Is it to do with zero padding?",
    "494290": "Its true the organizers claim the label is determined exclusively by the central 32px, but I think the way they phrased it in the description is not doing them or anyone else any favors. They say a single pixel of metastasis in the center will make the label positive, so are we trying to classify single pixels? \n\nThough the label may be determined by the center this does not imply there is no useful information in the rest of the image so discarding 89% of the pixels likely throws out too much.",
    "494364": "I tried center cropping based on the same ideas and it definitely works worse. Resizing to 196x196 or 244x244 works much much better. You should try that. \n\nP.S. I tested this with resnet50 and se_resnet50. In both case resizing to 196x196 or 244x244 worked tremendously better.",
    "494638": "HI, Jake. I have tried central crop 64 or 48, but it seems not as good as using the full size image(96x96). And resizing it to 224x224 may better if u use an imagenet pretrained models. And you can try ensemble,  which is a very good way to improve your score ([see this topic](https://www.kaggle.com/c/histopathologic-cancer-detection/discussion/83313))."
  },
  "source": "meta"
}