Hello all!
\n\nThere was a previous competition with ~35,000 fundus images with the same classification as for this competition (5 levels). Since it is such a large dataset, I have resized it and included in a Kaggle Dataset for use in Kaggle Kernels.
\n\nHere is the link!\nhttps://www.kaggle.com/tanlikesmath/diabetic-retinopathy-resized
\n\nThis includes two versions. One where I just resize to size=1024, and one where I crop as much black space as possible and then resize. The aspect ratios were maintained.
\n\nI also have a simple kernel for training a simple model on BINARY class predictions. I already have a model on multi-class but I have yet to clean up that kernel and also try it on this task. Be on the lookout for that in the next couple days!
\n\nHope this helps!
", "messages": [ { "id": 565493, "postDate": "2019-07-01T02:56:35.987Z", "content": "Hello all!
\n\nThere was a previous competition with ~35,000 fundus images with the same classification as for this competition (5 levels). Since it is such a large dataset, I have resized it and included in a Kaggle Dataset for use in Kaggle Kernels.
\n\nHere is the link!\nhttps://www.kaggle.com/tanlikesmath/diabetic-retinopathy-resized
\n\nThis includes two versions. One where I just resize to size=1024, and one where I crop as much black space as possible and then resize. The aspect ratios were maintained.
\n\nI also have a simple kernel for training a simple model on BINARY class predictions. I already have a model on multi-class but I have yet to clean up that kernel and also try it on this task. Be on the lookout for that in the next couple days!
\n\nHope this helps!
", "rawMarkdown": "Hello all!\n\nThere was a [previous competition](https://www.kaggle.com/c/diabetic-retinopathy-detection) with ~35,000 fundus images with the same classification as for this competition (5 levels). Since it is such a large dataset, I have resized it and included in a Kaggle Dataset for use in Kaggle Kernels.\n\nHere is the link!\nhttps://www.kaggle.com/tanlikesmath/diabetic-retinopathy-resized\n\nThis includes two versions. One where I just resize to size=1024, and one where I crop as much black space as possible and then resize. The aspect ratios were maintained. \n\nI also have a simple [kernel](https://www.kaggle.com/tanlikesmath/diabetic-retinopathy-resnet50-binary-cropped) for training a simple model on **BINARY** class predictions. I already have a model on multi-class but I have yet to clean up that kernel and also try it on this task. Be on the lookout for that in the next couple days!\n\nHope this helps! ", "votes": 81 }, { "id": 567784, "postDate": "2019-07-04T02:25:45.430Z", "content": "As promised, here is a starter kernel for training on the previous dataset and transfer learning for this task. It is probably not helpful in its current form due to the time limitations, but if you separate the two training steps into two kernels you might be able to achieve better results.
", "rawMarkdown": "As promised, here is a [starter kernel](https://www.kaggle.com/tanlikesmath/training-on-previous-dataset-for-aptos) for training on the previous dataset and transfer learning for this task. It is probably not helpful in its current form due to the time limitations, but if you separate the two training steps into two kernels you might be able to achieve better results.", "votes": 6 }, { "id": 568098, "postDate": "2019-07-04T11:51:05.800Z", "content": "Do you have script that you used for resizing ?
", "rawMarkdown": "Do you have script that you used for resizing ? ", "votes": 1, "replies": [ { "id": 568410, "postDate": "2019-07-04T21:57:07.970Z", "content": "I also would be interested in that :)
", "rawMarkdown": "I also would be interested in that :) " }, { "id": 568749, "postDate": "2019-07-05T11:32:34.350Z", "rawMarkdown": "", "isDeleted": true }, { "id": 570872, "postDate": "2019-07-08T21:32:13.373Z", "content": "I have fixed it. Sorry about that.
", "rawMarkdown": "I have fixed it. Sorry about that." } ] }, { "id": 568770, "postDate": "2019-07-05T12:32:13.400Z", "content": "Thanks for sharing and taking the time to do it, that competition data seems to have lower quality, but probably can help as a first stage pre-training.
", "rawMarkdown": "Thanks for sharing and taking the time to do it, that competition data seems to have lower quality, but probably can help as a first stage pre-training." }, { "id": 566637, "postDate": "2019-07-02T12:05:22.340Z", "content": "Thanks! That is what i was looking for! \nnow I have to figure out how to add these data to competition:))
", "rawMarkdown": "Thanks! That is what i was looking for! \nnow I have to figure out how to add these data to competition:))" }, { "id": 565884, "postDate": "2019-07-01T13:45:54.173Z", "content": "wow great!!
", "rawMarkdown": "wow great!!" }, { "id": 588049, "postDate": "2019-07-30T04:26:45.027Z", "rawMarkdown": "", "isDeleted": true }, { "id": 568053, "postDate": "2019-07-04T10:39:58.340Z", "content": "thanks
", "rawMarkdown": "thanks", "votes": 2 }, { "id": 616865, "postDate": "2019-09-03T14:41:53.530Z", "content": "Thanks for sharing.
", "rawMarkdown": "Thanks for sharing." }, { "id": 565838, "postDate": "2019-07-01T12:43:04.917Z", "content": "Thanks!
", "rawMarkdown": "Thanks!" } ], "comments": [ { "id": 567784, "author_name": "ilovescience", "author_url": "", "post_date": "2019-07-04T02:25:45.430000", "content": "As promised, here is a starter kernel for training on the previous dataset and transfer learning for this task. It is probably not helpful in its current form due to the time limitations, but if you separate the two training steps into two kernels you might be able to achieve better results.
", "votes": 6, "replies": [] }, { "id": 568098, "author_name": "DrHB", "author_url": "", "post_date": "2019-07-04T11:51:05.800000", "content": "Do you have script that you used for resizing ?
", "votes": 1, "replies": [ { "id": 568410, "author_name": "Federico Raimondi Cominesi", "author_url": "", "post_date": "2019-07-04T21:57:07.970000", "content": "I also would be interested in that :)
", "votes": 0, "replies": [] }, { "id": 568749, "author_name": "", "author_url": "", "post_date": "2019-07-05T11:32:34.350000", "content": "", "votes": 0, "replies": [] }, { "id": 570872, "author_name": "ilovescience", "author_url": "", "post_date": "2019-07-08T21:32:13.373000", "content": "I have fixed it. Sorry about that.
", "votes": 0, "replies": [] } ] }, { "id": 568770, "author_name": "DimitreOliveira", "author_url": "", "post_date": "2019-07-05T12:32:13.400000", "content": "Thanks for sharing and taking the time to do it, that competition data seems to have lower quality, but probably can help as a first stage pre-training.
", "votes": 0, "replies": [] }, { "id": 566637, "author_name": "Serge", "author_url": "", "post_date": "2019-07-02T12:05:22.340000", "content": "Thanks! That is what i was looking for! \nnow I have to figure out how to add these data to competition:))
", "votes": 0, "replies": [] }, { "id": 565884, "author_name": "DrHB", "author_url": "", "post_date": "2019-07-01T13:45:54.173000", "content": "wow great!!
", "votes": 0, "replies": [] }, { "id": 588049, "author_name": "", "author_url": "", "post_date": "2019-07-30T04:26:45.027000", "content": "", "votes": 0, "replies": [] }, { "id": 568053, "author_name": "Krishna Katyal", "author_url": "", "post_date": "2019-07-04T10:39:58.340000", "content": "thanks
", "votes": 2, "replies": [] }, { "id": 616865, "author_name": "Rashmiranjan Pradhan", "author_url": "", "post_date": "2019-09-03T14:41:53.530000", "content": "Thanks for sharing.
", "votes": 0, "replies": [] }, { "id": 565838, "author_name": "Bohdan Safoniuk", "author_url": "", "post_date": "2019-07-01T12:43:04.917000", "content": "Thanks!
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "565493": "Hello all!\n\nThere was a [previous competition](https://www.kaggle.com/c/diabetic-retinopathy-detection) with ~35,000 fundus images with the same classification as for this competition (5 levels). Since it is such a large dataset, I have resized it and included in a Kaggle Dataset for use in Kaggle Kernels.\n\nHere is the link!\nhttps://www.kaggle.com/tanlikesmath/diabetic-retinopathy-resized\n\nThis includes two versions. One where I just resize to size=1024, and one where I crop as much black space as possible and then resize. The aspect ratios were maintained. \n\nI also have a simple [kernel](https://www.kaggle.com/tanlikesmath/diabetic-retinopathy-resnet50-binary-cropped) for training a simple model on **BINARY** class predictions. I already have a model on multi-class but I have yet to clean up that kernel and also try it on this task. Be on the lookout for that in the next couple days!\n\nHope this helps! ", "567784": "As promised, here is a [starter kernel](https://www.kaggle.com/tanlikesmath/training-on-previous-dataset-for-aptos) for training on the previous dataset and transfer learning for this task. It is probably not helpful in its current form due to the time limitations, but if you separate the two training steps into two kernels you might be able to achieve better results.", "568098": "Do you have script that you used for resizing ? ", "568770": "Thanks for sharing and taking the time to do it, that competition data seems to have lower quality, but probably can help as a first stage pre-training.", "566637": "Thanks! That is what i was looking for! \nnow I have to figure out how to add these data to competition:))", "565884": "wow great!!", "588049": "", "568053": "thanks", "616865": "Thanks for sharing.", "565838": "Thanks!" } }