Congrats to all winners and everyone who finished this competition! It has been an amazing learning experience for me as I am still super new to image classification problems. Special thanks to @drhabib where I got the inspirition from for my approach and @ratthachat for his amazing preprocessing kernels.
\n\nSo here is my approach, which is quite simple:
\n\nThe model is EfficientNet B4 in Pytorch with Adam with image size of 256.
\n\nPreprocessing: \nCrop all images using the crop_image_from_gray function. \nFor 2019 validation data only, random center crop (1.1 to 1.3) is applied
Augmentation:\n \nRandomResizedCrop((256,256), scale=(0.5, 0.9)),\nRandomHorizontalFlip(),\nRandomRotation(360),\nColorJitter(contrast=(0.75,2.2))\n
Steps:\n1. Pretrain on previous competition data (2015) and validated by current training data (2019). \n a. Freeze all layers except the last, train for 5 epochs\n b. Unfreeze all layers and train for 15 epochs
\n\nOther things I tried but did not work for me:\n- Bigger image size and bigger EfficientNet models\n- Ben's preprocessing\n- CLAHE\n- Circle cropping
\n\nThanks to Kaggle, APTOS and everyone who contributed via Kernels and Discussions. I learned so much from you guys.
\n\nCode: https://github.com/qiletan/aptos2019-blindness-detection
", "messages": [ { "id": 621103, "postDate": "2019-09-08T07:49:35.137Z", "content": "Congrats to all winners and everyone who finished this competition! It has been an amazing learning experience for me as I am still super new to image classification problems. Special thanks to @drhabib where I got the inspirition from for my approach and @ratthachat for his amazing preprocessing kernels.
\n\nSo here is my approach, which is quite simple:
\n\nThe model is EfficientNet B4 in Pytorch with Adam with image size of 256.
\n\nPreprocessing: \nCrop all images using the crop_image_from_gray function. \nFor 2019 validation data only, random center crop (1.1 to 1.3) is applied
Augmentation:\n \nRandomResizedCrop((256,256), scale=(0.5, 0.9)),\nRandomHorizontalFlip(),\nRandomRotation(360),\nColorJitter(contrast=(0.75,2.2))\n
Steps:\n1. Pretrain on previous competition data (2015) and validated by current training data (2019). \n a. Freeze all layers except the last, train for 5 epochs\n b. Unfreeze all layers and train for 15 epochs
\n\nOther things I tried but did not work for me:\n- Bigger image size and bigger EfficientNet models\n- Ben's preprocessing\n- CLAHE\n- Circle cropping
\n\nThanks to Kaggle, APTOS and everyone who contributed via Kernels and Discussions. I learned so much from you guys.
\n\nCode: https://github.com/qiletan/aptos2019-blindness-detection
", "rawMarkdown": "Congrats to all winners and everyone who finished this competition! It has been an amazing learning experience for me as I am still super new to image classification problems. Special thanks to @drhabib where I got the inspirition from for my approach and @ratthachat for his amazing preprocessing kernels.\n\n\nSo here is my approach, which is quite simple:\n\nThe model is EfficientNet B4 in Pytorch with Adam with image size of 256.\n\n**Preprocessing:** \nCrop all images using the `crop_image_from_gray` function. \nFor 2019 validation data only, random center crop (1.1 to 1.3) is applied\n\n**Augmentation:**\n ```\nRandomResizedCrop((256,256), scale=(0.5, 0.9)),\nRandomHorizontalFlip(),\nRandomRotation(360),\nColorJitter(contrast=(0.75,2.2))\n```\n\n\n**Steps:**\n1. Pretrain on previous competition data (2015) and validated by current training data (2019). \n a. Freeze all layers except the last, train for 5 epochs\n b. Unfreeze all layers and train for 15 epochs\n\n2. Finetune using 2019 data. CV-5 is used and train for 7 epochs for each fold\n3. Optimised thresholds (credit goes to @abhishek)\n3. For inference, no TTA, and use simple average of the 5 CV models\n\n**Other things I tried but did not work for me:**\n- Bigger image size and bigger EfficientNet models\n- Ben's preprocessing\n- CLAHE\n- Circle cropping\n\n\n\nThanks to Kaggle, APTOS and everyone who contributed via Kernels and Discussions. I learned so much from you guys.\n\nCode: https://github.com/qiletan/aptos2019-blindness-detection\n\n\n\n\n\n", "votes": 6 }, { "id": 621949, "postDate": "2019-09-09T05:45:51.977Z", "rawMarkdown": "", "isDeleted": true }, { "id": 621410, "postDate": "2019-09-08T13:43:53.490Z", "content": "Thanks and see you next time @qiletan !
", "rawMarkdown": "Thanks and see you next time @qiletan !" }, { "id": 621398, "postDate": "2019-09-08T13:34:03.010Z", "content": "Congrats and thanks for sharing :)
", "rawMarkdown": "Congrats and thanks for sharing :)" } ], "comments": [ { "id": 621949, "author_name": "", "author_url": "", "post_date": "2019-09-09T05:45:51.977000", "content": "", "votes": 0, "replies": [] }, { "id": 621410, "author_name": "Neuron Engineer", "author_url": "", "post_date": "2019-09-08T13:43:53.490000", "content": "Thanks and see you next time @qiletan !
", "votes": 0, "replies": [] }, { "id": 621398, "author_name": "JM100", "author_url": "", "post_date": "2019-09-08T13:34:03.010000", "content": "Congrats and thanks for sharing :)
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "621103": "Congrats to all winners and everyone who finished this competition! It has been an amazing learning experience for me as I am still super new to image classification problems. Special thanks to @drhabib where I got the inspirition from for my approach and @ratthachat for his amazing preprocessing kernels.\n\n\nSo here is my approach, which is quite simple:\n\nThe model is EfficientNet B4 in Pytorch with Adam with image size of 256.\n\n**Preprocessing:** \nCrop all images using the `crop_image_from_gray` function. \nFor 2019 validation data only, random center crop (1.1 to 1.3) is applied\n\n**Augmentation:**\n ```\nRandomResizedCrop((256,256), scale=(0.5, 0.9)),\nRandomHorizontalFlip(),\nRandomRotation(360),\nColorJitter(contrast=(0.75,2.2))\n```\n\n\n**Steps:**\n1. Pretrain on previous competition data (2015) and validated by current training data (2019). \n a. Freeze all layers except the last, train for 5 epochs\n b. Unfreeze all layers and train for 15 epochs\n\n2. Finetune using 2019 data. CV-5 is used and train for 7 epochs for each fold\n3. Optimised thresholds (credit goes to @abhishek)\n3. For inference, no TTA, and use simple average of the 5 CV models\n\n**Other things I tried but did not work for me:**\n- Bigger image size and bigger EfficientNet models\n- Ben's preprocessing\n- CLAHE\n- Circle cropping\n\n\n\nThanks to Kaggle, APTOS and everyone who contributed via Kernels and Discussions. I learned so much from you guys.\n\nCode: https://github.com/qiletan/aptos2019-blindness-detection\n\n\n\n\n\n", "621949": "", "621410": "Thanks and see you next time @qiletan !", "621398": "Congrats and thanks for sharing :)" } }