{ "id": 107954, "title": "Congratulations to all ! 54th Solution.", "url": "/competitions/aptos2019-blindness-detection/discussion/107954", "author_name": "Chanhu", "post_date": "2019-09-08T06:04:22.313000", "votes": 12, "comment_count": 4, "views": 0, "content": "

Congratulations to all !\nThanks to my great teammates. (friends in real world)

\n\n
    \n
  1. Network : EfficientNetB4& EfficientNetB5 (Both Classification and regression.)
    2. \n
  2. Image size: B4: (256x256), B5: (328x328)
    3. \n
  3. Image Preprocessing: Cropping images Only.\n(Code from https://www.kaggle.com/benjaminwarner/starter-code-resized-15-19-blindness-images)
    4. \n
  4. Data Augment:\ntrain_transform = transforms.Compose([\n transforms.ColorJitter(brightness=0.45, contrast=0.45),\n transforms.RandomAffine(degrees=360, scale=(1.0, 1.3)),\n transforms.RandomHorizontalFlip(),\n transforms.RandomVerticalFlip(),\n transforms.ToTensor(),\n transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])
    5. \n
  5. Optimizer : Adam
    6. \n
  6. Loss : Classification: Cross entropy Loss\n Regression:MSE
    7. \n
  7. Learning Rate: Start from 0.001
    8. \n
  8. Learning Rate Scheduler：StepLR(B4-> StepSize:5, B5-> StepSize:10), gamma=0.1
    9. \n

  9. TTA: TTA 10 times by following code\ntest_transform = transforms.Compose([\ntransforms.RandomAffine(degrees=360),\ntransforms.RandomHorizontalFlip(),\ntransforms.RandomVerticalFlip(),\ntransforms.ToTensor(),\ntransforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])

    10. \n

  10. Score:\n -> Public LB:
    \n B4 Classification: 0.821
    \n B5 Classification: 0.825\n B4 Regression:0.823\n B5 Regression:0.825\n -> Private LB\n B4 Classification: 0.919
    \n B5 Classification: 0.921\n B4 Regression: 0.917\n B5 Regression: 0.921

    11. \n

  11. Final Submission:\nCombining B4 & B5 regression output together. （Public LB:0.835, Private 0.925)

    12. \n
", "messages": [ { "id": 620982, "postDate": "2019-09-08T06:04:22.313Z", "content": "

Congratulations to all !\nThanks to my great teammates. (friends in real world)

\n\n
    \n
  1. Network : EfficientNetB4& EfficientNetB5 (Both Classification and regression.)
    2. \n
  2. Image size: B4: (256x256), B5: (328x328)
    3. \n
  3. Image Preprocessing: Cropping images Only.\n(Code from https://www.kaggle.com/benjaminwarner/starter-code-resized-15-19-blindness-images)
    4. \n
  4. Data Augment:\ntrain_transform = transforms.Compose([\n transforms.ColorJitter(brightness=0.45, contrast=0.45),\n transforms.RandomAffine(degrees=360, scale=(1.0, 1.3)),\n transforms.RandomHorizontalFlip(),\n transforms.RandomVerticalFlip(),\n transforms.ToTensor(),\n transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])
    5. \n
  5. Optimizer : Adam
    6. \n
  6. Loss : Classification: Cross entropy Loss\n Regression:MSE
    7. \n
  7. Learning Rate: Start from 0.001
    8. \n
  8. Learning Rate Scheduler：StepLR(B4-> StepSize:5, B5-> StepSize:10), gamma=0.1
    9. \n

  9. TTA: TTA 10 times by following code\ntest_transform = transforms.Compose([\ntransforms.RandomAffine(degrees=360),\ntransforms.RandomHorizontalFlip(),\ntransforms.RandomVerticalFlip(),\ntransforms.ToTensor(),\ntransforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])

    10. \n

  10. Score:\n -> Public LB:
    \n B4 Classification: 0.821
    \n B5 Classification: 0.825\n B4 Regression:0.823\n B5 Regression:0.825\n -> Private LB\n B4 Classification: 0.919
    \n B5 Classification: 0.921\n B4 Regression: 0.917\n B5 Regression: 0.921

    11. \n

  11. Final Submission:\nCombining B4 & B5 regression output together. （Public LB:0.835, Private 0.925)

    12. \n
", "rawMarkdown": "Congratulations to all !\nThanks to my great teammates. (friends in real world)\n\n1. Network : EfficientNetB4& EfficientNetB5 (Both Classification and regression.)\n2. Image size: B4: (256x256), B5: (328x328)\n3. Image Preprocessing: Cropping images Only.\n(Code from https://www.kaggle.com/benjaminwarner/starter-code-resized-15-19-blindness-images) \n4. Data Augment:\n`train_transform = transforms.Compose([\n transforms.ColorJitter(brightness=0.45, contrast=0.45),\n transforms.RandomAffine(degrees=360, scale=(1.0, 1.3)),\n transforms.RandomHorizontalFlip(),\n transforms.RandomVerticalFlip(),\n transforms.ToTensor(),\n transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])`\n5. Optimizer : Adam\n6. Loss : Classification: Cross entropy Loss\n Regression:MSE\n7. Learning Rate: Start from 0.001\n8. Learning Rate Scheduler：StepLR(B4-> StepSize:5, B5-> StepSize:10), gamma=0.1\n9. TTA: TTA 10 times by following code\n `test_transform = transforms.Compose([\n transforms.RandomAffine(degrees=360),\n transforms.RandomHorizontalFlip(),\n transforms.RandomVerticalFlip(),\n transforms.ToTensor(),\n transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])`\n \n10. Score:\n -> Public LB: \n B4 Classification: 0.821 \n B5 Classification: 0.825\n B4 Regression:0.823\n B5 Regression:0.825\n -> Private LB\n B4 Classification: 0.919 \n B5 Classification: 0.921\n B4 Regression: 0.917\n B5 Regression: 0.921\n \n11. Final Submission:\nCombining B4 & B5 regression output together. （Public LB:0.835, Private 0.925)", "votes": 12 }, { "id": 621089, "postDate": "2019-09-08T07:34:38.950Z", "content": "

Congrats @chanhu and your teammates!\nCould I ask how’s the strategy of validation? I couldn’t make reliable validation set, and I believe this was one of discussion topics in this comp.

", "rawMarkdown": "Congrats @chanhu and your teammates!\nCould I ask how’s the strategy of validation? I couldn’t make reliable validation set, and I believe this was one of discussion topics in this comp.", "replies": [ { "id": 621233, "postDate": "2019-09-08T11:09:48.703Z", "content": "
    \n
  1. Combining 2019 and 2015 Data Together.
    2. \n
  2. Removing duplicated Image in 2019 Public Train Dataset.\n There are two type of duplicated Image.\n
    3. \n
  3. Data augmentation in validation.\n We found that Data augmentation in validation made model more stable.\n (the idea is based on Fixing the train-test resolution discrepancy)
    4. \n
", "rawMarkdown": "1. Combining 2019 and 2015 Data Together.\n2. Removing duplicated Image in 2019 Public Train Dataset.\n There are two type of duplicated Image.\n - Same Image with different Label.\n (keeping them, believe that there is same type of noise in private & public test set)\n - Same Image with same Label. (only keep 1 image before train_test_split, \n don‘t want same pictures which one is in train set, which one is in test set)\n3. Data augmentation in validation.\n We found that Data augmentation in validation made model more stable.\n (the idea is based on [Fixing the train-test resolution discrepancy](https://github.com/facebookresearch/FixRes))", "votes": 1 } ] }, { "id": 621955, "postDate": "2019-09-09T05:50:04.067Z", "rawMarkdown": "", "isDeleted": true }, { "id": 621049, "postDate": "2019-09-08T07:02:22.323Z", "content": "

Congrats and thank you for sharing.

", "rawMarkdown": "Congrats and thank you for sharing. ", "votes": 1 } ], "comments": [ { "id": 621089, "author_name": "Y. O.", "author_url": "", "post_date": "2019-09-08T07:34:38.950000", "content": "

Congrats @chanhu and your teammates!\nCould I ask how’s the strategy of validation? I couldn’t make reliable validation set, and I believe this was one of discussion topics in this comp.

", "votes": 0, "replies": [ { "id": 621233, "author_name": "Chanhu", "author_url": "", "post_date": "2019-09-08T11:09:48.703000", "content": "
    \n
  1. Combining 2019 and 2015 Data Together.
    2. \n
  2. Removing duplicated Image in 2019 Public Train Dataset.\n There are two type of duplicated Image.\n
    3. \n
  3. Data augmentation in validation.\n We found that Data augmentation in validation made model more stable.\n (the idea is based on Fixing the train-test resolution discrepancy)
    4. \n
", "votes": 1, "replies": [] } ] }, { "id": 621955, "author_name": "", "author_url": "", "post_date": "2019-09-09T05:50:04.067000", "content": "", "votes": 0, "replies": [] }, { "id": 621049, "author_name": "Anand Selvadurai", "author_url": "", "post_date": "2019-09-08T07:02:22.323000", "content": "

Congrats and thank you for sharing.

", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "620982": "Congratulations to all !\nThanks to my great teammates. (friends in real world)\n\n1. Network : EfficientNetB4& EfficientNetB5 (Both Classification and regression.)\n2. Image size: B4: (256x256), B5: (328x328)\n3. Image Preprocessing: Cropping images Only.\n(Code from https://www.kaggle.com/benjaminwarner/starter-code-resized-15-19-blindness-images) \n4. Data Augment:\n`train_transform = transforms.Compose([\n transforms.ColorJitter(brightness=0.45, contrast=0.45),\n transforms.RandomAffine(degrees=360, scale=(1.0, 1.3)),\n transforms.RandomHorizontalFlip(),\n transforms.RandomVerticalFlip(),\n transforms.ToTensor(),\n transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])`\n5. Optimizer : Adam\n6. Loss : Classification: Cross entropy Loss\n Regression:MSE\n7. Learning Rate: Start from 0.001\n8. Learning Rate Scheduler：StepLR(B4-> StepSize:5, B5-> StepSize:10), gamma=0.1\n9. TTA: TTA 10 times by following code\n `test_transform = transforms.Compose([\n transforms.RandomAffine(degrees=360),\n transforms.RandomHorizontalFlip(),\n transforms.RandomVerticalFlip(),\n transforms.ToTensor(),\n transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])`\n \n10. Score:\n -> Public LB: \n B4 Classification: 0.821 \n B5 Classification: 0.825\n B4 Regression:0.823\n B5 Regression:0.825\n -> Private LB\n B4 Classification: 0.919 \n B5 Classification: 0.921\n B4 Regression: 0.917\n B5 Regression: 0.921\n \n11. Final Submission:\nCombining B4 & B5 regression output together. （Public LB:0.835, Private 0.925)", "621089": "Congrats @chanhu and your teammates!\nCould I ask how’s the strategy of validation? I couldn’t make reliable validation set, and I believe this was one of discussion topics in this comp.", "621955": "", "621049": "Congrats and thank you for sharing. " } }