Thanks to Kaggle and Mayo Clinic, for this competition. Congratulations to all winners!
\nI expected a shakeup given 93% of test set was hidden but not a massive shakeup like we see. I had jumped 500 places to rank 52, winning my first medal (bronze), in a competition .
\nMy approach relied a lot on augmentations. Further details are below:
\nAugmentations:
\nTrain set:
Val/Test sets:
\nTraining:
\nPytorch EfficientNetB5 model
\nCrossEntropyLoss
\nAdam Optimizer. + CosineAnnealingLR
\n5 fold Stratified Split
\n10 Epochs per fold. For every fold, save the model based on the lowest loss
Inference:
\nFor each image, average the prediction from the 5 models obtained during the training phase)
Thanks to Kaggle and Mayo Clinic, for this competition. Congratulations to all winners!
\nI expected a shakeup given 93% of test set was hidden but not a massive shakeup like we see. I had jumped 500 places to rank 52, winning my first medal (bronze), in a competition .
\nMy approach relied a lot on augmentations. Further details are below:
\nAugmentations:
\nTrain set:
Val/Test sets:
\nTraining:
\nPytorch EfficientNetB5 model
\nCrossEntropyLoss
\nAdam Optimizer. + CosineAnnealingLR
\n5 fold Stratified Split
\n10 Epochs per fold. For every fold, save the model based on the lowest loss
Inference:
\nFor each image, average the prediction from the 5 models obtained during the training phase)
@karthikrg congrats on you results and you seemed to have made a simple approach work very well. Could you elaborate a bit what you mean by normalizing dataset using imagenet weights in augmentations?
", "rawMarkdown": "@karthikrg congrats on you results and you seemed to have made a simple approach work very well. Could you elaborate a bit what you mean by normalizing dataset using imagenet weights in augmentations? ", "replies": [ { "id": 1979260, "postDate": "2022-10-09T09:03:51.200Z", "content": "I meant to say Imagenet Stats: ([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]).
", "rawMarkdown": "I meant to say Imagenet Stats: ([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]).\n" }, { "id": 1979667, "postDate": "2022-10-09T15:51:03.957Z", "content": "@karthikrg I'm sorry but I still do not understand… are these predictions from imagenet weight model that you used in some fashion? Thanks for your patience with my slowness :)
", "rawMarkdown": "@karthikrg I'm sorry but I still do not understand... are these predictions from imagenet weight model that you used in some fashion? Thanks for your patience with my slowness :)" }, { "id": 1980698, "postDate": "2022-10-10T10:25:28.827Z", "content": "I used these stats to normalize images and see how the results are. I have updated my original post.
\nThere are opinions to the contrary, like discussion here.
", "rawMarkdown": "I used these stats to normalize images and see how the results are. I have updated my original post.\n\nThere are opinions to the contrary, like discussion [here](https://discuss.pytorch.org/t/discussion-why-normalise-according-to-imagenet-mean-and-std-dev-for-transfer-learning/115670/5).\n", "votes": 1 }, { "id": 1981070, "postDate": "2022-10-10T15:08:40.610Z", "content": "@karthikrg thanks, I get it now.
", "rawMarkdown": "@karthikrg thanks, I get it now." } ] } ], "comments": [ { "id": 1978629, "author_name": "tdiceman", "author_url": "", "post_date": "2022-10-08T20:33:20.993000", "content": "@karthikrg congrats on you results and you seemed to have made a simple approach work very well. Could you elaborate a bit what you mean by normalizing dataset using imagenet weights in augmentations?
", "votes": 0, "replies": [ { "id": 1979260, "author_name": "Karthik", "author_url": "", "post_date": "2022-10-09T09:03:51.200000", "content": "I meant to say Imagenet Stats: ([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]).
", "votes": 0, "replies": [] }, { "id": 1979667, "author_name": "tdiceman", "author_url": "", "post_date": "2022-10-09T15:51:03.957000", "content": "@karthikrg I'm sorry but I still do not understand… are these predictions from imagenet weight model that you used in some fashion? Thanks for your patience with my slowness :)
", "votes": 0, "replies": [] }, { "id": 1980698, "author_name": "Karthik", "author_url": "", "post_date": "2022-10-10T10:25:28.827000", "content": "I used these stats to normalize images and see how the results are. I have updated my original post.
\nThere are opinions to the contrary, like discussion here.
", "votes": 1, "replies": [] }, { "id": 1981070, "author_name": "tdiceman", "author_url": "", "post_date": "2022-10-10T15:08:40.610000", "content": "@karthikrg thanks, I get it now.
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "1977615": "Thanks to Kaggle and Mayo Clinic, for this competition. Congratulations to all winners!\n\nI expected a shakeup given 93% of test set was hidden but not a massive shakeup like we see. I had jumped 500 places to rank 52, winning my first medal (bronze), in a competition . \n\nMy approach relied a lot on augmentations. Further details are below:\n\n**Augmentations**:\nTrain set: \n- 512 resized RGB image\n- RandomHorizontalFlip\n- RandomRotation\n- FiveCrop\n- Normalization using ImageNet ~~weights ~~ stats\n\nVal/Test sets:\n- 512 resized RGB image\n- Normalization using ImageNet ~~weights ~~stats\n\n**Training**:\nPytorch EfficientNetB5 model\nCrossEntropyLoss\nAdam Optimizer. + CosineAnnealingLR\n5 fold Stratified Split\n10 Epochs per fold. For every fold, save the model based on the lowest loss\n\n**Inference**:\nFor each image, average the prediction from the 5 models obtained during the training phase) ", "1978629": "@karthikrg congrats on you results and you seemed to have made a simple approach work very well. Could you elaborate a bit what you mean by normalizing dataset using imagenet weights in augmentations? " } }