![\"Summary\"](\"https://i.ibb.co/VCns6cw/submission.png\")
Hi all! It was a very long and intensive 3 months)
\nFirst of all, many thanks to the organizers for this very challenging and interesting competition!
This was my first \"real\" competition at kaggle, so I'm glad to get a medal :-)
\nMy solution consists of several parts:
\n1) Cell segmentation: I trained MaskRCNN on the HPASegmentator predictions. The only reason for this was the inference speed. MaskRCNN works really faster. A few experiments showed that solution score stayed the same when swapping HPASegmentator with MaskRCNN.
\n2) Image Level classifiers:
\n3) Cell level classifiers:
\n I used PuzzleCAM paper approach for creating CAMs. And then I used them to create pseudo labels (at this point the labels were smooth - floats in [0.0, 1.0]) for each cell within images.
\n Next step was to manually set up thresholds for each class and get final pseudo labels for further training model.
4) Final submission I combined predictions from all 4 networks and get final predictions for each cell. The best result was with just an average of predictions from 4 models.
\n5) Training details
\nAll classifiers were trained with focal_loss + lovazh_loss (like in the winning solution of @bestfitting in previous HPA competition).
\nOptimizer: Adam
\nScheduler: ReduceOnPlatoue
\nI also used oversampling for rare classes and undersampling for too common classes.
Here is a summary:
\n
Thanks for reading and good luck in future competitions!)
", "messages": [ { "id": 1303667, "postDate": "2021-05-12T07:28:37.477Z", "content": "Hi all! It was a very long and intensive 3 months)
\nFirst of all, many thanks to the organizers for this very challenging and interesting competition!
This was my first \"real\" competition at kaggle, so I'm glad to get a medal :-)
\nMy solution consists of several parts:
\n1) Cell segmentation: I trained MaskRCNN on the HPASegmentator predictions. The only reason for this was the inference speed. MaskRCNN works really faster. A few experiments showed that solution score stayed the same when swapping HPASegmentator with MaskRCNN.
\n2) Image Level classifiers:
\n3) Cell level classifiers:
\n I used PuzzleCAM paper approach for creating CAMs. And then I used them to create pseudo labels (at this point the labels were smooth - floats in [0.0, 1.0]) for each cell within images.
\n Next step was to manually set up thresholds for each class and get final pseudo labels for further training model.
4) Final submission I combined predictions from all 4 networks and get final predictions for each cell. The best result was with just an average of predictions from 4 models.
\n5) Training details
\nAll classifiers were trained with focal_loss + lovazh_loss (like in the winning solution of @bestfitting in previous HPA competition).
\nOptimizer: Adam
\nScheduler: ReduceOnPlatoue
\nI also used oversampling for rare classes and undersampling for too common classes.
Here is a summary:
\n
Thanks for reading and good luck in future competitions!)
", "rawMarkdown": "Hi all! It was a very long and intensive 3 months)\nFirst of all, many thanks to the organizers for this very challenging and interesting competition!\n\nThis was my first \"real\" competition at kaggle, so I'm glad to get a medal :-)\n\nMy solution consists of several parts:\n\n1) **Cell segmentation**: I trained MaskRCNN on the HPASegmentator predictions. The only reason for this was the inference speed. MaskRCNN works really faster. A few experiments showed that solution score stayed the same when swapping HPASegmentator with MaskRCNN.\n\n2) **Image Level classifiers**:\n\n- EfficientNetB3 (1024x1024)\n- EfficientNetB3 (only green channel) (1024x1024) (thanks @h053473666 for this idea)\n\n3) **Cell level classifiers**:\n I used [PuzzleCAM paper](https://arxiv.org/abs/2101.11253) approach for creating CAMs. And then I used them to create pseudo labels (at this point the labels were smooth - floats in [0.0, 1.0]) for each cell within images.\n Next step was to manually set up thresholds for each class and get final pseudo labels for further training model.\n\n- EfficientNetB3 with dropout (256x256)\n- EfficientNetB3 with dropout (only green channel) (256x256)\n\n\n4) **Final submission** I combined predictions from all 4 networks and get final predictions for each cell. The best result was with just an average of predictions from 4 models.\n\n5) **Training details**\nAll classifiers were trained with focal_loss + lovazh_loss (like in the winning solution of @bestfitting in previous HPA competition).\nOptimizer: Adam\nScheduler: ReduceOnPlatoue\nI also used oversampling for rare classes and undersampling for too common classes.\n\nHere is a summary:\n\n\n\nThanks for reading and good luck in future competitions!)\n", "votes": 17 }, { "id": 1305574, "postDate": "2021-05-13T11:23:14.813Z", "content": "gratz with good performance outside MADE 💪 :D
", "rawMarkdown": "gratz with good performance outside MADE 💪 :D", "votes": 1, "replies": [ { "id": 1305750, "postDate": "2021-05-13T13:09:12.580Z", "content": "Thanks, as I see you also done great in Shopee competition) Congratulations with Expert tire)
", "rawMarkdown": "Thanks, as I see you also done great in Shopee competition) Congratulations with Expert tire)", "votes": 1 }, { "id": 1305773, "postDate": "2021-05-13T13:22:39.120Z", "content": "Well expert was even before MADE but thx 😄
", "rawMarkdown": "Well expert was even before MADE but thx 😄" } ] }, { "id": 1305031, "postDate": "2021-05-13T04:27:41.470Z", "content": "@mgurevich Congratulations and Thanks for sharing the approach
", "rawMarkdown": "@mgurevich Congratulations and Thanks for sharing the approach", "votes": 1 }, { "id": 1304399, "postDate": "2021-05-12T15:47:56.810Z", "content": "Congratulations on Silver medal!
", "rawMarkdown": "Congratulations on Silver medal!", "votes": 1, "replies": [ { "id": 1304733, "postDate": "2021-05-12T20:21:59.127Z", "content": "Thanks and congratulations to you too!)
", "rawMarkdown": "Thanks and congratulations to you too!)" } ] }, { "id": 1304143, "postDate": "2021-05-12T12:56:10.397Z", "content": "Thanks for sharing, nice summary figure!
", "rawMarkdown": "Thanks for sharing, nice summary figure!", "votes": 1, "replies": [ { "id": 1304732, "postDate": "2021-05-12T20:21:01.530Z", "content": "Thank you! Congratulations on the gold medal!)
", "rawMarkdown": "Thank you! Congratulations on the gold medal!)" } ] } ], "comments": [ { "id": 1305574, "author_name": "empty", "author_url": "", "post_date": "2021-05-13T11:23:14.813000", "content": "gratz with good performance outside MADE 💪 :D
", "votes": 1, "replies": [ { "id": 1305750, "author_name": "Mikhail Gurevich", "author_url": "", "post_date": "2021-05-13T13:09:12.580000", "content": "Thanks, as I see you also done great in Shopee competition) Congratulations with Expert tire)
", "votes": 1, "replies": [] }, { "id": 1305773, "author_name": "empty", "author_url": "", "post_date": "2021-05-13T13:22:39.120000", "content": "Well expert was even before MADE but thx 😄
", "votes": 0, "replies": [] } ] }, { "id": 1305031, "author_name": "Tensor Girl", "author_url": "", "post_date": "2021-05-13T04:27:41.470000", "content": "@mgurevich Congratulations and Thanks for sharing the approach
", "votes": 1, "replies": [] }, { "id": 1304399, "author_name": "Alien", "author_url": "", "post_date": "2021-05-12T15:47:56.810000", "content": "Congratulations on Silver medal!
", "votes": 1, "replies": [ { "id": 1304733, "author_name": "Mikhail Gurevich", "author_url": "", "post_date": "2021-05-12T20:21:59.127000", "content": "Thanks and congratulations to you too!)
", "votes": 0, "replies": [] } ] }, { "id": 1304143, "author_name": "corochann", "author_url": "", "post_date": "2021-05-12T12:56:10.397000", "content": "Thanks for sharing, nice summary figure!
", "votes": 1, "replies": [ { "id": 1304732, "author_name": "Mikhail Gurevich", "author_url": "", "post_date": "2021-05-12T20:21:01.530000", "content": "Thank you! Congratulations on the gold medal!)
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "1303667": "Hi all! It was a very long and intensive 3 months)\nFirst of all, many thanks to the organizers for this very challenging and interesting competition!\n\nThis was my first \"real\" competition at kaggle, so I'm glad to get a medal :-)\n\nMy solution consists of several parts:\n\n1) **Cell segmentation**: I trained MaskRCNN on the HPASegmentator predictions. The only reason for this was the inference speed. MaskRCNN works really faster. A few experiments showed that solution score stayed the same when swapping HPASegmentator with MaskRCNN.\n\n2) **Image Level classifiers**:\n\n- EfficientNetB3 (1024x1024)\n- EfficientNetB3 (only green channel) (1024x1024) (thanks @h053473666 for this idea)\n\n3) **Cell level classifiers**:\n I used [PuzzleCAM paper](https://arxiv.org/abs/2101.11253) approach for creating CAMs. And then I used them to create pseudo labels (at this point the labels were smooth - floats in [0.0, 1.0]) for each cell within images.\n Next step was to manually set up thresholds for each class and get final pseudo labels for further training model.\n\n- EfficientNetB3 with dropout (256x256)\n- EfficientNetB3 with dropout (only green channel) (256x256)\n\n\n4) **Final submission** I combined predictions from all 4 networks and get final predictions for each cell. The best result was with just an average of predictions from 4 models.\n\n5) **Training details**\nAll classifiers were trained with focal_loss + lovazh_loss (like in the winning solution of @bestfitting in previous HPA competition).\nOptimizer: Adam\nScheduler: ReduceOnPlatoue\nI also used oversampling for rare classes and undersampling for too common classes.\n\nHere is a summary:\n\n\n\nThanks for reading and good luck in future competitions!)\n", "1305574": "gratz with good performance outside MADE 💪 :D", "1305031": "@mgurevich Congratulations and Thanks for sharing the approach", "1304399": "Congratulations on Silver medal!", "1304143": "Thanks for sharing, nice summary figure!" } }