{ "id": 354857, "title": "2nd Place Solution", "url": "/competitions/hubmap-organ-segmentation/discussion/354857", "author_name": "Victor Durnov", "post_date": "2022-09-24T09:25:54.665000", "votes": 56, "comment_count": 11, "views": 0, "content": "

Big thanks to the organizers for the great competition!
\nAlso thanks to all competitors for sharing their experiments here, especially @hengck23

\n

Models

\n

In this competition the heavy encoders and larger resolution worked better. I've used 3 CNN encoders (efficientnet_b7, convnext_large, tf_efficientnetv2_l) and 1 transformer (coat_lite_medium). Coat performed the best as single model, but ensemble with CNNs scored more. Also tried few versions of swin v1 and v2, but it performed worse.

\n

All models trained on 3 input resolutions: 768 * 768, 1024 * 1024, 1472 * 1472 with 5 folds.

\n

Models also trained to predict organ and pixel_size. I think these aux outputs help to train more robust model. pixel_size calculated for resized input resolution and changed during training augmentations.

\n

Augmentations

\n

random cropping/padding
\nscaling
\nrotating
\nflipping
\ncolor changing
\nblur/noise
\nsaturation/brightness/contrast
\nelastic

\n

External data

\n

External data helped a lot here. I've download some HPA data with this notebook https://www.kaggle.com/code/carnozhao/hpa-data-download and picked some images manually from sources posted here and previous Hubmap and Panda competitions.

\n

Pseudo labeled

\n

All external data pseudo-labeled using ensemble of initial models.
\nAlso training data was pseudo-labeled and in 30% used as ground truth for training.

\n

Color transfering

\n

Training images recolored using this great notebook https://www.kaggle.com/code/gray98/stain-normalization-color-transfer with 3 different target images and used with 15% chance instead of original during training.

\n

Validation

\n

Test-time augmentations used on validation (flip, crop, padding) + external data also separated on folds and used as validation to get best checkpoints.

\n

Inference notebook: https://www.kaggle.com/code/victorsd/2nd-place-inference/notebook?scriptVersionId=106240458

", "messages": [ { "id": 1953175, "postDate": "2022-09-24T09:25:54.667Z", "content": "

Big thanks to the organizers for the great competition!
\nAlso thanks to all competitors for sharing their experiments here, especially @hengck23

\n

Models

\n

In this competition the heavy encoders and larger resolution worked better. I've used 3 CNN encoders (efficientnet_b7, convnext_large, tf_efficientnetv2_l) and 1 transformer (coat_lite_medium). Coat performed the best as single model, but ensemble with CNNs scored more. Also tried few versions of swin v1 and v2, but it performed worse.

\n

All models trained on 3 input resolutions: 768 * 768, 1024 * 1024, 1472 * 1472 with 5 folds.

\n

Models also trained to predict organ and pixel_size. I think these aux outputs help to train more robust model. pixel_size calculated for resized input resolution and changed during training augmentations.

\n

Augmentations

\n

random cropping/padding
\nscaling
\nrotating
\nflipping
\ncolor changing
\nblur/noise
\nsaturation/brightness/contrast
\nelastic

\n

External data

\n

External data helped a lot here. I've download some HPA data with this notebook https://www.kaggle.com/code/carnozhao/hpa-data-download and picked some images manually from sources posted here and previous Hubmap and Panda competitions.

\n

Pseudo labeled

\n

All external data pseudo-labeled using ensemble of initial models.
\nAlso training data was pseudo-labeled and in 30% used as ground truth for training.

\n

Color transfering

\n

Training images recolored using this great notebook https://www.kaggle.com/code/gray98/stain-normalization-color-transfer with 3 different target images and used with 15% chance instead of original during training.

\n

Validation

\n

Test-time augmentations used on validation (flip, crop, padding) + external data also separated on folds and used as validation to get best checkpoints.

\n

Inference notebook: https://www.kaggle.com/code/victorsd/2nd-place-inference/notebook?scriptVersionId=106240458

", "rawMarkdown": "Big thanks to the organizers for the great competition!\nAlso thanks to all competitors for sharing their experiments here, especially @hengck23\n\n# Models\n\nIn this competition the heavy encoders and larger resolution worked better. I've used 3 CNN encoders (efficientnet_b7, convnext_large, tf_efficientnetv2_l) and 1 transformer (coat_lite_medium). Coat performed the best as single model, but ensemble with CNNs scored more. Also tried few versions of swin v1 and v2, but it performed worse.\n\nAll models trained on 3 input resolutions: 768 * 768, 1024 * 1024, 1472 * 1472 with 5 folds.\n\nModels also trained to predict organ and pixel_size. I think these aux outputs help to train more robust model. pixel_size calculated for resized input resolution and changed during training augmentations.\n\n# Augmentations\n\nrandom cropping/padding\nscaling\nrotating\nflipping\ncolor changing\nblur/noise\nsaturation/brightness/contrast\nelastic\n\n# External data\n\nExternal data helped a lot here. I've download some HPA data with this notebook https://www.kaggle.com/code/carnozhao/hpa-data-download and picked some images manually from sources posted here and previous Hubmap and Panda competitions. \n\n# Pseudo labeled\n\nAll external data pseudo-labeled using ensemble of initial models. \nAlso training data was pseudo-labeled and in 30% used as ground truth for training.\n\n# Color transfering\n\nTraining images recolored using this great notebook https://www.kaggle.com/code/gray98/stain-normalization-color-transfer with 3 different target images and used with 15% chance instead of original during training.\n\n# Validation\n\nTest-time augmentations used on validation (flip, crop, padding) + external data also separated on folds and used as validation to get best checkpoints. \n\n\nInference notebook: https://www.kaggle.com/code/victorsd/2nd-place-inference/notebook?scriptVersionId=106240458", "votes": 56 }, { "id": 1956486, "postDate": "2022-09-26T13:30:06.060Z", "content": "

Congratulations , really good summary! Could you provide us with the code so we can understand the solution better ?
\nOne thing that i wanted to mention here,
\nModels also trained to predict organ and pixel_size. I think these aux outputs help to train more robust model. pixel_size calculated for resized input resolution and changed during training augmentations.
\n It was a brilliant idea!

", "rawMarkdown": "Congratulations , really good summary! Could you provide us with the code so we can understand the solution better ?\nOne thing that i wanted to mention here, \n`Models also trained to predict organ and pixel_size. I think these aux outputs help to train more robust model. pixel_size calculated for resized input resolution and changed during training augmentations.`\n It was a brilliant idea!", "votes": 1 }, { "id": 1953614, "postDate": "2022-09-24T15:58:28.807Z", "content": "

Congratulations! Will you post your training notebooks?

", "rawMarkdown": "Congratulations! Will you post your training notebooks?", "votes": 1, "replies": [ { "id": 1954096, "postDate": "2022-09-25T02:49:47.600Z", "content": "

I'll check this with the organizers when will prepare my solution. It is a mess of files/experiments now, not notebooks.

", "rawMarkdown": "I'll check this with the organizers when will prepare my solution. It is a mess of files/experiments now, not notebooks.", "votes": 2 } ] }, { "id": 1953184, "postDate": "2022-09-24T09:33:52.633Z", "content": "

Congratulations! I didn't realize that external data and pseudo label help such a lot, thanks for your sharing! :)

", "rawMarkdown": "Congratulations! I didn't realize that external data and pseudo label help such a lot, thanks for your sharing! :)", "votes": 1, "replies": [ { "id": 1953191, "postDate": "2022-09-24T09:42:12.743Z", "content": "

Thanks! Your stain normalization for test on inference is also interesting!

", "rawMarkdown": "Thanks! Your stain normalization for test on inference is also interesting!", "votes": 1 } ] }, { "id": 2004177, "postDate": "2022-10-26T06:08:21.870Z", "content": "

Great work. Can you share the loss function and LR scheduler used for training these models? Loss function is the most interesting thing with your aux outputs.

", "rawMarkdown": "Great work. Can you share the loss function and LR scheduler used for training these models? Loss function is the most interesting thing with your aux outputs. ", "replies": [ { "id": 2004203, "postDate": "2022-10-26T06:25:28.620Z", "content": "

dice+focal for segmentation, bce for organ classification, mse for pixel_size

", "rawMarkdown": "dice+focal for segmentation, bce for organ classification, mse for pixel_size" }, { "id": 2004205, "postDate": "2022-10-26T06:26:20.823Z", "content": "

lr = 1e-4, ReduceLROnPlateau

", "rawMarkdown": "lr = 1e-4, ReduceLROnPlateau", "votes": 1 }, { "id": 2004218, "postDate": "2022-10-26T06:32:54.270Z", "content": "

Thanks a lot for your quick response. Learned so many things here. I going to try to recreate your solution.

", "rawMarkdown": "Thanks a lot for your quick response. Learned so many things here. I going to try to recreate your solution." } ] }, { "id": 1957580, "postDate": "2022-09-27T04:27:18.430Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 1957692, "postDate": "2022-09-27T06:00:14.073Z", "content": "

2*a6000 (48gb). Really good GPU for larger models and input size.

", "rawMarkdown": "2*a6000 (48gb). Really good GPU for larger models and input size." } ] } ], "comments": [ { "id": 1956486, "author_name": "MohamedAmine SAIGHI", "author_url": "", "post_date": "2022-09-26T13:30:06.060000", "content": "

Congratulations , really good summary! Could you provide us with the code so we can understand the solution better ?
\nOne thing that i wanted to mention here,
\nModels also trained to predict organ and pixel_size. I think these aux outputs help to train more robust model. pixel_size calculated for resized input resolution and changed during training augmentations.
\n It was a brilliant idea!

", "votes": 1, "replies": [] }, { "id": 1953614, "author_name": "Swikwislkdjc", "author_url": "", "post_date": "2022-09-24T15:58:28.807000", "content": "

Congratulations! Will you post your training notebooks?

", "votes": 1, "replies": [ { "id": 1954096, "author_name": "Victor Durnov", "author_url": "", "post_date": "2022-09-25T02:49:47.600000", "content": "

I'll check this with the organizers when will prepare my solution. It is a mess of files/experiments now, not notebooks.

", "votes": 2, "replies": [] } ] }, { "id": 1953184, "author_name": "Rock", "author_url": "", "post_date": "2022-09-24T09:33:52.633000", "content": "

Congratulations! I didn't realize that external data and pseudo label help such a lot, thanks for your sharing! :)

", "votes": 1, "replies": [ { "id": 1953191, "author_name": "Victor Durnov", "author_url": "", "post_date": "2022-09-24T09:42:12.743000", "content": "

Thanks! Your stain normalization for test on inference is also interesting!

", "votes": 1, "replies": [] } ] }, { "id": 2004177, "author_name": "Mugdha Hardikar", "author_url": "", "post_date": "2022-10-26T06:08:21.870000", "content": "

Great work. Can you share the loss function and LR scheduler used for training these models? Loss function is the most interesting thing with your aux outputs.

", "votes": 0, "replies": [ { "id": 2004203, "author_name": "Victor Durnov", "author_url": "", "post_date": "2022-10-26T06:25:28.620000", "content": "

dice+focal for segmentation, bce for organ classification, mse for pixel_size

", "votes": 0, "replies": [] }, { "id": 2004205, "author_name": "Victor Durnov", "author_url": "", "post_date": "2022-10-26T06:26:20.823000", "content": "

lr = 1e-4, ReduceLROnPlateau

", "votes": 1, "replies": [] }, { "id": 2004218, "author_name": "Mugdha Hardikar", "author_url": "", "post_date": "2022-10-26T06:32:54.270000", "content": "

Thanks a lot for your quick response. Learned so many things here. I going to try to recreate your solution.

", "votes": 0, "replies": [] } ] }, { "id": 1957580, "author_name": "", "author_url": "", "post_date": "2022-09-27T04:27:18.430000", "content": "", "votes": 0, "replies": [ { "id": 1957692, "author_name": "Victor Durnov", "author_url": "", "post_date": "2022-09-27T06:00:14.073000", "content": "

2*a6000 (48gb). Really good GPU for larger models and input size.

", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "1953175": "Big thanks to the organizers for the great competition!\nAlso thanks to all competitors for sharing their experiments here, especially @hengck23\n\n# Models\n\nIn this competition the heavy encoders and larger resolution worked better. I've used 3 CNN encoders (efficientnet_b7, convnext_large, tf_efficientnetv2_l) and 1 transformer (coat_lite_medium). Coat performed the best as single model, but ensemble with CNNs scored more. Also tried few versions of swin v1 and v2, but it performed worse.\n\nAll models trained on 3 input resolutions: 768 * 768, 1024 * 1024, 1472 * 1472 with 5 folds.\n\nModels also trained to predict organ and pixel_size. I think these aux outputs help to train more robust model. pixel_size calculated for resized input resolution and changed during training augmentations.\n\n# Augmentations\n\nrandom cropping/padding\nscaling\nrotating\nflipping\ncolor changing\nblur/noise\nsaturation/brightness/contrast\nelastic\n\n# External data\n\nExternal data helped a lot here. I've download some HPA data with this notebook https://www.kaggle.com/code/carnozhao/hpa-data-download and picked some images manually from sources posted here and previous Hubmap and Panda competitions. \n\n# Pseudo labeled\n\nAll external data pseudo-labeled using ensemble of initial models. \nAlso training data was pseudo-labeled and in 30% used as ground truth for training.\n\n# Color transfering\n\nTraining images recolored using this great notebook https://www.kaggle.com/code/gray98/stain-normalization-color-transfer with 3 different target images and used with 15% chance instead of original during training.\n\n# Validation\n\nTest-time augmentations used on validation (flip, crop, padding) + external data also separated on folds and used as validation to get best checkpoints. \n\n\nInference notebook: https://www.kaggle.com/code/victorsd/2nd-place-inference/notebook?scriptVersionId=106240458", "1956486": "Congratulations , really good summary! Could you provide us with the code so we can understand the solution better ?\nOne thing that i wanted to mention here, \n`Models also trained to predict organ and pixel_size. I think these aux outputs help to train more robust model. pixel_size calculated for resized input resolution and changed during training augmentations.`\n It was a brilliant idea!", "1953614": "Congratulations! Will you post your training notebooks?", "1953184": "Congratulations! I didn't realize that external data and pseudo label help such a lot, thanks for your sharing! :)", "2004177": "Great work. Can you share the loss function and LR scheduler used for training these models? Loss function is the most interesting thing with your aux outputs. ", "1957580": "" } }