{ "id": 335026, "title": "HuBMAP + HPA Keras Segmentation", "url": "/competitions/hubmap-organ-segmentation/discussion/335026", "author_name": "", "post_date": "2022-07-04T10:00:55.162755Z", "votes": 2, "comment_count": 2, "views": 0, "content": "
    \n
  1. Create tfrecords
    \nhttps://www.kaggle.com/code/aleksandrkruchinin/hubmap-hpa-images-tfrecords
    2. \n
  2. Tfrecords dataset
    \nhttps://www.kaggle.com/datasets/aleksandrkruchinin/hubmaphpa-im512-reduce2
    3. \n
  3. Train
    \nhttps://www.kaggle.com/aleksandrkruchinin/tpu-hubmap-hpa-train-tensorflow
    4. \n
  4. Train models
    \nhttps://www.kaggle.com/datasets/aleksandrkruchinin/hubmap-keras-model-start-1/settings
    5. \n
  5. Submission
    \nhttps://www.kaggle.com/code/aleksandrkruchinin/gpu-hubmap-hpa-tensorflow-submission
    6. \n
", "messages": [ { "id": "1842814", "postDate": "07/04/2022 10:00:55", "content": "
    \n
  1. Create tfrecords
    \nhttps://www.kaggle.com/code/aleksandrkruchinin/hubmap-hpa-images-tfrecords
    2. \n
  2. Tfrecords dataset
    \nhttps://www.kaggle.com/datasets/aleksandrkruchinin/hubmaphpa-im512-reduce2
    3. \n
  3. Train
    \nhttps://www.kaggle.com/aleksandrkruchinin/tpu-hubmap-hpa-train-tensorflow
    4. \n
  4. Train models
    \nhttps://www.kaggle.com/datasets/aleksandrkruchinin/hubmap-keras-model-start-1/settings
    5. \n
  5. Submission
    \nhttps://www.kaggle.com/code/aleksandrkruchinin/gpu-hubmap-hpa-tensorflow-submission
    6. \n
", "rawMarkdown": "1. Create tfrecords\nhttps://www.kaggle.com/code/aleksandrkruchinin/hubmap-hpa-images-tfrecords\n2. Tfrecords dataset\nhttps://www.kaggle.com/datasets/aleksandrkruchinin/hubmaphpa-im512-reduce2\n3. Train\nhttps://www.kaggle.com/aleksandrkruchinin/tpu-hubmap-hpa-train-tensorflow\n4. Train models\nhttps://www.kaggle.com/datasets/aleksandrkruchinin/hubmap-keras-model-start-1/settings\n5. Submission\nhttps://www.kaggle.com/code/aleksandrkruchinin/gpu-hubmap-hpa-tensorflow-submission", "votes": null }, { "id": "1843902", "postDate": "07/05/2022 07:31:33", "content": "

Great job on putting together this getting started guide for the HuBMAP + HPA segmentation challenge! Your work will be a huge help to other members of the community who are just getting started with the challenge.

\n

A few things that might be helpful to add to your guide:

\n\n

Thanks again for your work on this and good luck to you in the challenge!

", "rawMarkdown": "Great job on putting together this getting started guide for the HuBMAP + HPA segmentation challenge! Your work will be a huge help to other members of the community who are just getting started with the challenge.\n\n**A few things that might be helpful to add to your guide:**\n\n- A brief overview of the overall pipeline you built. \n- A description for each of the links. Just so we know what you provide.\n\nThanks again for your work on this and good luck to you in the challenge!", "votes": null }, { "id": "1844072", "postDate": "07/05/2022 10:24:19", "content": "

This is a simple segmentation approach that was used in the competition:
\nhttps://www.kaggle.com/competitions/hubmap-kidney-segmentation
\nThe EfficientNet model is trained on data without regard to organ type and scale. Images are cut into blocks of size 1024, which are reduced by a factor of 2 to 512. In inference images are also cut into similar sizes. If they are less than 1024 then they are stretched to 1024.

\n
    \n
  1. https://www.kaggle.com/code/aleksandrkruchinin/hubmap-hpa-images-tfrecords
    \nCreating tfrecords for TensorFlow. The images are cut into 1024 blocks, which are reduced by a factor of 2 to 512.
    2. \n
  2. Training dataset received by the previous notebook
    \nhttps://www.kaggle.com/datasets/aleksandrkruchinin/hubmaphpa-im512-reduce2
    3. \n
  3. https://www.kaggle.com/aleksandrkruchinin/tpu-hubmap-hpa-train-tensorflow
    \nIn this notebook I was training our model for the number of 4 FOLDS. I saved the models with best val_diceCoefficient.
    \nGreat notebook with commentary explaining the training:
    \nhttps://www.kaggle.com/code/kool777/training-hubmap-eda-tf-keras-tpu/notebook
    4. \n
  4. https://www.kaggle.com/datasets/aleksandrkruchinin/hubmap-keras-model-start-1/settings
    \nDataset with models (saved from training notebook)
    5. \n
  5. https://www.kaggle.com/code/aleksandrkruchinin/gpu-hubmap-hpa-tensorflow-submission
    \nSubmittion notebook
    \nThe images are also cut into similar sizes and if they are less than 1024 then they are stretched to 1024. After resizing to 512x512 a prediction is made and using the average prediction from 4 models. Result on training data:
    \nId = 9777
    \nOriginal image, training mask, inference
    \n\"\"
    \nResult on the test image
    \n\"\"
    \nLB =0.5
    6. \n
", "rawMarkdown": "This is a simple segmentation approach that was used in the competition:\n[https://www.kaggle.com/competitions/hubmap-kidney-segmentation](url)\nThe EfficientNet model is trained on data without regard to organ type and scale. Images are cut into blocks of size 1024, which are reduced by a factor of 2 to 512. In inference images are also cut into similar sizes. If they are less than 1024 then they are stretched to 1024.\n\n1. [https://www.kaggle.com/code/aleksandrkruchinin/hubmap-hpa-images-tfrecords](url)\nCreating tfrecords for TensorFlow. The images are cut into 1024 blocks, which are reduced by a factor of 2 to 512.\n2. Training dataset received by the previous notebook\n[https://www.kaggle.com/datasets/aleksandrkruchinin/hubmaphpa-im512-reduce2](url)\n3. [https://www.kaggle.com/aleksandrkruchinin/tpu-hubmap-hpa-train-tensorflow](url)\nIn this notebook I was training our model for the number of 4 FOLDS. I saved the models with best val_diceCoefficient.\nGreat notebook with commentary explaining the training:\n[https://www.kaggle.com/code/kool777/training-hubmap-eda-tf-keras-tpu/notebook](url)\n4. [https://www.kaggle.com/datasets/aleksandrkruchinin/hubmap-keras-model-start-1/settings](url)\nDataset with models (saved from training notebook)\n5. https://www.kaggle.com/code/aleksandrkruchinin/gpu-hubmap-hpa-tensorflow-submission\nSubmittion notebook\nThe images are also cut into similar sizes and if they are less than 1024 then they are stretched to 1024. After resizing to 512x512 a prediction is made and using the average prediction from 4 models. Result on training data:\nId = 9777\nOriginal image, training mask, inference\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4159010%2F7933121532940e1ff30bcd7215fdf3de%2F2022-07-04_14-44-22.png?generation=1657016376257403&alt=media)\nResult on the test image\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4159010%2Fb261b831f47dec8e58f118ac7da56979%2F2022-07-04_14-31-03.png?generation=1657016456305852&alt=media)\nLB =0.5", "votes": null } ], "comments": [ { "id": 1843902, "author_name": "thedevastator", "author_url": "", "post_date": "07/05/2022 07:31:33", "content": "

Great job on putting together this getting started guide for the HuBMAP + HPA segmentation challenge! Your work will be a huge help to other members of the community who are just getting started with the challenge.

\n

A few things that might be helpful to add to your guide:

\n\n

Thanks again for your work on this and good luck to you in the challenge!

", "votes": null, "replies": [] }, { "id": 1844072, "author_name": "aleksandrkruchinin", "author_url": "", "post_date": "07/05/2022 10:24:19", "content": "

This is a simple segmentation approach that was used in the competition:
\nhttps://www.kaggle.com/competitions/hubmap-kidney-segmentation
\nThe EfficientNet model is trained on data without regard to organ type and scale. Images are cut into blocks of size 1024, which are reduced by a factor of 2 to 512. In inference images are also cut into similar sizes. If they are less than 1024 then they are stretched to 1024.

\n
    \n
  1. https://www.kaggle.com/code/aleksandrkruchinin/hubmap-hpa-images-tfrecords
    \nCreating tfrecords for TensorFlow. The images are cut into 1024 blocks, which are reduced by a factor of 2 to 512.
    2. \n
  2. Training dataset received by the previous notebook
    \nhttps://www.kaggle.com/datasets/aleksandrkruchinin/hubmaphpa-im512-reduce2
    3. \n
  3. https://www.kaggle.com/aleksandrkruchinin/tpu-hubmap-hpa-train-tensorflow
    \nIn this notebook I was training our model for the number of 4 FOLDS. I saved the models with best val_diceCoefficient.
    \nGreat notebook with commentary explaining the training:
    \nhttps://www.kaggle.com/code/kool777/training-hubmap-eda-tf-keras-tpu/notebook
    4. \n
  4. https://www.kaggle.com/datasets/aleksandrkruchinin/hubmap-keras-model-start-1/settings
    \nDataset with models (saved from training notebook)
    5. \n
  5. https://www.kaggle.com/code/aleksandrkruchinin/gpu-hubmap-hpa-tensorflow-submission
    \nSubmittion notebook
    \nThe images are also cut into similar sizes and if they are less than 1024 then they are stretched to 1024. After resizing to 512x512 a prediction is made and using the average prediction from 4 models. Result on training data:
    \nId = 9777
    \nOriginal image, training mask, inference
    \n\"\"
    \nResult on the test image
    \n\"\"
    \nLB =0.5
    6. \n
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1842814": "1. Create tfrecords\nhttps://www.kaggle.com/code/aleksandrkruchinin/hubmap-hpa-images-tfrecords\n2. Tfrecords dataset\nhttps://www.kaggle.com/datasets/aleksandrkruchinin/hubmaphpa-im512-reduce2\n3. Train\nhttps://www.kaggle.com/aleksandrkruchinin/tpu-hubmap-hpa-train-tensorflow\n4. Train models\nhttps://www.kaggle.com/datasets/aleksandrkruchinin/hubmap-keras-model-start-1/settings\n5. Submission\nhttps://www.kaggle.com/code/aleksandrkruchinin/gpu-hubmap-hpa-tensorflow-submission", "1843902": "Great job on putting together this getting started guide for the HuBMAP + HPA segmentation challenge! Your work will be a huge help to other members of the community who are just getting started with the challenge.\n\n**A few things that might be helpful to add to your guide:**\n\n- A brief overview of the overall pipeline you built. \n- A description for each of the links. Just so we know what you provide.\n\nThanks again for your work on this and good luck to you in the challenge!", "1844072": "This is a simple segmentation approach that was used in the competition:\n[https://www.kaggle.com/competitions/hubmap-kidney-segmentation](url)\nThe EfficientNet model is trained on data without regard to organ type and scale. Images are cut into blocks of size 1024, which are reduced by a factor of 2 to 512. In inference images are also cut into similar sizes. If they are less than 1024 then they are stretched to 1024.\n\n1. [https://www.kaggle.com/code/aleksandrkruchinin/hubmap-hpa-images-tfrecords](url)\nCreating tfrecords for TensorFlow. The images are cut into 1024 blocks, which are reduced by a factor of 2 to 512.\n2. Training dataset received by the previous notebook\n[https://www.kaggle.com/datasets/aleksandrkruchinin/hubmaphpa-im512-reduce2](url)\n3. [https://www.kaggle.com/aleksandrkruchinin/tpu-hubmap-hpa-train-tensorflow](url)\nIn this notebook I was training our model for the number of 4 FOLDS. I saved the models with best val_diceCoefficient.\nGreat notebook with commentary explaining the training:\n[https://www.kaggle.com/code/kool777/training-hubmap-eda-tf-keras-tpu/notebook](url)\n4. [https://www.kaggle.com/datasets/aleksandrkruchinin/hubmap-keras-model-start-1/settings](url)\nDataset with models (saved from training notebook)\n5. https://www.kaggle.com/code/aleksandrkruchinin/gpu-hubmap-hpa-tensorflow-submission\nSubmittion notebook\nThe images are also cut into similar sizes and if they are less than 1024 then they are stretched to 1024. After resizing to 512x512 a prediction is made and using the average prediction from 4 models. Result on training data:\nId = 9777\nOriginal image, training mask, inference\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4159010%2F7933121532940e1ff30bcd7215fdf3de%2F2022-07-04_14-44-22.png?generation=1657016376257403&alt=media)\nResult on the test image\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4159010%2Fb261b831f47dec8e58f118ac7da56979%2F2022-07-04_14-31-03.png?generation=1657016456305852&alt=media)\nLB =0.5" }, "source": "meta" }