{ "id": 238527, "title": "simple idea of 5th, tito's part", "url": "/competitions/hpa-single-cell-image-classification/discussion/238527", "author_name": "tito", "post_date": "2021-05-12T13:12:26.578000", "votes": 49, "comment_count": 14, "views": 0, "content": "

First of all I would like to thank my teammates @narsil and @tivfrvqhs5. I have had a very exciting 3 weeks teaming up with you guys.

\n

I would like to share my model idea briefly here.

\n

Classifier

\n

I created a classification model that predicts labels for each image (not for each cell).

\n

For inference, I extracted the cells one by one and made their augmented images:
\n\"inference

\n

This allowed me to solve this weakly supervised task as a normal classification task which is same as HPA2018 competition.
\nThis makes things very simple and allows me to reuse the HPA2018 solution.

\n

Segmenter

\n

The official segmenter is inaccurate for cells near the boundary (cells with no visible nucleus tend to be connected to other cells).
\nTo avoid this effect, I created a mmdetection model using cropped segmentation image.
\n\"segmentation
\nEven using official segmenter as traing label, score of this mmdetection mode is improved.
\nIn addition, this improved the inference speed and more ensembles enabled.

", "messages": [ { "id": 1304162, "postDate": "2021-05-12T13:12:26.577Z", "content": "

First of all I would like to thank my teammates @narsil and @tivfrvqhs5. I have had a very exciting 3 weeks teaming up with you guys.

\n

I would like to share my model idea briefly here.

\n

Classifier

\n

I created a classification model that predicts labels for each image (not for each cell).

\n

For inference, I extracted the cells one by one and made their augmented images:
\n\"inference

\n

This allowed me to solve this weakly supervised task as a normal classification task which is same as HPA2018 competition.
\nThis makes things very simple and allows me to reuse the HPA2018 solution.

\n

Segmenter

\n

The official segmenter is inaccurate for cells near the boundary (cells with no visible nucleus tend to be connected to other cells).
\nTo avoid this effect, I created a mmdetection model using cropped segmentation image.
\n\"segmentation
\nEven using official segmenter as traing label, score of this mmdetection mode is improved.
\nIn addition, this improved the inference speed and more ensembles enabled.

", "rawMarkdown": "First of all I would like to thank my teammates @narsil and @tivfrvqhs5. I have had a very exciting 3 weeks teaming up with you guys.\n\nI would like to share my model idea briefly here.\n\n## Classifier\nI created a classification model that predicts labels for each image (not for each cell).\n\nFor inference, I extracted the cells one by one and made their augmented images:\n![inference image](https://pbs.twimg.com/media/E1LWd4_VoAEzCUB?format=png&name=900x900)\n\nThis allowed me to solve this weakly supervised task as a normal classification task which is same as HPA2018 competition.\nThis makes things very simple and allows me to reuse the HPA2018 solution.\n\n\n## Segmenter\nThe official segmenter is inaccurate for cells near the boundary (cells with no visible nucleus tend to be connected to other cells).\nTo avoid this effect, I created a mmdetection model using cropped segmentation image.\n![segmentation for training image](https://pbs.twimg.com/media/E1Q3UZHUcAcSgPv?format=jpg&name=small)\nEven using official segmenter as traing label, score of this mmdetection mode is improved.\nIn addition, this improved the inference speed and more ensembles enabled.\n", "votes": 48 }, { "id": 1305745, "postDate": "2021-05-13T12:59:03.517Z", "content": "

Quite clever and simple! It will be interesting to test your augmented images in other solutions to get the best of all.

", "rawMarkdown": "Quite clever and simple! It will be interesting to test your augmented images in other solutions to get the best of all.", "votes": 1 }, { "id": 1305025, "postDate": "2021-05-13T04:26:03Z", "content": "

@its7171 Congratulations and Thanks for sharing the approach

", "rawMarkdown": "@its7171 Congratulations and Thanks for sharing the approach", "votes": 1 }, { "id": 1304798, "postDate": "2021-05-12T22:14:56.010Z", "content": "

Congrats! Interesting approach!

", "rawMarkdown": "Congrats! Interesting approach!", "votes": 1 }, { "id": 1304629, "postDate": "2021-05-12T18:41:17.730Z", "content": "

Congrats on the amazing performance and thanks for sharing, @its7171 ! Elegant!
\nReminded me of a quote by Leonardo da Vinci: \"Simplicity is the ultimate sophistication\"

", "rawMarkdown": "Congrats on the amazing performance and thanks for sharing, @its7171 ! Elegant!\nReminded me of a quote by Leonardo da Vinci: \"Simplicity is the ultimate sophistication\"", "votes": 1 }, { "id": 1304554, "postDate": "2021-05-12T17:35:24.493Z", "content": "

Nice! When we tried tiling a plot with a cropped cell and classifying such an image, we did not really see an improve in score. Should have done the resizing and rotations, apparently :-)

", "rawMarkdown": "Nice! When we tried tiling a plot with a cropped cell and classifying such an image, we did not really see an improve in score. Should have done the resizing and rotations, apparently :-)", "votes": 1 }, { "id": 1304450, "postDate": "2021-05-12T16:18:06.960Z", "content": "

Excellent solution! How long does it take to infer the private test dataset?

", "rawMarkdown": "Excellent solution! How long does it take to infer the private test dataset?", "votes": 1, "replies": [ { "id": 1304918, "postDate": "2021-05-13T01:51:29.173Z", "content": "

It took about 3-4 hours for 10 models.

", "rawMarkdown": "It took about 3-4 hours for 10 models.", "votes": 1 } ] }, { "id": 1304254, "postDate": "2021-05-12T14:09:36.387Z", "content": "

It was great working together!

", "rawMarkdown": "It was great working together! ", "votes": 1 }, { "id": 1304243, "postDate": "2021-05-12T14:04:28.587Z", "content": "

Congratulations!!! This is amazing - we did the same thing, but didn't think about resizing, after looking at your example it seems obvious!

", "rawMarkdown": "Congratulations!!! This is amazing - we did the same thing, but didn't think about resizing, after looking at your example it seems obvious!", "votes": 1 }, { "id": 1304168, "postDate": "2021-05-12T13:17:38.520Z", "content": "

Thanks for sharing. simple but powerful approach.

", "rawMarkdown": "Thanks for sharing. simple but powerful approach.", "votes": 1 }, { "id": 1304449, "postDate": "2021-05-12T16:16:52.870Z", "content": "

Very creative approach! Cool!
\nBut how did you deal with the long inference time, if you made each cell a whole image?

", "rawMarkdown": "Very creative approach! Cool!\nBut how did you deal with the long inference time, if you made each cell a whole image?", "votes": 2, "replies": [ { "id": 1304573, "postDate": "2021-05-12T17:48:40.613Z", "content": "

Actually, as one of our approaches we also classified every single cell (with other cells masked) on a 512x512 image and I think it took around 3-4 h to run the full prediction with 10 models (EfficientNet-b5 and v2s, eca-nfnet-l0 and l1). So, apparently the number of cells in the test set was small enough to fit in the 9 h.

", "rawMarkdown": "Actually, as one of our approaches we also classified every single cell (with other cells masked) on a 512x512 image and I think it took around 3-4 h to run the full prediction with 10 models (EfficientNet-b5 and v2s, eca-nfnet-l0 and l1). So, apparently the number of cells in the test set was small enough to fit in the 9 h.", "votes": 2 }, { "id": 1304919, "postDate": "2021-05-13T01:51:51.447Z", "content": "

Almost same inference time as Ilya's team.

\n
inference time: 3-4 hours for 10 models\nmodels: efficientnet-b0, inceptionv3, densenet121\nresolushon: 512, 960, 1280\n
", "rawMarkdown": "Almost same inference time as Ilya's team.\n\n```\ninference time: 3-4 hours for 10 models\nmodels: efficientnet-b0, inceptionv3, densenet121\nresolushon: 512, 960, 1280\n```", "votes": 2 } ] }, { "id": 1340617, "postDate": "2021-06-08T06:23:55.523Z", "content": "

Congratulations ! Would you be able to possibly share your code ? Will be nice to learn from. Thanks.

", "rawMarkdown": "Congratulations ! Would you be able to possibly share your code ? Will be nice to learn from. Thanks. " } ], "comments": [ { "id": 1305745, "author_name": "MPWARE", "author_url": "", "post_date": "2021-05-13T12:59:03.517000", "content": "

Quite clever and simple! It will be interesting to test your augmented images in other solutions to get the best of all.

", "votes": 1, "replies": [] }, { "id": 1305025, "author_name": "Tensor Girl", "author_url": "", "post_date": "2021-05-13T04:26:03", "content": "

@its7171 Congratulations and Thanks for sharing the approach

", "votes": 1, "replies": [] }, { "id": 1304798, "author_name": "corochann", "author_url": "", "post_date": "2021-05-12T22:14:56.010000", "content": "

Congrats! Interesting approach!

", "votes": 1, "replies": [] }, { "id": 1304629, "author_name": "Raman", "author_url": "", "post_date": "2021-05-12T18:41:17.730000", "content": "

Congrats on the amazing performance and thanks for sharing, @its7171 ! Elegant!
\nReminded me of a quote by Leonardo da Vinci: \"Simplicity is the ultimate sophistication\"

", "votes": 1, "replies": [] }, { "id": 1304554, "author_name": "Ilya Makarov", "author_url": "", "post_date": "2021-05-12T17:35:24.493000", "content": "

Nice! When we tried tiling a plot with a cropped cell and classifying such an image, we did not really see an improve in score. Should have done the resizing and rotations, apparently :-)

", "votes": 1, "replies": [] }, { "id": 1304450, "author_name": "Taku Hiraiwa", "author_url": "", "post_date": "2021-05-12T16:18:06.960000", "content": "

Excellent solution! How long does it take to infer the private test dataset?

", "votes": 1, "replies": [ { "id": 1304918, "author_name": "tito", "author_url": "", "post_date": "2021-05-13T01:51:29.173000", "content": "

It took about 3-4 hours for 10 models.

", "votes": 1, "replies": [] } ] }, { "id": 1304254, "author_name": "narsil (jobs-in-data.com)", "author_url": "", "post_date": "2021-05-12T14:09:36.387000", "content": "

It was great working together!

", "votes": 1, "replies": [] }, { "id": 1304243, "author_name": "Darek Kłeczek", "author_url": "", "post_date": "2021-05-12T14:04:28.587000", "content": "

Congratulations!!! This is amazing - we did the same thing, but didn't think about resizing, after looking at your example it seems obvious!

", "votes": 1, "replies": [] }, { "id": 1304168, "author_name": "yuvaramsingh", "author_url": "", "post_date": "2021-05-12T13:17:38.520000", "content": "

Thanks for sharing. simple but powerful approach.

", "votes": 1, "replies": [] }, { "id": 1304449, "author_name": "Guanshuo Xu", "author_url": "", "post_date": "2021-05-12T16:16:52.870000", "content": "

Very creative approach! Cool!
\nBut how did you deal with the long inference time, if you made each cell a whole image?

", "votes": 2, "replies": [ { "id": 1304573, "author_name": "Ilya Makarov", "author_url": "", "post_date": "2021-05-12T17:48:40.613000", "content": "

Actually, as one of our approaches we also classified every single cell (with other cells masked) on a 512x512 image and I think it took around 3-4 h to run the full prediction with 10 models (EfficientNet-b5 and v2s, eca-nfnet-l0 and l1). So, apparently the number of cells in the test set was small enough to fit in the 9 h.

", "votes": 2, "replies": [] }, { "id": 1304919, "author_name": "tito", "author_url": "", "post_date": "2021-05-13T01:51:51.447000", "content": "

Almost same inference time as Ilya's team.

\n
inference time: 3-4 hours for 10 models\nmodels: efficientnet-b0, inceptionv3, densenet121\nresolushon: 512, 960, 1280\n
", "votes": 2, "replies": [] } ] }, { "id": 1340617, "author_name": "Krishna ", "author_url": "", "post_date": "2021-06-08T06:23:55.523000", "content": "

Congratulations ! Would you be able to possibly share your code ? Will be nice to learn from. Thanks.

", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1304162": "First of all I would like to thank my teammates @narsil and @tivfrvqhs5. I have had a very exciting 3 weeks teaming up with you guys.\n\nI would like to share my model idea briefly here.\n\n## Classifier\nI created a classification model that predicts labels for each image (not for each cell).\n\nFor inference, I extracted the cells one by one and made their augmented images:\n![inference image](https://pbs.twimg.com/media/E1LWd4_VoAEzCUB?format=png&name=900x900)\n\nThis allowed me to solve this weakly supervised task as a normal classification task which is same as HPA2018 competition.\nThis makes things very simple and allows me to reuse the HPA2018 solution.\n\n\n## Segmenter\nThe official segmenter is inaccurate for cells near the boundary (cells with no visible nucleus tend to be connected to other cells).\nTo avoid this effect, I created a mmdetection model using cropped segmentation image.\n![segmentation for training image](https://pbs.twimg.com/media/E1Q3UZHUcAcSgPv?format=jpg&name=small)\nEven using official segmenter as traing label, score of this mmdetection mode is improved.\nIn addition, this improved the inference speed and more ensembles enabled.\n", "1305745": "Quite clever and simple! It will be interesting to test your augmented images in other solutions to get the best of all.", "1305025": "@its7171 Congratulations and Thanks for sharing the approach", "1304798": "Congrats! Interesting approach!", "1304629": "Congrats on the amazing performance and thanks for sharing, @its7171 ! Elegant!\nReminded me of a quote by Leonardo da Vinci: \"Simplicity is the ultimate sophistication\"", "1304554": "Nice! When we tried tiling a plot with a cropped cell and classifying such an image, we did not really see an improve in score. Should have done the resizing and rotations, apparently :-)", "1304450": "Excellent solution! How long does it take to infer the private test dataset?", "1304254": "It was great working together! ", "1304243": "Congratulations!!! This is amazing - we did the same thing, but didn't think about resizing, after looking at your example it seems obvious!", "1304168": "Thanks for sharing. simple but powerful approach.", "1304449": "Very creative approach! Cool!\nBut how did you deal with the long inference time, if you made each cell a whole image?", "1340617": "Congratulations ! Would you be able to possibly share your code ? Will be nice to learn from. Thanks. " } }