{ "id": 280137, "title": "Tutorial: detectron 2", "url": "/competitions/sartorius-cell-instance-segmentation/discussion/280137", "author_name": "", "post_date": "2021-10-20T16:45:01.647794500Z", "votes": 83, "comment_count": 9, "views": 0, "content": "

I've created a three part end-to-end tutorial using the detectron library that gives .217 LB score. It was more difficult to get a working baseline than in other competitions I've seen so far, but in the end I arrived with a quite short and clean solution, after I figured out how to use the tools.

\n

Here are the main things I stumbled upon and tips to overcome them:

\n\n

Resources:

\n", "messages": [ { "id": "1551473", "postDate": "10/20/2021 16:45:01", "content": "

I've created a three part end-to-end tutorial using the detectron library that gives .217 LB score. It was more difficult to get a working baseline than in other competitions I've seen so far, but in the end I arrived with a quite short and clean solution, after I figured out how to use the tools.

\n

Here are the main things I stumbled upon and tips to overcome them:

\n\n

Resources:

\n", "rawMarkdown": "I've created a three part end-to-end tutorial using the [detectron](https://github.com/facebookresearch/detectron2) library that gives .217 LB score. It was more difficult to get a working baseline than in other competitions I've seen so far, but in the end I arrived with a quite short and clean solution, after I figured out how to use the tools.\n\nHere are the main things I stumbled upon and tips to overcome them:\n- Dealing with masks can be tricky, I've encounterd few slightly different definitions of RLE, and it took me a while to debug that the one used in this competiotion is different than the one in the COCO dataset. It got me paranoid enough that I started displaying outputs on every step, but once I got all the annotations in the COCO json format the rest became much easier. I'm sharing the dataset so others can take adventage of it.\n- The actual training of a basic model was simple to do. I followed the official tutorial: https://colab.research.google.com/drive/16jcaJoc6bCFAQ96jDe2HwtXj7BMD_-m5 and it just worked. I'm not sure how hard it will be to further improve and optimize but getting an initial solution was easy.\n- It took me more than I have expected to get all the detectron dependencies work in the offline inference notebook. Maybe it's basic to others but I haven't done it before and had to experiment with diffent pip download and pip install incantations. Again I'm sharing the dataset so you can just use it.\n\n\nResources:\n- Notebooks: [Part 1](https://www.kaggle.com/slawekbiel/positive-score-with-detectron-1-3-input-data) [Part 2](https://www.kaggle.com/slawekbiel/positive-score-with-detectron-2-3-training) [Part 3](https://www.kaggle.com/slawekbiel/positive-score-with-detectron-3-3-inference)\n- Annotations in the COCO format: https://www.kaggle.com/slawekbiel/sartorius-cell-instance-segmentation-coco\n- Detectron library: https://www.kaggle.com/slawekbiel/detectron-05\n- Trained Mask RCNN weights: https://www.kaggle.com/slawekbiel/sartorius-models", "votes": null }, { "id": "1553307", "postDate": "10/22/2021 04:40:57", "content": "

great work. nice sharing.

", "rawMarkdown": "great work. nice sharing.", "votes": null }, { "id": "1554398", "postDate": "10/23/2021 04:09:02", "content": "

nice work. Thanks

", "rawMarkdown": "nice work. Thanks", "votes": null }, { "id": "1559426", "postDate": "10/27/2021 01:51:59", "content": "

great work, thanks for sharing

", "rawMarkdown": "great work, thanks for sharing", "votes": null }, { "id": "1591738", "postDate": "11/22/2021 15:24:06", "content": "

amazing work!

", "rawMarkdown": "amazing work!", "votes": null }, { "id": "1631639", "postDate": "12/28/2021 16:26:08", "content": "

hey :) can you please share how to get all the detectron dependencies work?

", "rawMarkdown": "hey :) can you please share how to get all the detectron dependencies work?", "votes": null }, { "id": "1631731", "postDate": "12/28/2021 18:11:13", "content": "

See the notebooks I linked.
\n “Part 3” does the offline installation.

", "rawMarkdown": "See the notebooks I linked.\n “Part 3” does the offline installation.", "votes": null }, { "id": "1633015", "postDate": "12/30/2021 10:20:45", "content": "

Good work!

", "rawMarkdown": "Good work!", "votes": null }, { "id": "1633069", "postDate": "12/30/2021 11:33:58", "content": "

thanks! i tried this code on my notebook and always get
\n\" error: could not create 'build': Read-only file system

\n
\n

ERROR: Failed building wheel for fvcore \"

\n

same for antlr4-python3-runtime

\n

have any idea how to make it work?

", "rawMarkdown": "thanks! i tried this code on my notebook and always get \n\" error: could not create 'build': Read-only file system\n ----------------------------------------\n ERROR: Failed building wheel for fvcore \"\n\n\nsame for antlr4-python3-runtime\n\nhave any idea how to make it work?", "votes": null }, { "id": "2942614", "postDate": "08/01/2024 00:28:26", "content": "

thank you, very good

", "rawMarkdown": "thank you, very good", "votes": null } ], "comments": [ { "id": 1553307, "author_name": "dragonzhang", "author_url": "", "post_date": "10/22/2021 04:40:57", "content": "

great work. nice sharing.

", "votes": null, "replies": [] }, { "id": 1554398, "author_name": "namgalielei", "author_url": "", "post_date": "10/23/2021 04:09:02", "content": "

nice work. Thanks

", "votes": null, "replies": [] }, { "id": 1559426, "author_name": "ramu199531", "author_url": "", "post_date": "10/27/2021 01:51:59", "content": "

great work, thanks for sharing

", "votes": null, "replies": [] }, { "id": 1591738, "author_name": "ccvipchenbin", "author_url": "", "post_date": "11/22/2021 15:24:06", "content": "

amazing work!

", "votes": null, "replies": [] }, { "id": 1631639, "author_name": "orlevy11", "author_url": "", "post_date": "12/28/2021 16:26:08", "content": "

hey :) can you please share how to get all the detectron dependencies work?

", "votes": null, "replies": [ { "id": 1631731, "author_name": "slawekbiel", "author_url": "", "post_date": "12/28/2021 18:11:13", "content": "

See the notebooks I linked.
\n “Part 3” does the offline installation.

", "votes": null, "replies": [] }, { "id": 1633069, "author_name": "orlevy11", "author_url": "", "post_date": "12/30/2021 11:33:58", "content": "

thanks! i tried this code on my notebook and always get
\n\" error: could not create 'build': Read-only file system

\n
\n

ERROR: Failed building wheel for fvcore \"

\n

same for antlr4-python3-runtime

\n

have any idea how to make it work?

", "votes": null, "replies": [] } ] }, { "id": 1633015, "author_name": "towhidultonmoy", "author_url": "", "post_date": "12/30/2021 10:20:45", "content": "

Good work!

", "votes": null, "replies": [] }, { "id": 2942614, "author_name": "liuweiq", "author_url": "", "post_date": "08/01/2024 00:28:26", "content": "

thank you, very good

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1551473": "I've created a three part end-to-end tutorial using the [detectron](https://github.com/facebookresearch/detectron2) library that gives .217 LB score. It was more difficult to get a working baseline than in other competitions I've seen so far, but in the end I arrived with a quite short and clean solution, after I figured out how to use the tools.\n\nHere are the main things I stumbled upon and tips to overcome them:\n- Dealing with masks can be tricky, I've encounterd few slightly different definitions of RLE, and it took me a while to debug that the one used in this competiotion is different than the one in the COCO dataset. It got me paranoid enough that I started displaying outputs on every step, but once I got all the annotations in the COCO json format the rest became much easier. I'm sharing the dataset so others can take adventage of it.\n- The actual training of a basic model was simple to do. I followed the official tutorial: https://colab.research.google.com/drive/16jcaJoc6bCFAQ96jDe2HwtXj7BMD_-m5 and it just worked. I'm not sure how hard it will be to further improve and optimize but getting an initial solution was easy.\n- It took me more than I have expected to get all the detectron dependencies work in the offline inference notebook. Maybe it's basic to others but I haven't done it before and had to experiment with diffent pip download and pip install incantations. Again I'm sharing the dataset so you can just use it.\n\n\nResources:\n- Notebooks: [Part 1](https://www.kaggle.com/slawekbiel/positive-score-with-detectron-1-3-input-data) [Part 2](https://www.kaggle.com/slawekbiel/positive-score-with-detectron-2-3-training) [Part 3](https://www.kaggle.com/slawekbiel/positive-score-with-detectron-3-3-inference)\n- Annotations in the COCO format: https://www.kaggle.com/slawekbiel/sartorius-cell-instance-segmentation-coco\n- Detectron library: https://www.kaggle.com/slawekbiel/detectron-05\n- Trained Mask RCNN weights: https://www.kaggle.com/slawekbiel/sartorius-models", "1553307": "great work. nice sharing.", "1554398": "nice work. Thanks", "1559426": "great work, thanks for sharing", "1591738": "amazing work!", "1631639": "hey :) can you please share how to get all the detectron dependencies work?", "1631731": "See the notebooks I linked.\n “Part 3” does the offline installation.", "1633015": "Good work!", "1633069": "thanks! i tried this code on my notebook and always get \n\" error: could not create 'build': Read-only file system\n ----------------------------------------\n ERROR: Failed building wheel for fvcore \"\n\n\nsame for antlr4-python3-runtime\n\nhave any idea how to make it work?", "2942614": "thank you, very good" }, "source": "meta" }