![\"Image\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F5570735%2Fd2f96fa3445f7c9685e60e12860883dd%2FV2.JPG?generation=1723826153577715&alt=media\")
After some positive feedback on the coordinate pretraining dataset, I have decided to release some more data.
\nThis time I made additions to the competition coordinate data by annotating the left side of each disc. This provides orientation information, and could be helpful for those using a 2-stage approach 😏
\n
I added this data here, and showed how to visualize examples here. Hope this helps!
", "messages": [ { "id": 2961480, "postDate": "2024-08-16T16:42:15.057Z", "content": "After some positive feedback on the coordinate pretraining dataset, I have decided to release some more data.
\nThis time I made additions to the competition coordinate data by annotating the left side of each disc. This provides orientation information, and could be helpful for those using a 2-stage approach 😏
\n
I added this data here, and showed how to visualize examples here. Hope this helps!
", "rawMarkdown": "After some positive feedback on the coordinate pretraining dataset, I have decided to release some more data. \n\nThis time I made additions to the competition coordinate data by annotating the left side of each disc. This provides orientation information, and could be helpful for those using a 2-stage approach 😏\n\n\n\nI added this data [here](https://www.kaggle.com/datasets/brendanartley/lumbar-coordinate-pretraining-dataset), and showed how to visualize examples [here](https://www.kaggle.com/code/brendanartley/lumbar-coordinate-dataset-code). Hope this helps!\n", "votes": 60 }, { "id": 2961524, "postDate": "2024-08-16T17:20:48.547Z", "content": "wow, it is unselfish. How many hours did you spend annotating it?!
\nBtw, have you solved the problem with sorting order? (Based on the leaderboard score, I guess you did.)
\ntrying to take advantage of your generosity, if it is not a secret, what cv score did you guys get for foraminal and subarticular?
Thanks @sergiosaharovskiy. It only took a few hours or so thanks to the VGG Image Annotator. We sort based on ImagePositionPatient in the same way as Ian Pan here.
\nCV is secret for now :)
", "rawMarkdown": "Thanks @sergiosaharovskiy. It only took a few hours or so thanks to the [VGG Image Annotator](https://www.robots.ox.ac.uk/~vgg/software/via/). We sort based on ImagePositionPatient in the same way as Ian Pan [here](https://www.kaggle.com/code/vaillant/cross-reference-images-in-different-mri-planes?scriptVersionId=182551992&cellId=2). \n\nCV is secret for now :)", "votes": 6, "replies": [ { "id": 2961617, "postDate": "2024-08-16T18:35:53.313Z", "content": "let me guess < 0.4 ? :D
", "rawMarkdown": "let me guess < 0.4 ? :D", "votes": 3 } ] } ] }, { "id": 2967453, "postDate": "2024-08-22T23:46:43.940Z", "content": "great work
", "rawMarkdown": "great work", "votes": 1 }, { "id": 2962712, "postDate": "2024-08-17T23:24:11.653Z", "content": "are you using a two stage approach for your work or one stage?
", "rawMarkdown": "are you using a two stage approach for your work or one stage?", "votes": 1, "replies": [ { "id": 2963646, "postDate": "2024-08-19T00:12:19.327Z", "content": "Two-stage.
", "rawMarkdown": "Two-stage. ", "votes": 2, "replies": [ { "id": 2963958, "postDate": "2024-08-19T12:16:42.003Z", "content": "Sorry, but is it a single model performing the detection, or are there multiple small expert models involved, or is it three models in total?
", "rawMarkdown": "Sorry, but is it a single model performing the detection, or are there multiple small expert models involved, or is it three models in total?" }, { "id": 2964036, "postDate": "2024-08-19T13:20:26.420Z", "content": "Hi @theexaltedone, thanks for the comments. I don’t mind sharing the general idea of our process, but I won’t get into the details. The dataset and the notebook I shared should give you the general idea.
", "rawMarkdown": "Hi @theexaltedone, thanks for the comments. I don’t mind sharing the general idea of our process, but I won’t get into the details. The dataset and the notebook I shared should give you the general idea.", "votes": 4 }, { "id": 2964086, "postDate": "2024-08-19T14:24:04.270Z", "content": "yh, thanks that would help me a lot.
", "rawMarkdown": "yh, thanks that would help me a lot.\n", "votes": 1 } ] } ] }, { "id": 2968045, "postDate": "2024-08-23T13:28:23.880Z", "content": "nice visualisation:
\nwhat did the model learns first?
\nl5/s1 … becuase it is the most imprtant point!
the next point is l4/l3 and then l1/l2
\nlast train iteration
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fbe922c7102cf6437985e7bd067b93d3d%2FSelection_999(5935).png?generation=1724419677302238&alt=media\")
intermediate …
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F7e8d42bc3e6fc4c8384c0c8d64e36552%2FSelection_999(5934).png?generation=1724419689158886&alt=media\")
first train iteration
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F4e07fe236dbd1bf2cbbb51cb4b5744a8%2FSelection_999(5933).png?generation=1724419700834470&alt=media\")
loss function bug
\nmake a bug and end up this
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fc284153118aa7839c6d0ea8c6044200d%2FSelection_999(5948).png?generation=1724487383709747&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F27c5d6ff39c361137368064f0c06f027%2FSelection_999(5947).png?generation=1724487394941357&alt=media\")
loss = F.mse_loss( xy_truth, (heatmap*coord).sum(...)) #this is suprsing learned (the horizontal line)\n\n\n loss = Jensen-Shannon divergence loss (heatmap, gaussian mask of truth xy) \n\n\n\n\n version\n = -. * /(sigma*sigma)\n = -. * /(sigma*sigma)\n version\n = -. * np.reciprocal(sigma*sigma) #sigma is integer=, and np.reciprocal() gives integer zero\n Jensen-Shannon divergence \nshape analysis:
\ncode from:
\nhttps://medium.com/@olga_kravchenko/generalized-procrustes-analysis-with-python-numpy-c571e8e8a421
left to right:
\nmean shape, align to center only, align to the center and scale, align to the center, scale, rotation
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F0f185e8152c32eed4cb622cdffc613ab%2FSelection_999(5924).png?generation=1724334688933737&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2Fe180ee6619b9d52bbb3584f5270fa396%2FSelection_999(5925).png?generation=1724342123517999&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F113660%2F4b58c01e78c49d39f6e92d6dbf43f777%2FSelection_999(5926).png?generation=1724358043209591&alt=media\")
what is the use?
\nwow, it is unselfish. How many hours did you spend annotating it?!
\nBtw, have you solved the problem with sorting order? (Based on the leaderboard score, I guess you did.)
\ntrying to take advantage of your generosity, if it is not a secret, what cv score did you guys get for foraminal and subarticular?
Thanks @sergiosaharovskiy. It only took a few hours or so thanks to the VGG Image Annotator. We sort based on ImagePositionPatient in the same way as Ian Pan here.
\nCV is secret for now :)
", "votes": 6, "replies": [ { "id": 2961617, "author_name": "SSS", "author_url": "", "post_date": "2024-08-16T18:35:53.313000", "content": "let me guess < 0.4 ? :D
", "votes": 3, "replies": [] } ] } ] }, { "id": 2967453, "author_name": "alka sharma", "author_url": "", "post_date": "2024-08-22T23:46:43.940000", "content": "great work
", "votes": 1, "replies": [] }, { "id": 2962712, "author_name": "Exalted Joseph", "author_url": "", "post_date": "2024-08-17T23:24:11.653000", "content": "are you using a two stage approach for your work or one stage?
", "votes": 1, "replies": [ { "id": 2963646, "author_name": "Bartley", "author_url": "", "post_date": "2024-08-19T00:12:19.327000", "content": "Two-stage.
", "votes": 2, "replies": [ { "id": 2963958, "author_name": "Exalted Joseph", "author_url": "", "post_date": "2024-08-19T12:16:42.003000", "content": "Sorry, but is it a single model performing the detection, or are there multiple small expert models involved, or is it three models in total?
", "votes": 0, "replies": [] }, { "id": 2964036, "author_name": "Bartley", "author_url": "", "post_date": "2024-08-19T13:20:26.420000", "content": "Hi @theexaltedone, thanks for the comments. I don’t mind sharing the general idea of our process, but I won’t get into the details. The dataset and the notebook I shared should give you the general idea.
", "votes": 4, "replies": [] }, { "id": 2964086, "author_name": "Exalted Joseph", "author_url": "", "post_date": "2024-08-19T14:24:04.270000", "content": "yh, thanks that would help me a lot.
", "votes": 1, "replies": [] } ] } ] }, { "id": 2968045, "author_name": "hengck23", "author_url": "", "post_date": "2024-08-23T13:28:23.880000", "content": "nice visualisation:
\nwhat did the model learns first?
\nl5/s1 … becuase it is the most imprtant point!
the next point is l4/l3 and then l1/l2
\nlast train iteration
\n
intermediate …
\n
first train iteration
\n
loss function bug
\nmake a bug and end up this
\n
loss = F.mse_loss( xy_truth, (heatmap*coord).sum(...)) #this is suprsing learned (the horizontal line)\n\n\n loss = Jensen-Shannon divergence loss (heatmap, gaussian mask of truth xy) \n\n\n\n\n version\n = -. * /(sigma*sigma)\n = -. * /(sigma*sigma)\n version\n = -. * np.reciprocal(sigma*sigma) #sigma is integer=, and np.reciprocal() gives integer zero\n Jensen-Shannon divergence \nshape analysis:
\ncode from:
\nhttps://medium.com/@olga_kravchenko/generalized-procrustes-analysis-with-python-numpy-c571e8e8a421
left to right:
\nmean shape, align to center only, align to the center and scale, align to the center, scale, rotation
what is the use?
\n