As this competition comes to a close I wanted to take a moment to reflect on all that I have learned, perhaps you would be willing to share what you learned too!
\nThis one was an interesting one for me as I have not taken part in the many medical image segmentation comps on Kaggle. I am really pleased with just how much I was able to pick up!
\n2 **Timm Models: ** I have used Timm models before but there are so many that it is tough to know all of them! I got to get a few good results with familiar models which is nice.
\n**Practice with Unet Architecture: ** Unets are incredibly powerful for most image tasks, yet I have not had many opportunities to work with them. This was a great opportunity to test it out and I was really impressed with the results!
**Working with 3D Data: ** This is the first time I have gotten to work with 3D data and it was a great time. I was really glad to see that there were so many effective ways to tackle this competition. I had some success with 2D, 2.5D, and 3D models.
I always try to do this at the end of competition because I always find myself chasing a medal and it can be easy to forget about the real purpose that brings us to Kaggle. Though the medals can be a reflection of learning they are not the only one. I am really happy with the amount I have learned in only 3 months so I can say I reached my goal on this one! I hope you all can as well! Please share some of your learnings below! Good luck on the private LB all!
", "messages": [ { "id": "2636205", "postDate": "02/04/2024 22:09:50", "content": "As this competition comes to a close I wanted to take a moment to reflect on all that I have learned, perhaps you would be willing to share what you learned too!
\nThis one was an interesting one for me as I have not taken part in the many medical image segmentation comps on Kaggle. I am really pleased with just how much I was able to pick up!
\n2 **Timm Models: ** I have used Timm models before but there are so many that it is tough to know all of them! I got to get a few good results with familiar models which is nice.
\n**Practice with Unet Architecture: ** Unets are incredibly powerful for most image tasks, yet I have not had many opportunities to work with them. This was a great opportunity to test it out and I was really impressed with the results!
**Working with 3D Data: ** This is the first time I have gotten to work with 3D data and it was a great time. I was really glad to see that there were so many effective ways to tackle this competition. I had some success with 2D, 2.5D, and 3D models.
I always try to do this at the end of competition because I always find myself chasing a medal and it can be easy to forget about the real purpose that brings us to Kaggle. Though the medals can be a reflection of learning they are not the only one. I am really happy with the amount I have learned in only 3 months so I can say I reached my goal on this one! I hope you all can as well! Please share some of your learnings below! Good luck on the private LB all!
", "rawMarkdown": "As this competition comes to a close I wanted to take a moment to reflect on all that I have learned, perhaps you would be willing to share what you learned too! \n\nThis one was an interesting one for me as I have not taken part in the many medical image segmentation comps on Kaggle. I am really pleased with just how much I was able to pick up! \n\n1. **Albumentations: ** I learned a lot about image augmentations which was essential to high quality models in this competition. I haven’t used this package before but liked it much more than the previous packages I have used for image work.\n\n2 **Timm Models: ** I have used Timm models before but there are so many that it is tough to know all of them! I got to get a few good results with familiar models which is nice. \n\n3. **Practice with Unet Architecture: ** Unets are incredibly powerful for most image tasks, yet I have not had many opportunities to work with them. This was a great opportunity to test it out and I was really impressed with the results!\n\n4. **Working with 3D Data: ** This is the first time I have gotten to work with 3D data and it was a great time. I was really glad to see that there were so many effective ways to tackle this competition. I had some success with 2D, 2.5D, and 3D models. \n\nI always try to do this at the end of competition because I always find myself chasing a medal and it can be easy to forget about the real purpose that brings us to Kaggle. Though the medals can be a reflection of learning they are not the only one. I am really happy with the amount I have learned in only 3 months so I can say I reached my goal on this one! I hope you all can as well! Please share some of your learnings below! Good luck on the private LB all!", "votes": null }, { "id": "2636209", "postDate": "02/04/2024 22:23:04", "content": "Timm is new for me, so I already learned something from this post! I'll have to check it out. I don't use Pytorch much as I learned Keras/tensorflow first, but Pytorch seems really popular here. It also seems like things like albumentations work a little more seamlessly with it. There are a lot more augmentations available than the basic keras preprocessing layers. Several folks have suggested augmentations have been key to good models for this competition. While I'm happy with the models I have been able to make in this comp, I do think something is missing and perhaps it was the augmentations.
", "rawMarkdown": "Timm is new for me, so I already learned something from this post! I'll have to check it out. I don't use Pytorch much as I learned Keras/tensorflow first, but Pytorch seems really popular here. It also seems like things like albumentations work a little more seamlessly with it. There are a lot more augmentations available than the basic keras preprocessing layers. Several folks have suggested augmentations have been key to good models for this competition. While I'm happy with the models I have been able to make in this comp, I do think something is missing and perhaps it was the augmentations.", "votes": null }, { "id": "2636213", "postDate": "02/04/2024 22:27:30", "content": "Yes PyTorch is just a bit more customizable so it normally has a higher upper bound with score! Augmentations are very useful for generalizing your models more!
", "rawMarkdown": "Yes PyTorch is just a bit more customizable so it normally has a higher upper bound with score! Augmentations are very useful for generalizing your models more!", "votes": null }, { "id": "2636214", "postDate": "02/04/2024 22:28:47", "content": "Good to know! Perhaps next competition I'll spend some time learning PyTorch and the libraries around it. Do you find the DataLoaders are of comparable efficiency to tf Datasets?
", "rawMarkdown": "Good to know! Perhaps next competition I'll spend some time learning PyTorch and the libraries around it. Do you find the DataLoaders are of comparable efficiency to tf Datasets?", "votes": null }, { "id": "2636306", "postDate": "02/05/2024 01:24:36", "content": "I would say it is fairly similar but I am sure that is more dependent on what you are doing and opinions
", "rawMarkdown": "I would say it is fairly similar but I am sure that is more dependent on what you are doing and opinions", "votes": null }, { "id": "2639173", "postDate": "02/06/2024 17:54:33", "content": "Hey! How do I submit a notebook? And what if my submission format is wrong? I am new.
", "rawMarkdown": "Hey! How do I submit a notebook? And what if my submission format is wrong? I am new.", "votes": null }, { "id": "2639197", "postDate": "02/06/2024 18:08:56", "content": "This is a hard competition to start on but if you are already in then I would start with the public notebooks and just try to understand what they are doing, this comp is not easy tho!
", "rawMarkdown": "This is a hard competition to start on but if you are already in then I would start with the public notebooks and just try to understand what they are doing, this comp is not easy tho!", "votes": null }, { "id": "2640421", "postDate": "02/06/2024 23:22:11", "content": "Ohh thnx I'll try that. How do I know which competitions are easy and which ones are hard?
", "rawMarkdown": "Ohh thnx I'll try that. How do I know which competitions are easy and which ones are hard?", "votes": null }, { "id": "2645158", "postDate": "02/10/2024 02:05:39", "content": "This competition is my first try with Unet and 3d data, previously I only have heard about Unet. In this competition I learned a useful tool was pyvista, I use it display the output and label, my model doesn't work really good, It just hit the inner side of the thick vessel exclude the thin vessel on my valid data, it got a bad score on LB, I did some bad process on the preprocessing stage . Thanks for sharing, I would try my best to study the good program and idea！
", "rawMarkdown": "This competition is my first try with Unet and 3d data, previously I only have heard about Unet. In this competition I learned a useful tool was pyvista, I use it display the output and label, my model doesn't work really good, It just hit the inner side of the thick vessel exclude the thin vessel on my valid data, it got a bad score on LB, I did some bad process on the preprocessing stage . Thanks for sharing, I would try my best to study the good program and idea！", "votes": null }, { "id": "2645176", "postDate": "02/10/2024 03:00:17", "content": "I mean honestly top 35% is not bad at all for the first time and I am sure you learned a lot!
", "rawMarkdown": "I mean honestly top 35% is not bad at all for the first time and I am sure you learned a lot!", "votes": null }, { "id": "2645235", "postDate": "02/10/2024 04:58:04", "content": "nono the score was the public notebook's, I forgot to select the final submission at the end of the competition, my submission have just 0.01 on the test set 😱
", "rawMarkdown": "nono the score was the public notebook's, I forgot to select the final submission at the end of the competition, my submission have just 0.01 on the test set 😱", "votes": null }, { "id": "2645242", "postDate": "02/10/2024 05:07:07", "content": "that's my bad😰
", "rawMarkdown": "that's my bad😰", "votes": null }, { "id": "2646492", "postDate": "02/11/2024 01:00:25", "content": "I love that you did a reflection on this competition, and I'll start implementing that routine once I get going with competitions! I've been in the data science space awhile, but I haven't used kaggle as fully as I would like. I'm getting my feet wet with competitions and so far I really like what I see. I've never worked with 3D data. Any insights you'd like to share?
", "rawMarkdown": "I love that you did a reflection on this competition, and I'll start implementing that routine once I get going with competitions! I've been in the data science space awhile, but I haven't used kaggle as fully as I would like. I'm getting my feet wet with competitions and so far I really like what I see. I've never worked with 3D data. Any insights you'd like to share?", "votes": null } ], "comments": [ { "id": 2636209, "author_name": "chemdatafarmer", "author_url": "", "post_date": "02/04/2024 22:23:04", "content": "Timm is new for me, so I already learned something from this post! I'll have to check it out. I don't use Pytorch much as I learned Keras/tensorflow first, but Pytorch seems really popular here. It also seems like things like albumentations work a little more seamlessly with it. There are a lot more augmentations available than the basic keras preprocessing layers. Several folks have suggested augmentations have been key to good models for this competition. While I'm happy with the models I have been able to make in this comp, I do think something is missing and perhaps it was the augmentations.
", "votes": null, "replies": [ { "id": 2636213, "author_name": "cody11null", "author_url": "", "post_date": "02/04/2024 22:27:30", "content": "Yes PyTorch is just a bit more customizable so it normally has a higher upper bound with score! Augmentations are very useful for generalizing your models more!
", "votes": null, "replies": [ { "id": 2636214, "author_name": "chemdatafarmer", "author_url": "", "post_date": "02/04/2024 22:28:47", "content": "Good to know! Perhaps next competition I'll spend some time learning PyTorch and the libraries around it. Do you find the DataLoaders are of comparable efficiency to tf Datasets?
", "votes": null, "replies": [ { "id": 2636306, "author_name": "cody11null", "author_url": "", "post_date": "02/05/2024 01:24:36", "content": "I would say it is fairly similar but I am sure that is more dependent on what you are doing and opinions
", "votes": null, "replies": [ { "id": 2639173, "author_name": "prabal0221", "author_url": "", "post_date": "02/06/2024 17:54:33", "content": "Hey! How do I submit a notebook? And what if my submission format is wrong? I am new.
", "votes": null, "replies": [ { "id": 2639197, "author_name": "cody11null", "author_url": "", "post_date": "02/06/2024 18:08:56", "content": "This is a hard competition to start on but if you are already in then I would start with the public notebooks and just try to understand what they are doing, this comp is not easy tho!
", "votes": null, "replies": [ { "id": 2640421, "author_name": "prabal0221", "author_url": "", "post_date": "02/06/2024 23:22:11", "content": "Ohh thnx I'll try that. How do I know which competitions are easy and which ones are hard?
", "votes": null, "replies": [] } ] } ] } ] } ] } ] } ] }, { "id": 2645158, "author_name": "athrunzala", "author_url": "", "post_date": "02/10/2024 02:05:39", "content": "This competition is my first try with Unet and 3d data, previously I only have heard about Unet. In this competition I learned a useful tool was pyvista, I use it display the output and label, my model doesn't work really good, It just hit the inner side of the thick vessel exclude the thin vessel on my valid data, it got a bad score on LB, I did some bad process on the preprocessing stage . Thanks for sharing, I would try my best to study the good program and idea！
", "votes": null, "replies": [ { "id": 2645176, "author_name": "cody11null", "author_url": "", "post_date": "02/10/2024 03:00:17", "content": "I mean honestly top 35% is not bad at all for the first time and I am sure you learned a lot!
", "votes": null, "replies": [ { "id": 2645235, "author_name": "athrunzala", "author_url": "", "post_date": "02/10/2024 04:58:04", "content": "nono the score was the public notebook's, I forgot to select the final submission at the end of the competition, my submission have just 0.01 on the test set 😱
", "votes": null, "replies": [ { "id": 2645242, "author_name": "athrunzala", "author_url": "", "post_date": "02/10/2024 05:07:07", "content": "that's my bad😰
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 2646492, "author_name": "milesmena3", "author_url": "", "post_date": "02/11/2024 01:00:25", "content": "I love that you did a reflection on this competition, and I'll start implementing that routine once I get going with competitions! I've been in the data science space awhile, but I haven't used kaggle as fully as I would like. I'm getting my feet wet with competitions and so far I really like what I see. I've never worked with 3D data. Any insights you'd like to share?
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2636205": "As this competition comes to a close I wanted to take a moment to reflect on all that I have learned, perhaps you would be willing to share what you learned too! \n\nThis one was an interesting one for me as I have not taken part in the many medical image segmentation comps on Kaggle. I am really pleased with just how much I was able to pick up! \n\n1. **Albumentations: ** I learned a lot about image augmentations which was essential to high quality models in this competition. I haven’t used this package before but liked it much more than the previous packages I have used for image work.\n\n2 **Timm Models: ** I have used Timm models before but there are so many that it is tough to know all of them! I got to get a few good results with familiar models which is nice. \n\n3. **Practice with Unet Architecture: ** Unets are incredibly powerful for most image tasks, yet I have not had many opportunities to work with them. This was a great opportunity to test it out and I was really impressed with the results!\n\n4. **Working with 3D Data: ** This is the first time I have gotten to work with 3D data and it was a great time. I was really glad to see that there were so many effective ways to tackle this competition. I had some success with 2D, 2.5D, and 3D models. \n\nI always try to do this at the end of competition because I always find myself chasing a medal and it can be easy to forget about the real purpose that brings us to Kaggle. Though the medals can be a reflection of learning they are not the only one. I am really happy with the amount I have learned in only 3 months so I can say I reached my goal on this one! I hope you all can as well! Please share some of your learnings below! Good luck on the private LB all!", "2636209": "Timm is new for me, so I already learned something from this post! I'll have to check it out. I don't use Pytorch much as I learned Keras/tensorflow first, but Pytorch seems really popular here. It also seems like things like albumentations work a little more seamlessly with it. There are a lot more augmentations available than the basic keras preprocessing layers. Several folks have suggested augmentations have been key to good models for this competition. While I'm happy with the models I have been able to make in this comp, I do think something is missing and perhaps it was the augmentations.", "2636213": "Yes PyTorch is just a bit more customizable so it normally has a higher upper bound with score! Augmentations are very useful for generalizing your models more!", "2636214": "Good to know! Perhaps next competition I'll spend some time learning PyTorch and the libraries around it. Do you find the DataLoaders are of comparable efficiency to tf Datasets?", "2636306": "I would say it is fairly similar but I am sure that is more dependent on what you are doing and opinions", "2639173": "Hey! How do I submit a notebook? And what if my submission format is wrong? I am new.", "2639197": "This is a hard competition to start on but if you are already in then I would start with the public notebooks and just try to understand what they are doing, this comp is not easy tho!", "2640421": "Ohh thnx I'll try that. How do I know which competitions are easy and which ones are hard?", "2645158": "This competition is my first try with Unet and 3d data, previously I only have heard about Unet. In this competition I learned a useful tool was pyvista, I use it display the output and label, my model doesn't work really good, It just hit the inner side of the thick vessel exclude the thin vessel on my valid data, it got a bad score on LB, I did some bad process on the preprocessing stage . Thanks for sharing, I would try my best to study the good program and idea！", "2645176": "I mean honestly top 35% is not bad at all for the first time and I am sure you learned a lot!", "2645235": "nono the score was the public notebook's, I forgot to select the final submission at the end of the competition, my submission have just 0.01 on the test set 😱", "2645242": "that's my bad😰", "2646492": "I love that you did a reflection on this competition, and I'll start implementing that routine once I get going with competitions! I've been in the data science space awhile, but I haven't used kaggle as fully as I would like. I'm getting my feet wet with competitions and so far I really like what I see. I've never worked with 3D data. Any insights you'd like to share?" }, "source": "meta" }