Hi everyone,
\nCongratulations to all the teams for your hard work tackling this novel and difficult challenge!
\nWe can't wait to find out more about what you all tried, what worked, and what didn't.
We put together the following (short) form to collect some general feedback on the challenge and to learn more about the methods tested.
\nWe invite you to fill it as it will give us valuable information and feedback on the relative performance of the different approaches.
\nHere is the link: https://forms.gle/AvR35rG8E3oyUuE79
We would also like to encourage all the teams to respond to this discussion thread to tell us what you tried, e.g., what worked or what didn't work.
\nFrom our past experience, sharing this information has been really valuable and insightful.
Finally, we recall that participants are highly encouraged to submit a working notes paper to CLEF, the deadline being extended to June 01.
\nMore information is provided in the following thread:
\nhttps://www.kaggle.com/competitions/geolifeclef-2022-lifeclef-2022-fgvc9/discussion/325984
All the best,
\nGeoLifeCLEF 2022 Challenge Organizers
Hi everyone,
\nCongratulations to all the teams for your hard work tackling this novel and difficult challenge!
\nWe can't wait to find out more about what you all tried, what worked, and what didn't.
We put together the following (short) form to collect some general feedback on the challenge and to learn more about the methods tested.
\nWe invite you to fill it as it will give us valuable information and feedback on the relative performance of the different approaches.
\nHere is the link: https://forms.gle/AvR35rG8E3oyUuE79
We would also like to encourage all the teams to respond to this discussion thread to tell us what you tried, e.g., what worked or what didn't work.
\nFrom our past experience, sharing this information has been really valuable and insightful.
Finally, we recall that participants are highly encouraged to submit a working notes paper to CLEF, the deadline being extended to June 01.
\nMore information is provided in the following thread:
\nhttps://www.kaggle.com/competitions/geolifeclef-2022-lifeclef-2022-fgvc9/discussion/325984
All the best,
\nGeoLifeCLEF 2022 Challenge Organizers
Hi, I am the third place in the competition, whether the winner has the award certificate?
", "rawMarkdown": "Hi, I am the third place in the competition, whether the winner has the award certificate?", "votes": null }, { "id": "1804004", "postDate": "05/28/2022 12:27:20", "content": "As I currently work on a tabular + goespatial problem, I was mainly interested in simple baseline (relying mostly on lat, long) and learning the basics of pytorch/computer vision. I got a decent knn baseline on lat / long only (still below RF but only 2%). My interpretation is that lot of bio features correlates with lat long locally.
\nI tried many things but nothing really beat a vanilla L2 knn with top_30 voting on the X neighbors (where X=1200 seems optimal). Other distances didn't work (L2 is a good approx of haversine, which is too slow), other features did't work (maybe altitude with a bit of scaling for 0.01%), other 'voting' rules didn't work (30 nearest target for exemple, weighting by different powers of distance, or some measure of local density). I tried to look for knn imrpovements: knn++ doesn't seems really related… kernel methods seemed interesting but I couln't find any easy to use implementation in python.
\nSo I end up with very vanilla sklearn knn that don't beat the RF baseline. I am not entirely convinced this is worth a working note… tell me and I'll write something early next week.
", "rawMarkdown": "As I currently work on a tabular + goespatial problem, I was mainly interested in simple baseline (relying mostly on lat, long) and learning the basics of pytorch/computer vision. I got a decent knn baseline on lat / long only (still below RF but only 2%). My interpretation is that lot of bio features correlates with lat long locally. \n\nI tried many things but nothing really beat a vanilla L2 knn with top_30 voting on the X neighbors (where X=1200 seems optimal). Other distances didn't work (L2 is a good approx of haversine, which is too slow), other features did't work (maybe altitude with a bit of scaling for 0.01%), other 'voting' rules didn't work (30 nearest target for exemple, weighting by different powers of distance, or some measure of local density). I tried to look for knn imrpovements: knn++ doesn't seems really related... kernel methods seemed interesting but I couln't find any easy to use implementation in python.\n\nSo I end up with very vanilla sklearn knn that don't beat the RF baseline. I am not entirely convinced this is worth a working note... tell me and I'll write something early next week.", "votes": null }, { "id": "1805864", "postDate": "05/30/2022 15:07:05", "content": "Hi Zhangxiaojuan, congrats for your third place! 😃
\nSorry, I'm not sure to understand your question. Are you referring to a certificate from Kaggle or a certificate provided by the organizers?
Hi Lucas, thanks a lot for your feedback!
\nInteresting to see that vanilla L2 kNN on coordinates works better than other more complicated kNN approaches.
\nI did manage to run a kNN on coordinates using proper Haversine distance (similarly K=1000 seemed good), it indeed gave very close results than your results using L2 distance (~79.9% top-30 error on the validation set).
\nI don't think this is indeed worth a working notes, it will take you a lot of time and effort.
\nThanks again for your participation, I hope you learned a few things! 😃
Thanks for the organisation. Yes I learned to quickstart a vision model on relatively high classification problem thanks to the shared baseline (maybe it would be a good idea to share one of the cnn baseline for new people). I think I might continue to look for relevant knn improvement and report next year :)
", "rawMarkdown": "Thanks for the organisation. Yes I learned to quickstart a vision model on relatively high classification problem thanks to the shared baseline (maybe it would be a good idea to share one of the cnn baseline for new people). I think I might continue to look for relevant knn improvement and report next year :)", "votes": null }, { "id": "1812807", "postDate": "06/06/2022 08:32:39", "content": "I referring to a certificate provided by the organizers, used to prove to my school that I participated in the GeoLifeCLEF 2022 Challenge and won third place
", "rawMarkdown": "I referring to a certificate provided by the organizers, used to prove to my school that I participated in the GeoLifeCLEF 2022 Challenge and won third place", "votes": null } ], "comments": [ { "id": 1801879, "author_name": "zhangxiaojuan", "author_url": "", "post_date": "05/26/2022 08:27:01", "content": "Hi, I am the third place in the competition, whether the winner has the award certificate?
", "votes": null, "replies": [ { "id": 1805864, "author_name": "tlorieul", "author_url": "", "post_date": "05/30/2022 15:07:05", "content": "Hi Zhangxiaojuan, congrats for your third place! 😃
\nSorry, I'm not sure to understand your question. Are you referring to a certificate from Kaggle or a certificate provided by the organizers?
I referring to a certificate provided by the organizers, used to prove to my school that I participated in the GeoLifeCLEF 2022 Challenge and won third place
", "votes": null, "replies": [] } ] }, { "id": 1804004, "author_name": "lucasmorin", "author_url": "", "post_date": "05/28/2022 12:27:20", "content": "As I currently work on a tabular + goespatial problem, I was mainly interested in simple baseline (relying mostly on lat, long) and learning the basics of pytorch/computer vision. I got a decent knn baseline on lat / long only (still below RF but only 2%). My interpretation is that lot of bio features correlates with lat long locally.
\nI tried many things but nothing really beat a vanilla L2 knn with top_30 voting on the X neighbors (where X=1200 seems optimal). Other distances didn't work (L2 is a good approx of haversine, which is too slow), other features did't work (maybe altitude with a bit of scaling for 0.01%), other 'voting' rules didn't work (30 nearest target for exemple, weighting by different powers of distance, or some measure of local density). I tried to look for knn imrpovements: knn++ doesn't seems really related… kernel methods seemed interesting but I couln't find any easy to use implementation in python.
\nSo I end up with very vanilla sklearn knn that don't beat the RF baseline. I am not entirely convinced this is worth a working note… tell me and I'll write something early next week.
", "votes": null, "replies": [ { "id": 1805871, "author_name": "tlorieul", "author_url": "", "post_date": "05/30/2022 15:16:03", "content": "Hi Lucas, thanks a lot for your feedback!
\nInteresting to see that vanilla L2 kNN on coordinates works better than other more complicated kNN approaches.
\nI did manage to run a kNN on coordinates using proper Haversine distance (similarly K=1000 seemed good), it indeed gave very close results than your results using L2 distance (~79.9% top-30 error on the validation set).
\nI don't think this is indeed worth a working notes, it will take you a lot of time and effort.
\nThanks again for your participation, I hope you learned a few things! 😃
Thanks for the organisation. Yes I learned to quickstart a vision model on relatively high classification problem thanks to the shared baseline (maybe it would be a good idea to share one of the cnn baseline for new people). I think I might continue to look for relevant knn improvement and report next year :)
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1800932": "Hi everyone,\n\nCongratulations to all the teams for your hard work tackling this novel and difficult challenge!\nWe can't wait to find out more about what you all tried, what worked, and what didn't.\n\nWe put together the **following (short) form to collect some general feedback on the challenge and to learn more about the methods tested.**\nWe invite you to fill it as it will give us valuable information and feedback on the relative performance of the different approaches.\nHere is the link: https://forms.gle/AvR35rG8E3oyUuE79\n\nWe would also like to encourage all the teams to respond to this discussion thread to tell us what you tried, e.g., what worked or what didn't work.\nFrom our past experience, sharing this information has been really valuable and insightful.\n\nFinally, we recall that participants are highly encouraged to **submit a working notes paper to CLEF, the deadline being extended to June 01.**\nMore information is provided in the following thread:\nhttps://www.kaggle.com/competitions/geolifeclef-2022-lifeclef-2022-fgvc9/discussion/325984\n\nAll the best,\nGeoLifeCLEF 2022 Challenge Organizers", "1801879": "Hi, I am the third place in the competition, whether the winner has the award certificate?", "1804004": "As I currently work on a tabular + goespatial problem, I was mainly interested in simple baseline (relying mostly on lat, long) and learning the basics of pytorch/computer vision. I got a decent knn baseline on lat / long only (still below RF but only 2%). My interpretation is that lot of bio features correlates with lat long locally. \n\nI tried many things but nothing really beat a vanilla L2 knn with top_30 voting on the X neighbors (where X=1200 seems optimal). Other distances didn't work (L2 is a good approx of haversine, which is too slow), other features did't work (maybe altitude with a bit of scaling for 0.01%), other 'voting' rules didn't work (30 nearest target for exemple, weighting by different powers of distance, or some measure of local density). I tried to look for knn imrpovements: knn++ doesn't seems really related... kernel methods seemed interesting but I couln't find any easy to use implementation in python.\n\nSo I end up with very vanilla sklearn knn that don't beat the RF baseline. I am not entirely convinced this is worth a working note... tell me and I'll write something early next week.", "1805864": "Hi Zhangxiaojuan, congrats for your third place! 😃\nSorry, I'm not sure to understand your question. Are you referring to a certificate from Kaggle or a certificate provided by the organizers?", "1805871": "Hi Lucas, thanks a lot for your feedback!\nInteresting to see that vanilla L2 kNN on coordinates works better than other more complicated kNN approaches.\nI did manage to run a kNN on coordinates using proper Haversine distance (similarly K=1000 seemed good), it indeed gave very close results than your results using L2 distance (~79.9% top-30 error on the validation set).\nI don't think this is indeed worth a working notes, it will take you a lot of time and effort.\nThanks again for your participation, I hope you learned a few things! 😃", "1806094": "Thanks for the organisation. Yes I learned to quickstart a vision model on relatively high classification problem thanks to the shared baseline (maybe it would be a good idea to share one of the cnn baseline for new people). I think I might continue to look for relevant knn improvement and report next year :)", "1812807": "I referring to a certificate provided by the organizers, used to prove to my school that I participated in the GeoLifeCLEF 2022 Challenge and won third place" }, "source": "meta" }