At the beginning of this competition, I was outraged to see how low my score on the LB was compared to the headliners. But, in fact, it wasn't! My score was correct! I didn't realize it until I submitted a simple average of my 2 best models: bagging works extremely well in this competition. As an example, I went from a score around 0.63 to 0.73.
\nIf I share, it's not to taunt people but rather to encourage those like me who are wondering how to break the 0.70 worm !
", "messages": [ { "id": "1277942", "postDate": "04/19/2021 11:50:36", "content": "At the beginning of this competition, I was outraged to see how low my score on the LB was compared to the headliners. But, in fact, it wasn't! My score was correct! I didn't realize it until I submitted a simple average of my 2 best models: bagging works extremely well in this competition. As an example, I went from a score around 0.63 to 0.73.
\nIf I share, it's not to taunt people but rather to encourage those like me who are wondering how to break the 0.70 worm !
", "rawMarkdown": "At the beginning of this competition, I was outraged to see how low my score on the LB was compared to the headliners. But, in fact, it wasn't! My score was correct! I didn't realize it until I submitted a simple average of my 2 best models: bagging works extremely well in this competition. As an example, I went from a score around **0.63** to **0.73**.\n\nIf I share, it's not to taunt people but rather to encourage those like me who are wondering how to break the **0.70** worm !", "votes": null }, { "id": "1278278", "postDate": "04/19/2021 17:56:43", "content": "do you apply bagging on score or at the end to the labels
", "rawMarkdown": "do you apply bagging on score or at the end to the labels", "votes": null }, { "id": "1278305", "postDate": "04/19/2021 18:14:44", "content": "I can see that details are very important to you … It was on raw scores.
", "rawMarkdown": "I can see that details are very important to you ... It was on raw scores.", "votes": null }, { "id": "1278319", "postDate": "04/19/2021 18:38:05", "content": "Thanks for sharing @kneroma !! I suppose you mean ensemble of 2 full kfold models
", "rawMarkdown": "Thanks for sharing @kneroma !! I suppose you mean ensemble of 2 full kfold models", "votes": null }, { "id": "1278330", "postDate": "04/19/2021 18:49:27", "content": "Thanks @imeintanis. Yes, 2 five kfold models.
", "rawMarkdown": "Thanks @imeintanis. Yes, 2 five kfold models.", "votes": null }, { "id": "1278343", "postDate": "04/19/2021 19:14:05", "content": "Thank you @kneroma
", "rawMarkdown": "Thank you @kneroma", "votes": null }, { "id": "1278355", "postDate": "04/19/2021 19:36:48", "content": "Good to see you sharing that much.
\nI was burned in previous comp, but I'll return to my level of sharing if this one ends well for me.
", "rawMarkdown": "Good to see you sharing that much.\n\nI was burned in previous comp, but I'll return to my level of sharing if this one ends well for me.", "votes": null }, { "id": "1278423", "postDate": "04/19/2021 21:45:29", "content": "Thanks @cpmpml . No matter if you share during this one or not, you still for sure one the best Kaggle voices !
", "rawMarkdown": "Thanks @cpmpml . No matter if you share during this one or not, you still for sure one the best Kaggle voices !", "votes": null }, { "id": "1279575", "postDate": "04/21/2021 04:24:17", "content": "Thank you for sharing.
", "rawMarkdown": "Thank you for sharing.", "votes": null }, { "id": "1279656", "postDate": "04/21/2021 06:13:30", "content": "Thank you as well @peigangshi
", "rawMarkdown": "Thank you as well @peigangshi", "votes": null }, { "id": "1279719", "postDate": "04/21/2021 07:26:48", "content": "for me a single model, without kfold - 0.72
\nensemble of 5 - 0.74
A single model 5 folds LB 0.75
", "rawMarkdown": "A single model 5 folds LB 0.75", "votes": null }, { "id": "1279761", "postDate": "04/21/2021 08:17:07", "content": "Nice one ! It means that there still room for improvements on the LB … Can't to wait to see how far we can go.
", "rawMarkdown": "Nice one ! It means that there still room for improvements on the LB ... Can't to wait to see how far we can go.", "votes": null }, { "id": "1279820", "postDate": "04/21/2021 08:50:30", "content": "I hope 0.74 is the limit ;)
", "rawMarkdown": "I ~~believe~~ hope 0.74 is the limit ;)", "votes": null }, { "id": "1279832", "postDate": "04/21/2021 09:00:58", "content": "How can I believe you @cpmpml while you already scored 0.75 😄 ?
", "rawMarkdown": "How can I believe you @cpmpml while you already scored **0.75** 😄 ?", "votes": null }, { "id": "1279840", "postDate": "04/21/2021 09:07:36", "content": "Thanks for sharing this.
", "rawMarkdown": "Thanks for sharing this.", "votes": null }, { "id": "1279841", "postDate": "04/21/2021 09:09:32", "content": "Thanks @vishalgrover5
", "rawMarkdown": "Thanks @vishalgrover5", "votes": null }, { "id": "1279850", "postDate": "04/21/2021 09:21:35", "content": "I was kidding ofc.
\nBut I wouldn't mind if no one else reaches 0.75 ;)
\nMaybe I was just extremely lucky. We will know at the end of competition.
", "rawMarkdown": "I was kidding ofc.\n\nBut I wouldn't mind if no one else reaches 0.75 ;)\n\nMaybe I was just extremely lucky. We will know at the end of competition.", "votes": null }, { "id": "1279892", "postDate": "04/21/2021 10:25:28", "content": "Oh my bad ! We'll see …
", "rawMarkdown": "Oh my bad ! We'll see ...", "votes": null }, { "id": "1280253", "postDate": "04/21/2021 18:36:51", "content": "@cpmpm l felt sorry for you! Even though I didn't nail it, I learned a lot, because that one comment from you kept me going. Thank you for that!
", "rawMarkdown": "cpmpm l felt sorry for you! Even though I didn't nail it, I learned a lot, because that one comment from you kept me going. Thank you for that!", "votes": null }, { "id": "1280267", "postDate": "04/21/2021 19:09:12", "content": "@nofreewill thanks! I think it helped many teams, but very few acknowledged it.
", "rawMarkdown": "nofreewill thanks! I think it helped many teams, but very few acknowledged it.", "votes": null }, { "id": "1280747", "postDate": "04/22/2021 09:47:54", "content": "done. The road to 0.77
", "rawMarkdown": "done. The road to 0.77", "votes": null }, { "id": "1280773", "postDate": "04/22/2021 10:28:11", "content": "Congratulations :) @vlomme
", "rawMarkdown": "Congratulations :) @vlomme", "votes": null }, { "id": "1280863", "postDate": "04/22/2021 12:39:52", "content": "Well done @vlomme
\nIt's time to resume submitting it seems. I hope my local CV was right for last two weeks…
", "rawMarkdown": "Well done @vlomme \n\nIt's time to resume submitting it seems. I hope my local CV was right for last two weeks...", "votes": null }, { "id": "1283410", "postDate": "04/24/2021 22:30:33", "content": "Do you mean with bagging a \"normal\" ensemble or the bagging (\"bootstrap aggregating\") method by training a set amount of random samples for each classifier?
", "rawMarkdown": "Do you mean with bagging a \"normal\" ensemble or the bagging (\"bootstrap aggregating\") method by training a set amount of random samples for each classifier?", "votes": null }, { "id": "1284410", "postDate": "04/25/2021 22:07:33", "content": "Coming soon.. boosting 0.80 best score …..be afraid?, ..it's a joke, I'm trying my best, my first time competing, but ensambles is the key… i wish u all the best!
", "rawMarkdown": "Coming soon.. boosting 0.80 best score .....be afraid?, ..it's a joke, I'm trying my best, my first time competing, but ensambles is the key... i wish u all the best!", "votes": null }, { "id": "1284592", "postDate": "04/26/2021 05:55:41", "content": "Show us!
", "rawMarkdown": "Show us!", "votes": null }, { "id": "1287296", "postDate": "04/28/2021 22:13:53", "content": "\n\ndone. The road to 0.77
\n
@vlomme you made it. Maybe time to stop and enjoy?
\nKidding of course, you keep me awake at night.
", "rawMarkdown": "> done. The road to 0.77\n\n@vlomme you made it. Maybe time to stop and enjoy?\n\n\nKidding of course, you keep me awake at night.", "votes": null }, { "id": "1287476", "postDate": "04/29/2021 05:51:37", "content": "@cpmpml Your result motivates me to keep working)
\nBut it takes a long time with gtx1060(
Maybe it's time to invest in a better card.
", "rawMarkdown": "Maybe it's time to invest in a better card.", "votes": null }, { "id": "1287548", "postDate": "04/29/2021 07:08:54", "content": "Absolutely, but not $4,000 for 3090
", "rawMarkdown": "Absolutely, but not $4,000 for 3090", "votes": null }, { "id": "1287580", "postDate": "04/29/2021 07:41:08", "content": "I did that for BMS xd
", "rawMarkdown": "I did that for BMS xd", "votes": null }, { "id": "1288268", "postDate": "04/29/2021 19:53:58", "content": "Nice to hear that! Thanks, @kneroma !
", "rawMarkdown": "Nice to hear that! Thanks, @kneroma !", "votes": null }, { "id": "1288465", "postDate": "04/30/2021 03:26:10", "content": "\n\ndone. The road to 0.77
\n
@vladimirsydor well done too, you also keep me awake.
\nAnd I am sure several others will soon make it to 0.77, then pass my current score.
\nWho will be first to 0.80?
", "rawMarkdown": "> done. The road to 0.77\n\n@vladimirsydor well done too, you also keep me awake.\n\nAnd I am sure several others will soon make it to 0.77, then pass my current score.\n\nWho will be first to 0.80?", "votes": null }, { "id": "1288548", "postDate": "04/30/2021 05:35:56", "content": "@brunosan2018 , is your 0.80 ready ?
", "rawMarkdown": "brunosan2018 , is your 0.80 ready ?", "votes": null }, { "id": "1288934", "postDate": "04/30/2021 13:38:05", "content": "Thank you as well dear @doanquanvietnamca
", "rawMarkdown": "Thank you as well dear @doanquanvietnamca", "votes": null }, { "id": "1289040", "postDate": "04/30/2021 15:25:58", "content": "Hello, I have a question, how many kinds of birds did you use to train? for 30 classes I have an f1 0.65 to 0.73 … 30 classes okay? please help me :( need a hands
", "rawMarkdown": "Hello, I have a question, how many kinds of birds did you use to train? for 30 classes I have an f1 0.65 to 0.73 ... 30 classes okay? please help me :( need a hands", "votes": null }, { "id": "1289055", "postDate": "04/30/2021 15:45:38", "content": "No, 30 classes are not Ok. You must use the train_audios folder as well. It contains 397 birds.
", "rawMarkdown": "No, 30 classes are not Ok. You must use the **train_audios** folder as well. It contains ~~297~~ 397 birds.", "votes": null }, { "id": "1292162", "postDate": "05/03/2021 17:51:24", "content": "Hey, thank you for your pipeline, it's really good.
\nI'm trying to ensemble models on raw scores, my models have from 0.64 to 0.7 on CV and 0.58 to 0.64 on LB.
\nBut I can't seem to get good ensembles, the validation scores with low thresh go really bad and decent with higher thresh (but not better than any single model). Do you have any idea what may be causing this?
Do you include some diversity in your models ? For diversity, you can vary your training pipeline, use different backbone models, add some augs … usually diverse models would give better ensemble.
", "rawMarkdown": "Do you include some diversity in your models ? For diversity, you can vary your training pipeline, use different backbone models, add some augs ... usually diverse models would give better ensemble.", "votes": null }, { "id": "1292253", "postDate": "05/03/2021 19:31:16", "content": "Thank you, I am using different backbones but haven't tried augs yet. Will do…
\nGood luck and thank you again for your great contribution!
Thank you as well @flrotm and good luck :)
", "rawMarkdown": "Thank you as well @flrotm and good luck :)", "votes": null }, { "id": "1295167", "postDate": "05/06/2021 08:55:54", "content": "I have same question. It seems the former to me.
", "rawMarkdown": "I have same question. It seems the former to me.", "votes": null }, { "id": "1295180", "postDate": "05/06/2021 09:07:35", "content": "I think so too, unfortunately no one replied yet.
", "rawMarkdown": "I think so too, unfortunately no one replied yet.", "votes": null }, { "id": "1295195", "postDate": "05/06/2021 09:18:01", "content": "You are right. On kaggle forums, bagging usually means averaging predictions of several models. It is the simplest form of ensembling.
", "rawMarkdown": "You are right. On kaggle forums, bagging usually means averaging predictions of several models. It is the simplest form of ensembling.", "votes": null }, { "id": "1297315", "postDate": "05/07/2021 23:00:46", "content": "Hmm, I thought it's 397:
\ndf_train = pd.read_csv(\"../input/birdclef-2021/train_metadata.csv\")
\nLABEL_IDS = {label: label_id for label_id,label in enumerate(sorted(df_train[\"primary_label\"].unique()))}
\nlen(LABEL_IDS)
If anybody is looking for a teammate, please contact me from my profile.
", "rawMarkdown": "Hmm, I thought it's 397:\n`df_train = pd.read_csv(\"../input/birdclef-2021/train_metadata.csv\")`\n`LABEL_IDS = {label: label_id for label_id,label in enumerate(sorted(df_train[\"primary_label\"].unique()))}`\n`len(LABEL_IDS)`\n\n---\n\n*If anybody is looking for a teammate, please contact me from my [profile](https://www.kaggle.com/philipkd)*.", "votes": null }, { "id": "1297358", "postDate": "05/08/2021 00:26:38", "content": "@philipkd yeah you're right, my bad !
", "rawMarkdown": "philipkd yeah you're right, my bad !", "votes": null }, { "id": "1298881", "postDate": "05/09/2021 09:43:40", "content": "0.79 aka Return of the King
", "rawMarkdown": "0.79 aka Return of the King", "votes": null }, { "id": "1298953", "postDate": "05/09/2021 11:03:52", "content": "Hello , did you use voting?
", "rawMarkdown": "Hello , did you use voting?", "votes": null }, { "id": "1299216", "postDate": "05/09/2021 14:46:27", "content": "No. Just simple average.
", "rawMarkdown": "No. Just simple average.", "votes": null }, { "id": "1300476", "postDate": "05/10/2021 13:45:35", "content": "Wow,Thank you for your valuable experience.
", "rawMarkdown": "Wow,Thank you for your valuable experience.", "votes": null }, { "id": "1306658", "postDate": "05/14/2021 01:33:49", "content": "When I ensembled 61 and 59, I got 58 lb…..
", "rawMarkdown": "When I ensembled 61 and 59, I got 58 lb.....", "votes": null }, { "id": "1307933", "postDate": "05/14/2021 19:31:30", "content": "How do you ensemble?
", "rawMarkdown": "How do you ensemble?", "votes": null }, { "id": "1308081", "postDate": "05/15/2021 01:04:16", "content": "Hello,I only average each model's output score as a final out put,then use threshold to decide the categories.
", "rawMarkdown": "Hello,I only average each model's output score as a final out put,then use threshold to decide the categories.", "votes": null }, { "id": "1315883", "postDate": "05/20/2021 07:02:21", "content": "GM magic! I am on the way.
", "rawMarkdown": "GM magic! I am on the way.", "votes": null }, { "id": "1316191", "postDate": "05/20/2021 11:10:15", "content": "@cpmpml , what's the best score you have achieved by just using audio clip ? no soundscapes or metadata.
", "rawMarkdown": "cpmpml , what's the best score you have achieved by just using audio clip ? no soundscapes or metadata.", "votes": null }, { "id": "1316193", "postDate": "05/20/2021 11:11:12", "content": "You will know soon, shortly after the competition deadline ;)
", "rawMarkdown": "You will know soon, shortly after the competition deadline ;)", "votes": null }, { "id": "1316194", "postDate": "05/20/2021 11:15:45", "content": "Sure. Just getting started late
", "rawMarkdown": "Sure. Just getting started late", "votes": null } ], "comments": [ { "id": 1278278, "author_name": "arunodhayan", "author_url": "", "post_date": "04/19/2021 17:56:43", "content": "do you apply bagging on score or at the end to the labels
", "votes": null, "replies": [ { "id": 1278305, "author_name": "kneroma", "author_url": "", "post_date": "04/19/2021 18:14:44", "content": "I can see that details are very important to you … It was on raw scores.
", "votes": null, "replies": [] }, { "id": 1278343, "author_name": "arunodhayan", "author_url": "", "post_date": "04/19/2021 19:14:05", "content": "Thank you @kneroma
", "votes": null, "replies": [] } ] }, { "id": 1278319, "author_name": "imeintanis", "author_url": "", "post_date": "04/19/2021 18:38:05", "content": "Thanks for sharing @kneroma !! I suppose you mean ensemble of 2 full kfold models
", "votes": null, "replies": [ { "id": 1278330, "author_name": "kneroma", "author_url": "", "post_date": "04/19/2021 18:49:27", "content": "Thanks @imeintanis. Yes, 2 five kfold models.
", "votes": null, "replies": [] } ] }, { "id": 1278355, "author_name": "cpmpml", "author_url": "", "post_date": "04/19/2021 19:36:48", "content": "Good to see you sharing that much.
\nI was burned in previous comp, but I'll return to my level of sharing if this one ends well for me.
", "votes": null, "replies": [ { "id": 1278423, "author_name": "kneroma", "author_url": "", "post_date": "04/19/2021 21:45:29", "content": "Thanks @cpmpml . No matter if you share during this one or not, you still for sure one the best Kaggle voices !
", "votes": null, "replies": [] }, { "id": 1280253, "author_name": "nofreewill", "author_url": "", "post_date": "04/21/2021 18:36:51", "content": "@cpmpm l felt sorry for you! Even though I didn't nail it, I learned a lot, because that one comment from you kept me going. Thank you for that!
", "votes": null, "replies": [] }, { "id": 1280267, "author_name": "cpmpml", "author_url": "", "post_date": "04/21/2021 19:09:12", "content": "@nofreewill thanks! I think it helped many teams, but very few acknowledged it.
", "votes": null, "replies": [] } ] }, { "id": 1279575, "author_name": "peigangshi", "author_url": "", "post_date": "04/21/2021 04:24:17", "content": "Thank you for sharing.
", "votes": null, "replies": [ { "id": 1279656, "author_name": "kneroma", "author_url": "", "post_date": "04/21/2021 06:13:30", "content": "Thank you as well @peigangshi
", "votes": null, "replies": [] } ] }, { "id": 1279719, "author_name": "vlomme", "author_url": "", "post_date": "04/21/2021 07:26:48", "content": "for me a single model, without kfold - 0.72
\nensemble of 5 - 0.74
Nice one ! It means that there still room for improvements on the LB … Can't to wait to see how far we can go.
", "votes": null, "replies": [] }, { "id": 1279820, "author_name": "cpmpml", "author_url": "", "post_date": "04/21/2021 08:50:30", "content": "I hope 0.74 is the limit ;)
", "votes": null, "replies": [] }, { "id": 1279832, "author_name": "kneroma", "author_url": "", "post_date": "04/21/2021 09:00:58", "content": "How can I believe you @cpmpml while you already scored 0.75 😄 ?
", "votes": null, "replies": [] }, { "id": 1279850, "author_name": "cpmpml", "author_url": "", "post_date": "04/21/2021 09:21:35", "content": "I was kidding ofc.
\nBut I wouldn't mind if no one else reaches 0.75 ;)
\nMaybe I was just extremely lucky. We will know at the end of competition.
", "votes": null, "replies": [] }, { "id": 1279892, "author_name": "kneroma", "author_url": "", "post_date": "04/21/2021 10:25:28", "content": "Oh my bad ! We'll see …
", "votes": null, "replies": [] }, { "id": 1280747, "author_name": "vlomme", "author_url": "", "post_date": "04/22/2021 09:47:54", "content": "done. The road to 0.77
", "votes": null, "replies": [] }, { "id": 1280773, "author_name": "kneroma", "author_url": "", "post_date": "04/22/2021 10:28:11", "content": "Congratulations :) @vlomme
", "votes": null, "replies": [] }, { "id": 1280863, "author_name": "cpmpml", "author_url": "", "post_date": "04/22/2021 12:39:52", "content": "Well done @vlomme
\nIt's time to resume submitting it seems. I hope my local CV was right for last two weeks…
", "votes": null, "replies": [] }, { "id": 1287296, "author_name": "cpmpml", "author_url": "", "post_date": "04/28/2021 22:13:53", "content": "\n\ndone. The road to 0.77
\n
@vlomme you made it. Maybe time to stop and enjoy?
\nKidding of course, you keep me awake at night.
", "votes": null, "replies": [] }, { "id": 1287476, "author_name": "vlomme", "author_url": "", "post_date": "04/29/2021 05:51:37", "content": "@cpmpml Your result motivates me to keep working)
\nBut it takes a long time with gtx1060(
Maybe it's time to invest in a better card.
", "votes": null, "replies": [] }, { "id": 1287548, "author_name": "vlomme", "author_url": "", "post_date": "04/29/2021 07:08:54", "content": "Absolutely, but not $4,000 for 3090
", "votes": null, "replies": [] }, { "id": 1287580, "author_name": "nofreewill", "author_url": "", "post_date": "04/29/2021 07:41:08", "content": "I did that for BMS xd
", "votes": null, "replies": [] }, { "id": 1288465, "author_name": "cpmpml", "author_url": "", "post_date": "04/30/2021 03:26:10", "content": "\n\ndone. The road to 0.77
\n
@vladimirsydor well done too, you also keep me awake.
\nAnd I am sure several others will soon make it to 0.77, then pass my current score.
\nWho will be first to 0.80?
", "votes": null, "replies": [] }, { "id": 1288548, "author_name": "nofreewill", "author_url": "", "post_date": "04/30/2021 05:35:56", "content": "@brunosan2018 , is your 0.80 ready ?
", "votes": null, "replies": [] }, { "id": 1298881, "author_name": "nofreewill", "author_url": "", "post_date": "05/09/2021 09:43:40", "content": "0.79 aka Return of the King
", "votes": null, "replies": [] } ] }, { "id": 1279742, "author_name": "cpmpml", "author_url": "", "post_date": "04/21/2021 07:57:53", "content": "A single model 5 folds LB 0.75
", "votes": null, "replies": [ { "id": 1315883, "author_name": "shigengtian", "author_url": "", "post_date": "05/20/2021 07:02:21", "content": "GM magic! I am on the way.
", "votes": null, "replies": [] }, { "id": 1316191, "author_name": "rsinda", "author_url": "", "post_date": "05/20/2021 11:10:15", "content": "@cpmpml , what's the best score you have achieved by just using audio clip ? no soundscapes or metadata.
", "votes": null, "replies": [] }, { "id": 1316193, "author_name": "cpmpml", "author_url": "", "post_date": "05/20/2021 11:11:12", "content": "You will know soon, shortly after the competition deadline ;)
", "votes": null, "replies": [] }, { "id": 1316194, "author_name": "rsinda", "author_url": "", "post_date": "05/20/2021 11:15:45", "content": "Sure. Just getting started late
", "votes": null, "replies": [] } ] }, { "id": 1279840, "author_name": "vishalgrover5", "author_url": "", "post_date": "04/21/2021 09:07:36", "content": "Thanks for sharing this.
", "votes": null, "replies": [ { "id": 1279841, "author_name": "kneroma", "author_url": "", "post_date": "04/21/2021 09:09:32", "content": "Thanks @vishalgrover5
", "votes": null, "replies": [] } ] }, { "id": 1283410, "author_name": "aliabdin1", "author_url": "", "post_date": "04/24/2021 22:30:33", "content": "Do you mean with bagging a \"normal\" ensemble or the bagging (\"bootstrap aggregating\") method by training a set amount of random samples for each classifier?
", "votes": null, "replies": [ { "id": 1295167, "author_name": "yamanity", "author_url": "", "post_date": "05/06/2021 08:55:54", "content": "I have same question. It seems the former to me.
", "votes": null, "replies": [] }, { "id": 1295180, "author_name": "aliabdin1", "author_url": "", "post_date": "05/06/2021 09:07:35", "content": "I think so too, unfortunately no one replied yet.
", "votes": null, "replies": [] }, { "id": 1295195, "author_name": "cpmpml", "author_url": "", "post_date": "05/06/2021 09:18:01", "content": "You are right. On kaggle forums, bagging usually means averaging predictions of several models. It is the simplest form of ensembling.
", "votes": null, "replies": [] } ] }, { "id": 1284410, "author_name": "brunosan2018", "author_url": "", "post_date": "04/25/2021 22:07:33", "content": "Coming soon.. boosting 0.80 best score …..be afraid?, ..it's a joke, I'm trying my best, my first time competing, but ensambles is the key… i wish u all the best!
", "votes": null, "replies": [ { "id": 1284592, "author_name": "cpmpml", "author_url": "", "post_date": "04/26/2021 05:55:41", "content": "Show us!
", "votes": null, "replies": [] } ] }, { "id": 1288268, "author_name": "doanquanvietnamca", "author_url": "", "post_date": "04/29/2021 19:53:58", "content": "Nice to hear that! Thanks, @kneroma !
", "votes": null, "replies": [ { "id": 1288934, "author_name": "kneroma", "author_url": "", "post_date": "04/30/2021 13:38:05", "content": "Thank you as well dear @doanquanvietnamca
", "votes": null, "replies": [] } ] }, { "id": 1289040, "author_name": "brunosan2018", "author_url": "", "post_date": "04/30/2021 15:25:58", "content": "Hello, I have a question, how many kinds of birds did you use to train? for 30 classes I have an f1 0.65 to 0.73 … 30 classes okay? please help me :( need a hands
", "votes": null, "replies": [ { "id": 1289055, "author_name": "kneroma", "author_url": "", "post_date": "04/30/2021 15:45:38", "content": "No, 30 classes are not Ok. You must use the train_audios folder as well. It contains 397 birds.
", "votes": null, "replies": [] }, { "id": 1297315, "author_name": "philipkd", "author_url": "", "post_date": "05/07/2021 23:00:46", "content": "Hmm, I thought it's 397:
\ndf_train = pd.read_csv(\"../input/birdclef-2021/train_metadata.csv\")
\nLABEL_IDS = {label: label_id for label_id,label in enumerate(sorted(df_train[\"primary_label\"].unique()))}
\nlen(LABEL_IDS)
If anybody is looking for a teammate, please contact me from my profile.
", "votes": null, "replies": [] }, { "id": 1297358, "author_name": "kneroma", "author_url": "", "post_date": "05/08/2021 00:26:38", "content": "@philipkd yeah you're right, my bad !
", "votes": null, "replies": [] } ] }, { "id": 1292162, "author_name": "flrotm", "author_url": "", "post_date": "05/03/2021 17:51:24", "content": "Hey, thank you for your pipeline, it's really good.
\nI'm trying to ensemble models on raw scores, my models have from 0.64 to 0.7 on CV and 0.58 to 0.64 on LB.
\nBut I can't seem to get good ensembles, the validation scores with low thresh go really bad and decent with higher thresh (but not better than any single model). Do you have any idea what may be causing this?
Do you include some diversity in your models ? For diversity, you can vary your training pipeline, use different backbone models, add some augs … usually diverse models would give better ensemble.
", "votes": null, "replies": [] } ] }, { "id": 1292253, "author_name": "flrotm", "author_url": "", "post_date": "05/03/2021 19:31:16", "content": "Thank you, I am using different backbones but haven't tried augs yet. Will do…
\nGood luck and thank you again for your great contribution!
Thank you as well @flrotm and good luck :)
", "votes": null, "replies": [] } ] }, { "id": 1298953, "author_name": "zekunn", "author_url": "", "post_date": "05/09/2021 11:03:52", "content": "Hello , did you use voting?
", "votes": null, "replies": [ { "id": 1299216, "author_name": "kneroma", "author_url": "", "post_date": "05/09/2021 14:46:27", "content": "No. Just simple average.
", "votes": null, "replies": [] } ] }, { "id": 1300476, "author_name": "majunfu", "author_url": "", "post_date": "05/10/2021 13:45:35", "content": "Wow,Thank you for your valuable experience.
", "votes": null, "replies": [] }, { "id": 1306658, "author_name": "zekunn", "author_url": "", "post_date": "05/14/2021 01:33:49", "content": "When I ensembled 61 and 59, I got 58 lb…..
", "votes": null, "replies": [ { "id": 1307933, "author_name": "cpmpml", "author_url": "", "post_date": "05/14/2021 19:31:30", "content": "How do you ensemble?
", "votes": null, "replies": [] }, { "id": 1308081, "author_name": "zekunn", "author_url": "", "post_date": "05/15/2021 01:04:16", "content": "Hello,I only average each model's output score as a final out put,then use threshold to decide the categories.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1277942": "At the beginning of this competition, I was outraged to see how low my score on the LB was compared to the headliners. But, in fact, it wasn't! My score was correct! I didn't realize it until I submitted a simple average of my 2 best models: bagging works extremely well in this competition. As an example, I went from a score around **0.63** to **0.73**.\n\nIf I share, it's not to taunt people but rather to encourage those like me who are wondering how to break the **0.70** worm !", "1278278": "do you apply bagging on score or at the end to the labels", "1278305": "I can see that details are very important to you ... It was on raw scores.", "1278319": "Thanks for sharing @kneroma !! I suppose you mean ensemble of 2 full kfold models", "1278330": "Thanks @imeintanis. Yes, 2 five kfold models.", "1278343": "Thank you @kneroma", "1278355": "Good to see you sharing that much.\n\nI was burned in previous comp, but I'll return to my level of sharing if this one ends well for me.", "1278423": "Thanks @cpmpml . No matter if you share during this one or not, you still for sure one the best Kaggle voices !", "1279575": "Thank you for sharing.", "1279656": "Thank you as well @peigangshi", "1279719": "for me a single model, without kfold - 0.72 \nensemble of 5 - 0.74", "1279742": "A single model 5 folds LB 0.75", "1279761": "Nice one ! It means that there still room for improvements on the LB ... Can't to wait to see how far we can go.", "1279820": "I ~~believe~~ hope 0.74 is the limit ;)", "1279832": "How can I believe you @cpmpml while you already scored **0.75** 😄 ?", "1279840": "Thanks for sharing this.", "1279841": "Thanks @vishalgrover5", "1279850": "I was kidding ofc.\n\nBut I wouldn't mind if no one else reaches 0.75 ;)\n\nMaybe I was just extremely lucky. We will know at the end of competition.", "1279892": "Oh my bad ! We'll see ...", "1280253": "cpmpm l felt sorry for you! Even though I didn't nail it, I learned a lot, because that one comment from you kept me going. Thank you for that!", "1280267": "nofreewill thanks! I think it helped many teams, but very few acknowledged it.", "1280747": "done. The road to 0.77", "1280773": "Congratulations :) @vlomme", "1280863": "Well done @vlomme \n\nIt's time to resume submitting it seems. I hope my local CV was right for last two weeks...", "1283410": "Do you mean with bagging a \"normal\" ensemble or the bagging (\"bootstrap aggregating\") method by training a set amount of random samples for each classifier?", "1284410": "Coming soon.. boosting 0.80 best score .....be afraid?, ..it's a joke, I'm trying my best, my first time competing, but ensambles is the key... i wish u all the best!", "1284592": "Show us!", "1287296": "> done. The road to 0.77\n\n@vlomme you made it. Maybe time to stop and enjoy?\n\n\nKidding of course, you keep me awake at night.", "1287476": "cpmpml Your result motivates me to keep working)\nBut it takes a long time with gtx1060(", "1287537": "Maybe it's time to invest in a better card.", "1287548": "Absolutely, but not $4,000 for 3090", "1287580": "I did that for BMS xd", "1288268": "Nice to hear that! Thanks, @kneroma !", "1288465": "> done. The road to 0.77\n\n@vladimirsydor well done too, you also keep me awake.\n\nAnd I am sure several others will soon make it to 0.77, then pass my current score.\n\nWho will be first to 0.80?", "1288548": "brunosan2018 , is your 0.80 ready ?", "1288934": "Thank you as well dear @doanquanvietnamca", "1289040": "Hello, I have a question, how many kinds of birds did you use to train? for 30 classes I have an f1 0.65 to 0.73 ... 30 classes okay? please help me :( need a hands", "1289055": "No, 30 classes are not Ok. You must use the **train_audios** folder as well. It contains ~~297~~ 397 birds.", "1292162": "Hey, thank you for your pipeline, it's really good.\nI'm trying to ensemble models on raw scores, my models have from 0.64 to 0.7 on CV and 0.58 to 0.64 on LB.\nBut I can't seem to get good ensembles, the validation scores with low thresh go really bad and decent with higher thresh (but not better than any single model). Do you have any idea what may be causing this?", "1292243": "Do you include some diversity in your models ? For diversity, you can vary your training pipeline, use different backbone models, add some augs ... usually diverse models would give better ensemble.", "1292253": "Thank you, I am using different backbones but haven't tried augs yet. Will do...\nGood luck and thank you again for your great contribution!", "1292383": "Thank you as well @flrotm and good luck :)", "1295167": "I have same question. It seems the former to me.", "1295180": "I think so too, unfortunately no one replied yet.", "1295195": "You are right. On kaggle forums, bagging usually means averaging predictions of several models. It is the simplest form of ensembling.", "1297315": "Hmm, I thought it's 397:\n`df_train = pd.read_csv(\"../input/birdclef-2021/train_metadata.csv\")`\n`LABEL_IDS = {label: label_id for label_id,label in enumerate(sorted(df_train[\"primary_label\"].unique()))}`\n`len(LABEL_IDS)`\n\n---\n\n*If anybody is looking for a teammate, please contact me from my [profile](https://www.kaggle.com/philipkd)*.", "1297358": "philipkd yeah you're right, my bad !", "1298881": "0.79 aka Return of the King", "1298953": "Hello , did you use voting?", "1299216": "No. Just simple average.", "1300476": "Wow,Thank you for your valuable experience.", "1306658": "When I ensembled 61 and 59, I got 58 lb.....", "1307933": "How do you ensemble?", "1308081": "Hello,I only average each model's output score as a final out put,then use threshold to decide the categories.", "1315883": "GM magic! I am on the way.", "1316191": "cpmpml , what's the best score you have achieved by just using audio clip ? no soundscapes or metadata.", "1316193": "You will know soon, shortly after the competition deadline ;)", "1316194": "Sure. Just getting started late" }, "source": "meta" }