LB RunningTime
\n0.848 ~23min
\n0.842 ~16min
\nwhat's your best single model scored and how long it take?
LB RunningTime
\n0.848 ~23min
\n0.842 ~16min
\nwhat's your best single model scored and how long it take?
Update: Single best model efficientnet_b0 [Running Time: ~10 Mins]
\nLB: 0.90
\nTrain Audio + Train soundscapes
", "rawMarkdown": "Update: Single best model efficientnet_b0 [Running Time: ~10 Mins]\n\nLB: 0.90\n\nTrain Audio + Train soundscapes", "votes": 8, "replies": [ { "id": 3193272, "postDate": "2025-05-04T06:38:08.640Z", "content": "Wow, I would swear I have tried everything and anything. I cant seems to get any positives out of train soundscapes. Very impressive
", "rawMarkdown": "Wow, I would swear I have tried everything and anything. I cant seems to get any positives out of train soundscapes. Very impressive", "replies": [ { "id": 3193709, "postDate": "2025-05-04T21:56:23.133Z", "content": "I think I am over-fitting to the public LB, Shake up is coming. 😂
", "rawMarkdown": "I think I am over-fitting to the public LB, Shake up is coming. 😂" } ] } ] }, { "id": 3164424, "postDate": "2025-03-31T17:51:17.153Z", "content": "LB = 0.858 Efficientnet B0 (Openvino) Running Time Around 13 Minutes - No Threads, Simple Loop based submission for now.
\nLB = 0.859 Efficientnet B3 (Openvino) Running Time Around 25 Minutes - No Threads, Simple Loop based submission for now.
\nModels are really unstable, and highly depends on batch distribution in mixup.
", "rawMarkdown": "LB = 0.858 Efficientnet B0 (Openvino) Running Time Around 13 Minutes - No Threads, Simple Loop based submission for now.\n\nLB = 0.859 Efficientnet B3 (Openvino) Running Time Around 25 Minutes - No Threads, Simple Loop based submission for now.\n\n\nModels are really unstable, and highly depends on batch distribution in mixup. ", "votes": 9, "replies": [ { "id": 3166988, "postDate": "2025-04-01T07:40:33.747Z", "content": "Thanks for sharing. Is this using all the data for training?
", "rawMarkdown": "Thanks for sharing. Is this using all the data for training?", "replies": [ { "id": 3167150, "postDate": "2025-04-01T10:00:43.070Z", "content": "Just with Train Audios.
\nStill trying to figure out a way to use Soundscapes.
Are these results based on SED?
", "rawMarkdown": "Are these results based on SED?" }, { "id": 3167273, "postDate": "2025-04-01T12:17:43.600Z", "rawMarkdown": "", "isDeleted": true }, { "id": 3167276, "postDate": "2025-04-01T12:18:49.283Z", "content": "So do I, I have tried several methods to utilize unlabeled soundscapes, but none of them can improve my LB :(.
", "rawMarkdown": "So do I, I have tried several methods to utilize unlabeled soundscapes, but none of them can improve my LB :(." } ] }, { "id": 3170736, "postDate": "2025-04-05T00:34:46.607Z", "content": "single fold b0 getting .858?? Assuming current best is an ensmble of folds of these 2. Thats insane my b0 is not even close haha
", "rawMarkdown": "single fold b0 getting .858?? Assuming current best is an ensmble of folds of these 2. Thats insane my b0 is not even close haha", "votes": 1 } ] }, { "id": 3180715, "postDate": "2025-04-17T03:17:05.700Z", "content": "Finally, I have a single CNN Efficientnet-B0 with LB: 0.875. (Running Time ~ 15 Mins).
\nI have tried a lot of ways to use pseudo labels, but I don't think it's helping at all, it just lowers the LB.
", "rawMarkdown": "Finally, I have a single CNN Efficientnet-B0 with LB: 0.875. (Running Time ~ 15 Mins).\n\n\nI have tried a lot of ways to use pseudo labels, but I don't think it's helping at all, it just lowers the LB. ", "votes": 5, "replies": [ { "id": 3180744, "postDate": "2025-04-17T04:28:51.747Z", "content": "Amazing result! I have used unlabel data and it help me increase my lb score ~0.015.
", "rawMarkdown": "Amazing result! I have used unlabel data and it help me increase my lb score ~0.015.", "replies": [ { "id": 3180779, "postDate": "2025-04-17T05:25:58.650Z", "content": "Did you use that unlabeled data as pseudo label or some other way?
", "rawMarkdown": "Did you use that unlabeled data as pseudo label or some other way?" }, { "id": 3180785, "postDate": "2025-04-17T05:39:27.673Z", "content": "Emm.. In my case, generally speaking, I think it can't be called as pseudo label.
", "rawMarkdown": "Emm.. In my case, generally speaking, I think it can't be called as pseudo label." }, { "id": 3180787, "postDate": "2025-04-17T05:41:49.107Z", "content": "Cool. Thanks.
", "rawMarkdown": "Cool. Thanks." } ] } ] }, { "id": 3178587, "postDate": "2025-04-14T11:02:32.583Z", "content": "0.845 b0 tuning public notebook hyper-parameters only.
", "rawMarkdown": "0.845 b0 tuning public notebook hyper-parameters only.", "votes": 5 }, { "id": 3191953, "postDate": "2025-05-02T10:00:50.870Z", "content": "My best single‐model submission:
\nModel: SED architecture (eca_nfnet_l0)
\nLB Score: 0.890
\nRunning Time: ~14 min
Nice, so far I have 0.878 in ~12min. Are you using the unlabeled data?
", "rawMarkdown": "Nice, so far I have 0.878 in ~12min. Are you using the unlabeled data?", "votes": 2, "replies": [ { "id": 3193944, "postDate": "2025-05-05T08:37:19.553Z", "content": "Yes, I'm using the unlabeled data. Incorporating it into my training pipeline has significantly boosted my LB score.
", "rawMarkdown": "Yes, I'm using the unlabeled data. Incorporating it into my training pipeline has significantly boosted my LB score.", "votes": 1 } ] }, { "id": 3192050, "postDate": "2025-05-02T12:13:31.143Z", "content": "my best model 0.870 ~23min.
", "rawMarkdown": "my best model 0.870 ~23min.", "votes": 3, "replies": [ { "id": 3193945, "postDate": "2025-05-05T08:37:23.687Z", "content": "Great result!
", "rawMarkdown": "Great result! " } ] }, { "id": 3196499, "postDate": "2025-05-07T05:21:59.023Z", "content": "\n\nMy best single‐model submission:
\nModel: SED architecture (eca_nfnet_l0)
\n
\n LB Score: 0.890
\n Running Time: ~14 min
May I ask how your eca_nfnet_l0 worked, i tried it but cannot convert it into openvion
", "rawMarkdown": "> My best single‐model submission:\n> \n> Model: SED architecture (eca_nfnet_l0)\n> LB Score: 0.890\n> Running Time: ~14 min\n\nMay I ask how your eca_nfnet_l0 worked, i tried it but cannot convert it into openvion ", "votes": 1, "replies": [ { "id": 3196526, "postDate": "2025-05-07T05:55:01.907Z", "content": "I saw previous competition solution, they use OpenVINO for Efficientnet, and ONNX for eca_nfnet_l0
", "rawMarkdown": "I saw previous competition solution, they use OpenVINO for Efficientnet, and ONNX for eca_nfnet_l0", "votes": 2 }, { "id": 3196538, "postDate": "2025-05-07T06:10:21.890Z", "content": "eca_nfnet_l0 works with latest openvino
", "rawMarkdown": "eca_nfnet_l0 works with latest openvino", "votes": 2 }, { "id": 3196543, "postDate": "2025-05-07T06:15:35.833Z", "content": "thanks so mcuh for your reply
", "rawMarkdown": "thanks so mcuh for your reply", "votes": 1 }, { "id": 3200780, "postDate": "2025-05-13T03:54:10.197Z", "content": "I'm afraid I haven't tried OpenVINO yet.
", "rawMarkdown": "I'm afraid I haven't tried OpenVINO yet." } ] }, { "id": 3200784, "postDate": "2025-05-13T03:59:32.040Z", "content": "Update
\nMy best single-model submission:
\nModel: SED architecture (tf_efficientnetv2_s.in21k)
\nLB Score: 0.903
\nRunning Time: ~12 min
do you train on different segment length compared with when you predict?
", "rawMarkdown": "do you train on different segment length compared with when you predict?" }, { "id": 3200974, "postDate": "2025-05-13T09:31:22.570Z", "content": "Yes, I do use different segment lengths
", "rawMarkdown": "Yes, I do use different segment lengths", "votes": 2 } ] } ] }, { "id": 3161810, "postDate": "2025-03-28T12:18:35.970Z", "content": "LB=0.834, Running Time ~6 min, single model without any ensembling.
", "rawMarkdown": "LB=0.834, Running Time ~6 min, single model without any ensembling.", "votes": 3, "replies": [ { "id": 3161814, "postDate": "2025-03-28T12:24:59.967Z", "content": "Wow~ very efficient model. Is it efficientnetb0?
", "rawMarkdown": "Wow~ very efficient model. Is it efficientnetb0?", "replies": [ { "id": 3161844, "postDate": "2025-03-28T13:30:29.240Z", "content": "It's some kind of custom 'Frankenstein-like' architecture.
\nI'm using ThreadPoolExecutor for parallel processing of the data. However, I don't use quantization and inference accelerators like ONNX or OpenVINO since we have 90 min and it seems more than enough.
Update, LB:0.902 and ~20min running time
", "rawMarkdown": "Update, LB:0.902 and ~20min running time", "votes": 1, "replies": [ { "id": 3209840, "postDate": "2025-05-26T11:50:25.727Z", "content": "SED or CNN？
", "rawMarkdown": "SED or CNN？", "replies": [ { "id": 3210315, "postDate": "2025-05-27T03:12:06.867Z", "content": "SED with efv2b3 backbone
", "rawMarkdown": "SED with efv2b3 backbone", "votes": 1 } ] }, { "id": 3209963, "postDate": "2025-05-26T14:56:32.603Z", "content": "same results
", "rawMarkdown": "same results\n" } ] }, { "id": 3174342, "postDate": "2025-04-09T02:09:51.297Z", "content": "Finally a single model EfficientNet B3 (LB 0.864). 😪
\nTrain Data Only.
So, now your best score (0.872) is a model ensemble version? I haven't tried ensemble yet. Try to improve my best single model but it has been stuck at 0.849 for two weeks already. 😴
", "rawMarkdown": "So, now your best score (0.872) is a model ensemble version? I haven't tried ensemble yet. Try to improve my best single model but it has been stuck at 0.849 for two weeks already. 😴", "replies": [ { "id": 3174361, "postDate": "2025-04-09T02:39:37.953Z", "content": "Unfortunately the ensemble result didn’t change by adding this model or removing other models.
\nMy best LB score is with an ensemble of 2 x B0 and 2 x B3 (scores 0.854, 0.858, 0.856, 0.859) and ensemble scores 0.870 with post processing 0.872.
", "rawMarkdown": "Unfortunately the ensemble result didn’t change by adding this model or removing other models. \n\nMy best LB score is with an ensemble of 2 x B0 and 2 x B3 (scores 0.854, 0.858, 0.856, 0.859) and ensemble scores 0.870 with post processing 0.872.\n", "votes": 3 }, { "id": 3174366, "postDate": "2025-04-09T02:54:10.483Z", "content": "I just submit my first ensemble version and it has finished running few minutes ago. (seresnext, efv2b3, ecanfnetl0) and it scored 0.846, which is slight worse than my efv2b3 model (0.849) . I expect it would be higher than 0.849 but it is not. Seems like there maybe some bug in my seresnext or ecanfnetl0 model (I have not submit them yet).
", "rawMarkdown": "I just submit my first ensemble version and it has finished running few minutes ago. (seresnext, efv2b3, ecanfnetl0) and it scored 0.846, which is slight worse than my efv2b3 model (0.849) . I expect it would be higher than 0.849 but it is not. Seems like there maybe some bug in my seresnext or ecanfnetl0 model (I have not submit them yet)." }, { "id": 3174369, "postDate": "2025-04-09T02:59:51.717Z", "content": "I see,
\nEnsemble with logits was not working for me in this competition against my experiments.
\nSo, I had to ensemble after sigmoid.
\nNot sure, if that's the case for you.
Thank you for your advice~ I do use the probility instead of logits to ensemble model, so there maybe some other problem I guess.
", "rawMarkdown": "Thank you for your advice~ I do use the probility instead of logits to ensemble model, so there maybe some other problem I guess." }, { "id": 3176007, "postDate": "2025-04-10T20:04:47.813Z", "content": "I think, it's quite reasonable that ensembling with logits doesn't work, because the logits of each model in the ensemble may have different scales. For example, suppose we have multiplied the model's logits by some factor, i.e. 10. The ROCAUC score of this particular model remains the same. However, the weight of the model in the ensemble is increased a lot. By applying the sigmoid we put the model's output into the uniform scale, and all the models play equivalent role in the ensemble.
", "rawMarkdown": "I think, it's quite reasonable that ensembling with logits doesn't work, because the logits of each model in the ensemble may have different scales. For example, suppose we have multiplied the model's logits by some factor, i.e. 10. The ROCAUC score of this particular model remains the same. However, the weight of the model in the ensemble is increased a lot. By applying the sigmoid we put the model's output into the uniform scale, and all the models play equivalent role in the ensemble." }, { "id": 3182144, "postDate": "2025-04-18T20:00:15.913Z", "content": "Each model in the ensemble is 1 fold or you are using multiple folds/model (e.g. 5 folds/model)?
", "rawMarkdown": "Each model in the ensemble is 1 fold or you are using multiple folds/model (e.g. 5 folds/model)?" } ] }, { "id": 3175194, "postDate": "2025-04-09T20:54:48.617Z", "content": "Thanks for sharing! Would you mind clarifying what Train Data Only setting refers to? From your public notebook last year, I noticed that you used 5-fold cross-validation and only included the first 5 seconds of each audio sample. Does this approach correspond to the Train Data Only setting, and the score is from one of the five folds? Thank you!
by Train Data Only he means he is not using the train_soundscapes currently. I another post of his he mention that for training he is using random 5s clip from the track and for testing he is using the first 5s clip from a track.
It means, no train_soundscapes.
\nIt's random 5 seconds, first 5 seconds are performing much worse.
Thank you!
", "rawMarkdown": "Thank you!" } ] } ] }, { "id": 3177766, "postDate": "2025-04-13T08:38:39.683Z", "content": "Hi everyone, interesting discussion on single model performance!
\nI'm feeling a bit stuck myself. I've been working off a baseline and haven't managed to push the score much beyond that yet (around the low 0.8x range). I tried a few things like using a 3-channel input (mel spec + MFCCs + delta MFCCs) but that didn't seem to help, unfortunately.
\nLooking at the summary of past top solutions (like the one I compiled kaggle discussion), it's clear that pseudo-labeling on the unlabeled soundscapes was almost universally beneficial last year. Handling secondary labels effectively also seems important.
\nMy plan is to try implementing pseudo-labeling next, and then explore ideas like weighted loss for secondary labels or maybe even triplet loss if I get adventurous.
\nGiven that I'm currently plateauing with the basics, where would you suggest focusing efforts right now for the potentially biggest impact? Is diving straight into pseudo-labelling the best next step, or should I perhaps revisit data augmentation, model architecture choices, or something else first?
\nAppreciate any insights from those who've managed to break through similar plateaus!
", "rawMarkdown": "Hi everyone, interesting discussion on single model performance!\n\nI'm feeling a bit stuck myself. I've been working off a baseline and haven't managed to push the score much beyond that yet (around the low 0.8x range). I tried a few things like using a 3-channel input (mel spec + MFCCs + delta MFCCs) but that didn't seem to help, unfortunately.\n\nLooking at the summary of past top solutions (like the one I compiled [kaggle discussion](https://www.kaggle.com/competitions/birdclef-2025/discussion/572928)), it's clear that pseudo-labeling on the unlabeled soundscapes was almost universally beneficial last year. Handling secondary labels effectively also seems important.\n\nMy plan is to try implementing pseudo-labeling next, and then explore ideas like weighted loss for secondary labels or maybe even triplet loss if I get adventurous.\n\nGiven that I'm currently plateauing with the basics, where would you suggest focusing efforts *right now* for the potentially biggest impact? Is diving straight into pseudo-labelling the best next step, or should I perhaps revisit data augmentation, model architecture choices, or something else first?\n\nAppreciate any insights from those who've managed to break through similar plateaus!", "replies": [ { "id": 3177840, "postDate": "2025-04-13T10:19:56.157Z", "content": "\nI just created a post, to give you some ideas. https://www.kaggle.com/competitions/birdclef-2025/discussion/573066
", "rawMarkdown": "Hi @devasypatel23 \n\nI just created a post, to give you some ideas. [https://www.kaggle.com/competitions/birdclef-2025/discussion/573066](https://www.kaggle.com/competitions/birdclef-2025/discussion/573066)", "replies": [ { "id": 3178154, "postDate": "2025-04-13T19:35:38.490Z", "content": "Hey @salmanahmedtamu thanks for the effort, I read you post, I'll surely try Focal loss and other mentioned things next, Meanwhile I just tried CE loss as suggested by last year's 1st place team, but wasn't able to find much success, will play around with it for a while maybe try to tune the params a bit
", "rawMarkdown": "Hey @salmanahmedtamu thanks for the effort, I read you post, I'll surely try Focal loss and other mentioned things next, Meanwhile I just tried CE loss as suggested by last year's 1st place team, but wasn't able to find much success, will play around with it for a while maybe try to tune the params a bit" } ] } ] }, { "id": 3179729, "postDate": "2025-04-15T16:16:47.527Z", "content": "0.807, <1 min
\nI'm still trying to catch up to the public baselines & understand all the performance gaps, but hey at least I get fast feedback when submitting! All 60ish of my submissions have used GPU with a b0 model (see https://www.kaggle.com/code/robbynevels/bc25-infer-baseline-gpu/)
", "rawMarkdown": "0.807, <1 min \n\nI'm still trying to catch up to the public baselines & understand all the performance gaps, but hey at least I get fast feedback when submitting! All 60ish of my submissions have used GPU with a b0 model (see https://www.kaggle.com/code/robbynevels/bc25-infer-baseline-gpu/)", "votes": -1 }, { "id": 3202890, "postDate": "2025-05-16T04:07:55.877Z", "content": "Update: LB: 0.889, Running Time: ~20min
", "rawMarkdown": "Update: LB: 0.889, Running Time: ~20min" }, { "id": 3174971, "postDate": "2025-04-09T16:09:02.743Z", "content": "Update. Best model scored 0.860, running time ~19min
", "rawMarkdown": "Update. Best model scored 0.860, running time ~19min", "replies": [ { "id": 3176066, "postDate": "2025-04-10T23:40:30.753Z", "content": "Is it SED or CNN?
\nI haven't tried SED yet.
SED， I haven't tried CNN
", "rawMarkdown": "SED， I haven't tried CNN" } ] }, { "id": 3176275, "postDate": "2025-04-11T07:19:47.913Z", "content": "My best model scored 85.5 in 13 minutes, but the training is unstable, small changes lead to drastic fluctuations in the score. Model soup helped improve the score but didn’t contribute to training stability.
", "rawMarkdown": "My best model scored 85.5 in 13 minutes, but the training is unstable, small changes lead to drastic fluctuations in the score. Model soup helped improve the score but didn’t contribute to training stability." } ] }, { "id": 3162157, "postDate": "2025-03-28T21:30:00.063Z", "content": "16 and 23 minutes? Is this efficientnet b0?
", "rawMarkdown": "16 and 23 minutes? Is this efficientnet b0?", "replies": [ { "id": 3162247, "postDate": "2025-03-29T02:04:26.830Z", "content": "Efficientnetv2b3 backbone with SED model architecture
", "rawMarkdown": "Efficientnetv2b3 backbone with SED model architecture", "votes": 2 } ] }, { "id": 3161801, "postDate": "2025-03-28T11:59:18.593Z", "content": "What do you mean by a single model? N folds / full train / single fold?
", "rawMarkdown": "What do you mean by a single model? **N** folds / full train / single fold?", "replies": [ { "id": 3161811, "postDate": "2025-03-28T12:19:15.903Z", "content": "Single fold or full train，I use all data to train without validation.
", "rawMarkdown": "Single fold or full train，I use all data to train without validation. " } ] }, { "id": 3180086, "postDate": "2025-04-16T06:05:21.967Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 3192450, "author_name": "Salman Ahmed", "author_url": "", "post_date": "2025-05-02T20:57:12.070000", "content": "Update: Single best model efficientnet_b0 [Running Time: ~10 Mins]
\nLB: 0.90
\nTrain Audio + Train soundscapes
", "votes": 8, "replies": [ { "id": 3193272, "author_name": "Fir_las_47", "author_url": "", "post_date": "2025-05-04T06:38:08.640000", "content": "Wow, I would swear I have tried everything and anything. I cant seems to get any positives out of train soundscapes. Very impressive
", "votes": 0, "replies": [ { "id": 3193709, "author_name": "Salman Ahmed", "author_url": "", "post_date": "2025-05-04T21:56:23.133000", "content": "I think I am over-fitting to the public LB, Shake up is coming. 😂
", "votes": 0, "replies": [] } ] } ] }, { "id": 3164424, "author_name": "Salman Ahmed", "author_url": "", "post_date": "2025-03-31T17:51:17.153000", "content": "LB = 0.858 Efficientnet B0 (Openvino) Running Time Around 13 Minutes - No Threads, Simple Loop based submission for now.
\nLB = 0.859 Efficientnet B3 (Openvino) Running Time Around 25 Minutes - No Threads, Simple Loop based submission for now.
\nModels are really unstable, and highly depends on batch distribution in mixup.
", "votes": 9, "replies": [ { "id": 3166988, "author_name": "Hicham Bellafkir", "author_url": "", "post_date": "2025-04-01T07:40:33.747000", "content": "Thanks for sharing. Is this using all the data for training?
", "votes": 0, "replies": [ { "id": 3167150, "author_name": "Salman Ahmed", "author_url": "", "post_date": "2025-04-01T10:00:43.070000", "content": "Just with Train Audios.
\nStill trying to figure out a way to use Soundscapes.
Are these results based on SED?
", "votes": 0, "replies": [] }, { "id": 3167273, "author_name": "", "author_url": "", "post_date": "2025-04-01T12:17:43.600000", "content": "", "votes": 0, "replies": [] }, { "id": 3167276, "author_name": "Starry", "author_url": "", "post_date": "2025-04-01T12:18:49.283000", "content": "So do I, I have tried several methods to utilize unlabeled soundscapes, but none of them can improve my LB :(.
", "votes": 0, "replies": [] } ] }, { "id": 3170736, "author_name": "Fir_las_47", "author_url": "", "post_date": "2025-04-05T00:34:46.607000", "content": "single fold b0 getting .858?? Assuming current best is an ensmble of folds of these 2. Thats insane my b0 is not even close haha
", "votes": 1, "replies": [] } ] }, { "id": 3180715, "author_name": "Salman Ahmed", "author_url": "", "post_date": "2025-04-17T03:17:05.700000", "content": "Finally, I have a single CNN Efficientnet-B0 with LB: 0.875. (Running Time ~ 15 Mins).
\nI have tried a lot of ways to use pseudo labels, but I don't think it's helping at all, it just lowers the LB.
", "votes": 5, "replies": [ { "id": 3180744, "author_name": "Starry", "author_url": "", "post_date": "2025-04-17T04:28:51.747000", "content": "Amazing result! I have used unlabel data and it help me increase my lb score ~0.015.
", "votes": 0, "replies": [ { "id": 3180779, "author_name": "Salman Ahmed", "author_url": "", "post_date": "2025-04-17T05:25:58.650000", "content": "Did you use that unlabeled data as pseudo label or some other way?
", "votes": 0, "replies": [] }, { "id": 3180785, "author_name": "Starry", "author_url": "", "post_date": "2025-04-17T05:39:27.673000", "content": "Emm.. In my case, generally speaking, I think it can't be called as pseudo label.
", "votes": 0, "replies": [] }, { "id": 3180787, "author_name": "Salman Ahmed", "author_url": "", "post_date": "2025-04-17T05:41:49.107000", "content": "Cool. Thanks.
", "votes": 0, "replies": [] } ] } ] }, { "id": 3178587, "author_name": "yuanzhe zhou", "author_url": "", "post_date": "2025-04-14T11:02:32.583000", "content": "0.845 b0 tuning public notebook hyper-parameters only.
", "votes": 5, "replies": [] }, { "id": 3191953, "author_name": "MYSO", "author_url": "", "post_date": "2025-05-02T10:00:50.870000", "content": "My best single‐model submission:
\nModel: SED architecture (eca_nfnet_l0)
\nLB Score: 0.890
\nRunning Time: ~14 min
Nice, so far I have 0.878 in ~12min. Are you using the unlabeled data?
", "votes": 2, "replies": [ { "id": 3193944, "author_name": "MYSO", "author_url": "", "post_date": "2025-05-05T08:37:19.553000", "content": "Yes, I'm using the unlabeled data. Incorporating it into my training pipeline has significantly boosted my LB score.
", "votes": 1, "replies": [] } ] }, { "id": 3192050, "author_name": "Starry", "author_url": "", "post_date": "2025-05-02T12:13:31.143000", "content": "my best model 0.870 ~23min.
", "votes": 3, "replies": [ { "id": 3193945, "author_name": "MYSO", "author_url": "", "post_date": "2025-05-05T08:37:23.687000", "content": "Great result!
", "votes": 0, "replies": [] } ] }, { "id": 3196499, "author_name": "HZM", "author_url": "", "post_date": "2025-05-07T05:21:59.023000", "content": "\n\nMy best single‐model submission:
\nModel: SED architecture (eca_nfnet_l0)
\n
\n LB Score: 0.890
\n Running Time: ~14 min
May I ask how your eca_nfnet_l0 worked, i tried it but cannot convert it into openvion
", "votes": 1, "replies": [ { "id": 3196526, "author_name": "lhwcv", "author_url": "", "post_date": "2025-05-07T05:55:01.907000", "content": "I saw previous competition solution, they use OpenVINO for Efficientnet, and ONNX for eca_nfnet_l0
", "votes": 2, "replies": [] }, { "id": 3196538, "author_name": "Salman Ahmed", "author_url": "", "post_date": "2025-05-07T06:10:21.890000", "content": "eca_nfnet_l0 works with latest openvino
", "votes": 2, "replies": [] }, { "id": 3196543, "author_name": "HZM", "author_url": "", "post_date": "2025-05-07T06:15:35.833000", "content": "thanks so mcuh for your reply
", "votes": 1, "replies": [] }, { "id": 3200780, "author_name": "MYSO", "author_url": "", "post_date": "2025-05-13T03:54:10.197000", "content": "I'm afraid I haven't tried OpenVINO yet.
", "votes": 0, "replies": [] } ] }, { "id": 3200784, "author_name": "MYSO", "author_url": "", "post_date": "2025-05-13T03:59:32.040000", "content": "Update
\nMy best single-model submission:
\nModel: SED architecture (tf_efficientnetv2_s.in21k)
\nLB Score: 0.903
\nRunning Time: ~12 min
do you train on different segment length compared with when you predict?
", "votes": 0, "replies": [] }, { "id": 3200974, "author_name": "MYSO", "author_url": "", "post_date": "2025-05-13T09:31:22.570000", "content": "Yes, I do use different segment lengths
", "votes": 2, "replies": [] } ] } ] }, { "id": 3161810, "author_name": "Konstantin Dmitriev", "author_url": "", "post_date": "2025-03-28T12:18:35.970000", "content": "LB=0.834, Running Time ~6 min, single model without any ensembling.
", "votes": 3, "replies": [ { "id": 3161814, "author_name": "Starry", "author_url": "", "post_date": "2025-03-28T12:24:59.967000", "content": "Wow~ very efficient model. Is it efficientnetb0?
", "votes": 0, "replies": [ { "id": 3161844, "author_name": "Konstantin Dmitriev", "author_url": "", "post_date": "2025-03-28T13:30:29.240000", "content": "It's some kind of custom 'Frankenstein-like' architecture.
\nI'm using ThreadPoolExecutor for parallel processing of the data. However, I don't use quantization and inference accelerators like ONNX or OpenVINO since we have 90 min and it seems more than enough.
Update, LB:0.902 and ~20min running time
", "votes": 1, "replies": [ { "id": 3209840, "author_name": "lhwcv", "author_url": "", "post_date": "2025-05-26T11:50:25.727000", "content": "SED or CNN？
", "votes": 0, "replies": [ { "id": 3210315, "author_name": "Starry", "author_url": "", "post_date": "2025-05-27T03:12:06.867000", "content": "SED with efv2b3 backbone
", "votes": 1, "replies": [] } ] }, { "id": 3209963, "author_name": "HZM", "author_url": "", "post_date": "2025-05-26T14:56:32.603000", "content": "same results
", "votes": 0, "replies": [] } ] }, { "id": 3174342, "author_name": "Salman Ahmed", "author_url": "", "post_date": "2025-04-09T02:09:51.297000", "content": "Finally a single model EfficientNet B3 (LB 0.864). 😪
\nTrain Data Only.
So, now your best score (0.872) is a model ensemble version? I haven't tried ensemble yet. Try to improve my best single model but it has been stuck at 0.849 for two weeks already. 😴
", "votes": 0, "replies": [ { "id": 3174361, "author_name": "Salman Ahmed", "author_url": "", "post_date": "2025-04-09T02:39:37.953000", "content": "Unfortunately the ensemble result didn’t change by adding this model or removing other models.
\nMy best LB score is with an ensemble of 2 x B0 and 2 x B3 (scores 0.854, 0.858, 0.856, 0.859) and ensemble scores 0.870 with post processing 0.872.
", "votes": 3, "replies": [] }, { "id": 3174366, "author_name": "Starry", "author_url": "", "post_date": "2025-04-09T02:54:10.483000", "content": "I just submit my first ensemble version and it has finished running few minutes ago. (seresnext, efv2b3, ecanfnetl0) and it scored 0.846, which is slight worse than my efv2b3 model (0.849) . I expect it would be higher than 0.849 but it is not. Seems like there maybe some bug in my seresnext or ecanfnetl0 model (I have not submit them yet).
", "votes": 0, "replies": [] }, { "id": 3174369, "author_name": "Salman Ahmed", "author_url": "", "post_date": "2025-04-09T02:59:51.717000", "content": "I see,
\nEnsemble with logits was not working for me in this competition against my experiments.
\nSo, I had to ensemble after sigmoid.
\nNot sure, if that's the case for you.
Thank you for your advice~ I do use the probility instead of logits to ensemble model, so there maybe some other problem I guess.
", "votes": 0, "replies": [] }, { "id": 3176007, "author_name": "Konstantin Dmitriev", "author_url": "", "post_date": "2025-04-10T20:04:47.813000", "content": "I think, it's quite reasonable that ensembling with logits doesn't work, because the logits of each model in the ensemble may have different scales. For example, suppose we have multiplied the model's logits by some factor, i.e. 10. The ROCAUC score of this particular model remains the same. However, the weight of the model in the ensemble is increased a lot. By applying the sigmoid we put the model's output into the uniform scale, and all the models play equivalent role in the ensemble.
", "votes": 0, "replies": [] }, { "id": 3182144, "author_name": "Gabriel Preda", "author_url": "", "post_date": "2025-04-18T20:00:15.913000", "content": "Each model in the ensemble is 1 fold or you are using multiple folds/model (e.g. 5 folds/model)?
", "votes": 0, "replies": [] } ] }, { "id": 3175194, "author_name": "Zhongkai Shangguan", "author_url": "", "post_date": "2025-04-09T20:54:48.617000", "content": "Thanks for sharing! Would you mind clarifying what Train Data Only setting refers to? From your public notebook last year, I noticed that you used 5-fold cross-validation and only included the first 5 seconds of each audio sample. Does this approach correspond to the Train Data Only setting, and the score is from one of the five folds? Thank you!
by Train Data Only he means he is not using the train_soundscapes currently. I another post of his he mention that for training he is using random 5s clip from the track and for testing he is using the first 5s clip from a track.
It means, no train_soundscapes.
\nIt's random 5 seconds, first 5 seconds are performing much worse.
Thank you!
", "votes": 0, "replies": [] } ] } ] }, { "id": 3177766, "author_name": "Devasy Patel23", "author_url": "", "post_date": "2025-04-13T08:38:39.683000", "content": "Hi everyone, interesting discussion on single model performance!
\nI'm feeling a bit stuck myself. I've been working off a baseline and haven't managed to push the score much beyond that yet (around the low 0.8x range). I tried a few things like using a 3-channel input (mel spec + MFCCs + delta MFCCs) but that didn't seem to help, unfortunately.
\nLooking at the summary of past top solutions (like the one I compiled kaggle discussion), it's clear that pseudo-labeling on the unlabeled soundscapes was almost universally beneficial last year. Handling secondary labels effectively also seems important.
\nMy plan is to try implementing pseudo-labeling next, and then explore ideas like weighted loss for secondary labels or maybe even triplet loss if I get adventurous.
\nGiven that I'm currently plateauing with the basics, where would you suggest focusing efforts right now for the potentially biggest impact? Is diving straight into pseudo-labelling the best next step, or should I perhaps revisit data augmentation, model architecture choices, or something else first?
\nAppreciate any insights from those who've managed to break through similar plateaus!
", "votes": 0, "replies": [ { "id": 3177840, "author_name": "Salman Ahmed", "author_url": "", "post_date": "2025-04-13T10:19:56.157000", "content": "\nI just created a post, to give you some ideas. https://www.kaggle.com/competitions/birdclef-2025/discussion/573066
", "votes": 0, "replies": [ { "id": 3178154, "author_name": "Devasy Patel23", "author_url": "", "post_date": "2025-04-13T19:35:38.490000", "content": "Hey @salmanahmedtamu thanks for the effort, I read you post, I'll surely try Focal loss and other mentioned things next, Meanwhile I just tried CE loss as suggested by last year's 1st place team, but wasn't able to find much success, will play around with it for a while maybe try to tune the params a bit
", "votes": 0, "replies": [] } ] } ] }, { "id": 3179729, "author_name": "thacrobatheskis", "author_url": "", "post_date": "2025-04-15T16:16:47.527000", "content": "0.807, <1 min
\nI'm still trying to catch up to the public baselines & understand all the performance gaps, but hey at least I get fast feedback when submitting! All 60ish of my submissions have used GPU with a b0 model (see https://www.kaggle.com/code/robbynevels/bc25-infer-baseline-gpu/)
", "votes": -1, "replies": [] }, { "id": 3202890, "author_name": "Starry", "author_url": "", "post_date": "2025-05-16T04:07:55.877000", "content": "Update: LB: 0.889, Running Time: ~20min
", "votes": 0, "replies": [] }, { "id": 3174971, "author_name": "Starry", "author_url": "", "post_date": "2025-04-09T16:09:02.743000", "content": "Update. Best model scored 0.860, running time ~19min
", "votes": 0, "replies": [ { "id": 3176066, "author_name": "Salman Ahmed", "author_url": "", "post_date": "2025-04-10T23:40:30.753000", "content": "Is it SED or CNN?
\nI haven't tried SED yet.
SED， I haven't tried CNN
", "votes": 0, "replies": [] } ] }, { "id": 3176275, "author_name": "Hicham Bellafkir", "author_url": "", "post_date": "2025-04-11T07:19:47.913000", "content": "My best model scored 85.5 in 13 minutes, but the training is unstable, small changes lead to drastic fluctuations in the score. Model soup helped improve the score but didn’t contribute to training stability.
", "votes": 0, "replies": [] } ] }, { "id": 3162157, "author_name": "Fir_las_47", "author_url": "", "post_date": "2025-03-28T21:30:00.063000", "content": "16 and 23 minutes? Is this efficientnet b0?
", "votes": 0, "replies": [ { "id": 3162247, "author_name": "Starry", "author_url": "", "post_date": "2025-03-29T02:04:26.830000", "content": "Efficientnetv2b3 backbone with SED model architecture
", "votes": 2, "replies": [] } ] }, { "id": 3161801, "author_name": "Yurnero", "author_url": "", "post_date": "2025-03-28T11:59:18.593000", "content": "What do you mean by a single model? N folds / full train / single fold?
", "votes": 0, "replies": [ { "id": 3161811, "author_name": "Starry", "author_url": "", "post_date": "2025-03-28T12:19:15.903000", "content": "Single fold or full train，I use all data to train without validation.
", "votes": 0, "replies": [] } ] }, { "id": 3180086, "author_name": "", "author_url": "", "post_date": "2025-04-16T06:05:21.967000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "3161673": " LB RunningTime \n 0.848 ~23min \n 0.842 ~16min \n\nwhat's your best single model scored and how long it take?\n", "3192450": "Update: Single best model efficientnet_b0 [Running Time: ~10 Mins]\n\nLB: 0.90\n\nTrain Audio + Train soundscapes", "3164424": "LB = 0.858 Efficientnet B0 (Openvino) Running Time Around 13 Minutes - No Threads, Simple Loop based submission for now.\n\nLB = 0.859 Efficientnet B3 (Openvino) Running Time Around 25 Minutes - No Threads, Simple Loop based submission for now.\n\n\nModels are really unstable, and highly depends on batch distribution in mixup. ", "3180715": "Finally, I have a single CNN Efficientnet-B0 with LB: 0.875. (Running Time ~ 15 Mins).\n\n\nI have tried a lot of ways to use pseudo labels, but I don't think it's helping at all, it just lowers the LB. ", "3178587": "0.845 b0 tuning public notebook hyper-parameters only.", "3191953": "My best single‐model submission:\n\nModel: SED architecture (eca_nfnet_l0)\nLB Score: 0.890\nRunning Time: ~14 min", "3161810": "LB=0.834, Running Time ~6 min, single model without any ensembling.", "3209141": "Update, LB:0.902 and ~20min running time", "3174342": "Finally a single model EfficientNet B3 (LB 0.864). 😪\nTrain Data Only.", "3177766": "Hi everyone, interesting discussion on single model performance!\n\nI'm feeling a bit stuck myself. I've been working off a baseline and haven't managed to push the score much beyond that yet (around the low 0.8x range). I tried a few things like using a 3-channel input (mel spec + MFCCs + delta MFCCs) but that didn't seem to help, unfortunately.\n\nLooking at the summary of past top solutions (like the one I compiled [kaggle discussion](https://www.kaggle.com/competitions/birdclef-2025/discussion/572928)), it's clear that pseudo-labeling on the unlabeled soundscapes was almost universally beneficial last year. Handling secondary labels effectively also seems important.\n\nMy plan is to try implementing pseudo-labeling next, and then explore ideas like weighted loss for secondary labels or maybe even triplet loss if I get adventurous.\n\nGiven that I'm currently plateauing with the basics, where would you suggest focusing efforts *right now* for the potentially biggest impact? Is diving straight into pseudo-labelling the best next step, or should I perhaps revisit data augmentation, model architecture choices, or something else first?\n\nAppreciate any insights from those who've managed to break through similar plateaus!", "3179729": "0.807, <1 min \n\nI'm still trying to catch up to the public baselines & understand all the performance gaps, but hey at least I get fast feedback when submitting! All 60ish of my submissions have used GPU with a b0 model (see https://www.kaggle.com/code/robbynevels/bc25-infer-baseline-gpu/)", "3202890": "Update: LB: 0.889, Running Time: ~20min", "3174971": "Update. Best model scored 0.860, running time ~19min", "3162157": "16 and 23 minutes? Is this efficientnet b0?", "3161801": "What do you mean by a single model? **N** folds / full train / single fold?", "3180086": "" } }