For me,
\nFor me,
\nfor my experiments
\none fold max score 0.80+ with CV 0.97. - will try SED Thanks.
\n@lihaoweicvch is local CV matching with LB? whats your best local CV/LB?
", "rawMarkdown": "for my experiments\n- CE better than BCE.\n- Pseudo labels improved. \n- random crop without human voice improved.\n\none fold max score 0.80+ with CV 0.97. - will try SED Thanks.\n\n@lihaoweicvch is local CV matching with LB? whats your best local CV/LB?", "votes": null }, { "id": "3194934", "postDate": "05/06/2025 12:47:33", "content": "Not matching，so I didn’t make local validation, CV 0.97, LB 0.85 That’s my very early version
", "rawMarkdown": "Not matching，so I didn’t make local validation, CV 0.97, LB 0.85 That’s my very early version", "votes": null }, { "id": "3195002", "postDate": "05/06/2025 14:14:43", "content": "Thank for sharing this.
\nI just did a single experiment with SED.
\nPseudo Label works.
\nI haven't tried CE, for me FocalBCE works.
Do you use all the dataset to train model instead of cross validation?
", "rawMarkdown": "Do you use all the dataset to train model instead of cross validation?", "votes": null }, { "id": "3195041", "postDate": "05/06/2025 15:08:06", "content": "\n\nPseudo Label is useful on some configurations
\n
For me, it's on some models
", "rawMarkdown": "> Pseudo Label is useful on some configurations\n\nFor me, it's on some models", "votes": null }, { "id": "3195087", "postDate": "05/06/2025 16:03:36", "content": "In my tests, CE usually outperforms BCE. I'm still trying out Focal BCE now.
", "rawMarkdown": "In my tests, CE usually outperforms BCE. I'm still trying out Focal BCE now.", "votes": null }, { "id": "3195330", "postDate": "05/06/2025 22:34:29", "content": "Cool. Please share if it worked for you or not.
", "rawMarkdown": "Cool. Please share if it worked for you or not.", "votes": null }, { "id": "3195345", "postDate": "05/06/2025 23:17:06", "content": "Yes I use all
", "rawMarkdown": "Yes I use all", "votes": null }, { "id": "3195387", "postDate": "05/07/2025 01:03:49", "content": "\nI have a question, when you created pseudo labels for unlabeled_soundscapes, how did you apply CE loss, if no specie was detected? Did you set a threshold to include or exclude those chunks?
\nIf multiple species are detected in unlabeled_soundscapes, did you normalized it to have the sum of 1 for passing in to CE loss?
", "rawMarkdown": "Hi @lihaoweicvch \n\nI have a question, when you created pseudo labels for unlabeled_soundscapes, how did you apply CE loss, if no specie was detected? Did you set a threshold to include or exclude those chunks?\n\nIf multiple species are detected in unlabeled_soundscapes, did you normalized it to have the sum of 1 for passing in to CE loss?", "votes": null }, { "id": "3195397", "postDate": "05/07/2025 01:41:01", "content": "Thanks for replying. Do you also use average model method? such as ema, swa, model soup……
", "rawMarkdown": "Thanks for replying. Do you also use average model method? such as ema, swa, model soup......", "votes": null }, { "id": "3196417", "postDate": "05/07/2025 02:50:52", "content": "I include those chunks which has species, exclude no species trunks but see them as background noise on some configuration.
\nI didn't normalized it to have the sum of 1.
\nwhat about yours?
not yet, try later
", "rawMarkdown": "not yet, try later", "votes": null }, { "id": "3196420", "postDate": "05/07/2025 02:56:27", "content": "I see, I tried CE earlier in the competition and then dropped the idea because it can produce noisy predictions.
\nI just started training again with a CE pipeline by normalizing labels to be 1. Let’s see how it works.
", "rawMarkdown": "I see, I tried CE earlier in the competition and then dropped the idea because it can produce noisy predictions.\n\nI just started training again with a CE pipeline by normalizing labels to be 1. Let’s see how it works.", "votes": null }, { "id": "3196483", "postDate": "05/07/2025 04:49:42", "content": "Well it scored 0.752 xD
", "rawMarkdown": "Well it scored 0.752 xD", "votes": null }, { "id": "3196487", "postDate": "05/07/2025 04:55:39", "content": "I use loss like this.
\n ():\n log_probs = torch.nn.functional.log_softmax(logits, dim=)\n loss = -torch.(soft_targets * log_probs, dim=)\n loss.mean()\n@salmanahmedtamu I just tried CNN + CE, its predictions really noisy, many false positives, now I'am trying FocalBCE to see if it's better for CNN.
", "rawMarkdown": "salmanahmedtamu I just tried CNN + CE, its predictions really noisy, many false positives, now I'am trying FocalBCE to see if it's better for CNN.", "votes": null }, { "id": "3196529", "postDate": "05/07/2025 05:57:58", "content": "I use FocalLossBCE from your notebook: https://www.kaggle.com/code/salmanahmedtamu/training-0-65-0-66
\nMay I ask your BCE and Focal weight? default is 1.0 and 1.0
Yes, both weights are set to 1
", "rawMarkdown": "Yes, both weights are set to 1", "votes": null }, { "id": "3196651", "postDate": "05/07/2025 09:12:21", "content": "Nice work Thanks for sharing this
", "rawMarkdown": "Nice work Thanks for sharing this", "votes": null }, { "id": "3197305", "postDate": "05/08/2025 03:30:45", "content": "How to build a SED model?
", "rawMarkdown": "How to build a SED model?", "votes": null }, { "id": "3197346", "postDate": "05/08/2025 05:01:25", "content": "https://www.kaggle.com/competitions/birdclef-2025/discussion/576765
", "rawMarkdown": "https://www.kaggle.com/competitions/birdclef-2025/discussion/576765", "votes": null }, { "id": "3198809", "postDate": "05/10/2025 02:58:46", "content": "Have you tried cleaning vocal segments and using FocalBCE? It doesn't work for me.
", "rawMarkdown": "Have you tried cleaning vocal segments and using FocalBCE? It doesn't work for me.", "votes": null }, { "id": "3198868", "postDate": "05/10/2025 05:15:23", "content": "Yup, that's the solution which scored 0.907 LB.
", "rawMarkdown": "Yup, that's the solution which scored 0.907 LB.", "votes": null }, { "id": "3198869", "postDate": "05/10/2025 05:16:25", "content": "You need to consider the noisy audios in the train set, specially for the CSA recordings.
", "rawMarkdown": "You need to consider the noisy audios in the train set, specially for the CSA recordings.", "votes": null }, { "id": "3199561", "postDate": "05/11/2025 06:11:24", "content": "does CE + Sigmoid works better than CE + softmax for you?
", "rawMarkdown": "does CE + Sigmoid works better than CE + softmax for you?", "votes": null }, { "id": "3199817", "postDate": "05/11/2025 13:54:17", "content": "I didn’t try CE + softmax (infer), I guess there maybe more false positives, I”ll try tmr and report to here. what about yours
", "rawMarkdown": "I didn’t try CE + softmax (infer), I guess there maybe more false positives, I”ll try tmr and report to here. what about yours", "votes": null }, { "id": "3199928", "postDate": "05/11/2025 17:59:04", "content": "I am not able to get more than 0.752 with CE + sigmoid. 🥲
", "rawMarkdown": "I am not able to get more than 0.752 with CE + sigmoid. 🥲", "votes": null }, { "id": "3199929", "postDate": "05/11/2025 18:00:42", "content": "Yeah, it might work, if you have nocall as well in your classes.
0.9 drop to 0.88 if using softmax for me
", "rawMarkdown": "0.9 drop to 0.88 if using softmax for me", "votes": null }, { "id": "3200447", "postDate": "05/12/2025 15:27:33", "content": "Hi @salmanahmedtamu! What is nocall here? Is it a fake class of having no bird calls in the current 5 sec interval made by preprocessing of the training data?
how to deal with length less 5s segments?😭
", "rawMarkdown": "how to deal with length less 5s segments?😭", "votes": null }, { "id": "3201626", "postDate": "05/14/2025 07:02:43", "content": "simply pad it to 5s or repeat that segment to 5s.
", "rawMarkdown": "simply pad it to 5s or repeat that segment to 5s.", "votes": null }, { "id": "3201655", "postDate": "05/14/2025 07:56:24", "content": "i do pad it to 5s its hurts pb
", "rawMarkdown": "i do pad it to 5s its hurts pb", "votes": null }, { "id": "3201859", "postDate": "05/14/2025 13:47:10", "content": "Could you please provide some more detailed information about sed?
", "rawMarkdown": "Could you please provide some more detailed information about sed?", "votes": null }, { "id": "3201870", "postDate": "05/14/2025 14:03:33", "content": "nothing worked for me after overfitting to 0.845, it seems to be overfitting very hard 🤔
", "rawMarkdown": "nothing worked for me after overfitting to 0.845, it seems to be overfitting very hard 🤔", "votes": null }, { "id": "3202144", "postDate": "05/14/2025 23:25:19", "content": "The following methods worked well for me:
\nHi, I tried Removing human voices too, But got bad rewards from pb score, May you provide mroe detail, thanks.
", "rawMarkdown": "Hi, I tried Removing human voices too, But got bad rewards from pb score, May you provide mroe detail, thanks.", "votes": null }, { "id": "3202179", "postDate": "05/15/2025 02:04:14", "content": "Is pink noise and log-mel spectrogram really helpful for you?
", "rawMarkdown": "Is pink noise and log-mel spectrogram really helpful for you?", "votes": null }, { "id": "3202203", "postDate": "05/15/2025 03:02:25", "content": "Did you delete the entire file if it contained human voices?
\nIt might be better to remove only the frames that contain human voices within the file.
What’s the performance difference between SED and CNN in your experiments?
", "rawMarkdown": "What’s the performance difference between SED and CNN in your experiments?", "votes": null }, { "id": "3202406", "postDate": "05/15/2025 10:19:12", "content": "Compared to a CNN baseline, the SED model improved the LB score by about 0.03 in my experiments.
", "rawMarkdown": "Compared to a CNN baseline, the SED model improved the LB score by about 0.03 in my experiments.", "votes": null }, { "id": "3202450", "postDate": "05/15/2025 11:38:12", "content": "Thank you. Is this SED with CE same as @lihaoweicvch mentioned?
\nIt's not working for me. 😅
I'm only using FocalLoss(BCE)
", "rawMarkdown": "I'm only using FocalLoss(BCE)", "votes": null }, { "id": "3202489", "postDate": "05/15/2025 13:18:00", "content": "Thanks for sharing.
", "rawMarkdown": "Thanks for sharing.", "votes": null }, { "id": "3202552", "postDate": "05/15/2025 14:45:17", "content": "I remove only the frames that contain human voices within the file
", "rawMarkdown": "I remove only the frames that contain human voices within the file", "votes": null }, { "id": "3203071", "postDate": "05/16/2025 09:12:35", "content": "The following methods worked for me
\nis it a single model?
", "rawMarkdown": "is it a single model?", "votes": null }, { "id": "3203335", "postDate": "05/16/2025 15:27:24", "content": "I get a slightly better score when promoting orthogonal matrices. Not sure if that still works when the performance is better than mine.
", "rawMarkdown": "I get a slightly better score when promoting orthogonal matrices. Not sure if that still works when the performance is better than mine.", "votes": null }, { "id": "3203337", "postDate": "05/16/2025 15:30:23", "content": "\"Post processing and predictions smoothing\" what do I need to google in order to find useful ideas for that?
", "rawMarkdown": "\"Post processing and predictions smoothing\" what do I need to google in order to find useful ideas for that?", "votes": null }, { "id": "3209255", "postDate": "05/25/2025 13:05:11", "content": "wow just to confirm: 0.03? or 0.003?
", "rawMarkdown": "wow just to confirm: 0.03? or 0.003?", "votes": null }, { "id": "3209262", "postDate": "05/25/2025 13:17:04", "content": "It’s 0.03, but that score comes from the very early stage of my experiments. After that, I used only the SED model.
", "rawMarkdown": "It’s 0.03, but that score comes from the very early stage of my experiments. After that, I used only the SED model.", "votes": null }, { "id": "3209266", "postDate": "05/25/2025 13:27:16", "content": "I see, that's a lot of improvement. I jsut start trying SED, but not too much improvement on our side.
", "rawMarkdown": "I see, that's a lot of improvement. I jsut start trying SED, but not too much improvement on our side.", "votes": null }, { "id": "3210155", "postDate": "05/26/2025 19:58:01", "content": "I haven't been able to get a boost from this either, but I'm still working on it.
", "rawMarkdown": "I haven't been able to get a boost from this either, but I'm still working on it.", "votes": null }, { "id": "3210156", "postDate": "05/26/2025 20:00:23", "content": "For the without human voice, did you just sample from regions that the VAD didn't detect as speech?
", "rawMarkdown": "For the without human voice, did you just sample from regions that the VAD didn't detect as speech?", "votes": null } ], "comments": [ { "id": 3194913, "author_name": "seshurajup", "author_url": "", "post_date": "05/06/2025 12:17:34", "content": "for my experiments
\none fold max score 0.80+ with CV 0.97. - will try SED Thanks.
\n@lihaoweicvch is local CV matching with LB? whats your best local CV/LB?
", "votes": null, "replies": [ { "id": 3194934, "author_name": "lihaoweicvch", "author_url": "", "post_date": "05/06/2025 12:47:33", "content": "Not matching，so I didn’t make local validation, CV 0.97, LB 0.85 That’s my very early version
", "votes": null, "replies": [] }, { "id": 3210156, "author_name": "willrice", "author_url": "", "post_date": "05/26/2025 20:00:23", "content": "For the without human voice, did you just sample from regions that the VAD didn't detect as speech?
", "votes": null, "replies": [] } ] }, { "id": 3195002, "author_name": "salmanahmedtamu", "author_url": "", "post_date": "05/06/2025 14:14:43", "content": "Thank for sharing this.
\nI just did a single experiment with SED.
\nPseudo Label works.
\nI haven't tried CE, for me FocalBCE works.
Do you use all the dataset to train model instead of cross validation?
", "votes": null, "replies": [ { "id": 3195345, "author_name": "lihaoweicvch", "author_url": "", "post_date": "05/06/2025 23:17:06", "content": "Yes I use all
", "votes": null, "replies": [ { "id": 3195397, "author_name": "i2nfinit3y", "author_url": "", "post_date": "05/07/2025 01:41:01", "content": "Thanks for replying. Do you also use average model method? such as ema, swa, model soup……
", "votes": null, "replies": [ { "id": 3196418, "author_name": "lihaoweicvch", "author_url": "", "post_date": "05/07/2025 02:51:21", "content": "not yet, try later
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 3195041, "author_name": "snorfyang", "author_url": "", "post_date": "05/06/2025 15:08:06", "content": "\n\nPseudo Label is useful on some configurations
\n
For me, it's on some models
", "votes": null, "replies": [] }, { "id": 3195087, "author_name": "digiranger", "author_url": "", "post_date": "05/06/2025 16:03:36", "content": "In my tests, CE usually outperforms BCE. I'm still trying out Focal BCE now.
", "votes": null, "replies": [ { "id": 3195330, "author_name": "salmanahmedtamu", "author_url": "", "post_date": "05/06/2025 22:34:29", "content": "Cool. Please share if it worked for you or not.
", "votes": null, "replies": [] } ] }, { "id": 3195387, "author_name": "salmanahmedtamu", "author_url": "", "post_date": "05/07/2025 01:03:49", "content": "\nI have a question, when you created pseudo labels for unlabeled_soundscapes, how did you apply CE loss, if no specie was detected? Did you set a threshold to include or exclude those chunks?
\nIf multiple species are detected in unlabeled_soundscapes, did you normalized it to have the sum of 1 for passing in to CE loss?
", "votes": null, "replies": [ { "id": 3196417, "author_name": "lihaoweicvch", "author_url": "", "post_date": "05/07/2025 02:50:52", "content": "I include those chunks which has species, exclude no species trunks but see them as background noise on some configuration.
\nI didn't normalized it to have the sum of 1.
\nwhat about yours?
I see, I tried CE earlier in the competition and then dropped the idea because it can produce noisy predictions.
\nI just started training again with a CE pipeline by normalizing labels to be 1. Let’s see how it works.
", "votes": null, "replies": [] }, { "id": 3196483, "author_name": "salmanahmedtamu", "author_url": "", "post_date": "05/07/2025 04:49:42", "content": "Well it scored 0.752 xD
", "votes": null, "replies": [] }, { "id": 3196487, "author_name": "salmanahmedtamu", "author_url": "", "post_date": "05/07/2025 04:55:39", "content": "I use loss like this.
\n ():\n log_probs = torch.nn.functional.log_softmax(logits, dim=)\n loss = -torch.(soft_targets * log_probs, dim=)\n loss.mean()\n@salmanahmedtamu I just tried CNN + CE, its predictions really noisy, many false positives, now I'am trying FocalBCE to see if it's better for CNN.
", "votes": null, "replies": [ { "id": 3196529, "author_name": "lihaoweicvch", "author_url": "", "post_date": "05/07/2025 05:57:58", "content": "I use FocalLossBCE from your notebook: https://www.kaggle.com/code/salmanahmedtamu/training-0-65-0-66
\nMay I ask your BCE and Focal weight? default is 1.0 and 1.0
Yes, both weights are set to 1
", "votes": null, "replies": [] } ] } ] } ] } ] } ] }, { "id": 3196651, "author_name": "sayakmukherjee05", "author_url": "", "post_date": "05/07/2025 09:12:21", "content": "Nice work Thanks for sharing this
", "votes": null, "replies": [] }, { "id": 3197305, "author_name": "chenbaoying", "author_url": "", "post_date": "05/08/2025 03:30:45", "content": "How to build a SED model?
", "votes": null, "replies": [ { "id": 3197346, "author_name": "lihaoweicvch", "author_url": "", "post_date": "05/08/2025 05:01:25", "content": "https://www.kaggle.com/competitions/birdclef-2025/discussion/576765
", "votes": null, "replies": [] } ] }, { "id": 3198809, "author_name": "shenshiyun", "author_url": "", "post_date": "05/10/2025 02:58:46", "content": "Have you tried cleaning vocal segments and using FocalBCE? It doesn't work for me.
", "votes": null, "replies": [ { "id": 3198868, "author_name": "salmanahmedtamu", "author_url": "", "post_date": "05/10/2025 05:15:23", "content": "Yup, that's the solution which scored 0.907 LB.
", "votes": null, "replies": [] }, { "id": 3198869, "author_name": "salmanahmedtamu", "author_url": "", "post_date": "05/10/2025 05:16:25", "content": "You need to consider the noisy audios in the train set, specially for the CSA recordings.
", "votes": null, "replies": [] } ] }, { "id": 3199561, "author_name": "salmanahmedtamu", "author_url": "", "post_date": "05/11/2025 06:11:24", "content": "does CE + Sigmoid works better than CE + softmax for you?
", "votes": null, "replies": [ { "id": 3199817, "author_name": "lihaoweicvch", "author_url": "", "post_date": "05/11/2025 13:54:17", "content": "I didn’t try CE + softmax (infer), I guess there maybe more false positives, I”ll try tmr and report to here. what about yours
", "votes": null, "replies": [ { "id": 3199928, "author_name": "salmanahmedtamu", "author_url": "", "post_date": "05/11/2025 17:59:04", "content": "I am not able to get more than 0.752 with CE + sigmoid. 🥲
", "votes": null, "replies": [] }, { "id": 3199929, "author_name": "salmanahmedtamu", "author_url": "", "post_date": "05/11/2025 18:00:42", "content": "Yeah, it might work, if you have nocall as well in your classes.
0.9 drop to 0.88 if using softmax for me
", "votes": null, "replies": [] }, { "id": 3200447, "author_name": "veshkinartem", "author_url": "", "post_date": "05/12/2025 15:27:33", "content": "Hi @salmanahmedtamu! What is nocall here? Is it a fake class of having no bird calls in the current 5 sec interval made by preprocessing of the training data?
how to deal with length less 5s segments?😭
", "votes": null, "replies": [ { "id": 3201626, "author_name": "fangsionfang", "author_url": "", "post_date": "05/14/2025 07:02:43", "content": "simply pad it to 5s or repeat that segment to 5s.
", "votes": null, "replies": [ { "id": 3201655, "author_name": "aichangeworld", "author_url": "", "post_date": "05/14/2025 07:56:24", "content": "i do pad it to 5s its hurts pb
", "votes": null, "replies": [] } ] } ] }, { "id": 3201859, "author_name": "yandsbnb666zhang", "author_url": "", "post_date": "05/14/2025 13:47:10", "content": "Could you please provide some more detailed information about sed?
", "votes": null, "replies": [] }, { "id": 3201870, "author_name": "yuanzhezhou", "author_url": "", "post_date": "05/14/2025 14:03:33", "content": "nothing worked for me after overfitting to 0.845, it seems to be overfitting very hard 🤔
", "votes": null, "replies": [] }, { "id": 3202144, "author_name": "myso1987", "author_url": "", "post_date": "05/14/2025 23:25:19", "content": "The following methods worked well for me:
\nHi, I tried Removing human voices too, But got bad rewards from pb score, May you provide mroe detail, thanks.
", "votes": null, "replies": [ { "id": 3202203, "author_name": "myso1987", "author_url": "", "post_date": "05/15/2025 03:02:25", "content": "Did you delete the entire file if it contained human voices?
\nIt might be better to remove only the frames that contain human voices within the file.
I remove only the frames that contain human voices within the file
", "votes": null, "replies": [ { "id": 3210155, "author_name": "willrice", "author_url": "", "post_date": "05/26/2025 19:58:01", "content": "I haven't been able to get a boost from this either, but I'm still working on it.
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 3202215, "author_name": "salmanahmedtamu", "author_url": "", "post_date": "05/15/2025 03:39:15", "content": "What’s the performance difference between SED and CNN in your experiments?
", "votes": null, "replies": [ { "id": 3202406, "author_name": "myso1987", "author_url": "", "post_date": "05/15/2025 10:19:12", "content": "Compared to a CNN baseline, the SED model improved the LB score by about 0.03 in my experiments.
", "votes": null, "replies": [ { "id": 3202450, "author_name": "salmanahmedtamu", "author_url": "", "post_date": "05/15/2025 11:38:12", "content": "Thank you. Is this SED with CE same as @lihaoweicvch mentioned?
\nIt's not working for me. 😅
I'm only using FocalLoss(BCE)
", "votes": null, "replies": [ { "id": 3202489, "author_name": "salmanahmedtamu", "author_url": "", "post_date": "05/15/2025 13:18:00", "content": "Thanks for sharing.
", "votes": null, "replies": [] } ] } ] }, { "id": 3209255, "author_name": "leonshangguan", "author_url": "", "post_date": "05/25/2025 13:05:11", "content": "wow just to confirm: 0.03? or 0.003?
", "votes": null, "replies": [ { "id": 3209262, "author_name": "myso1987", "author_url": "", "post_date": "05/25/2025 13:17:04", "content": "It’s 0.03, but that score comes from the very early stage of my experiments. After that, I used only the SED model.
", "votes": null, "replies": [ { "id": 3209266, "author_name": "leonshangguan", "author_url": "", "post_date": "05/25/2025 13:27:16", "content": "I see, that's a lot of improvement. I jsut start trying SED, but not too much improvement on our side.
", "votes": null, "replies": [] } ] } ] } ] } ] } ] }, { "id": 3202179, "author_name": "shenshiyun", "author_url": "", "post_date": "05/15/2025 02:04:14", "content": "Is pink noise and log-mel spectrogram really helpful for you?
", "votes": null, "replies": [] }, { "id": 3203071, "author_name": "sayedathar11", "author_url": "", "post_date": "05/16/2025 09:12:35", "content": "The following methods worked for me
\nis it a single model?
", "votes": null, "replies": [] }, { "id": 3203337, "author_name": "tim6502", "author_url": "", "post_date": "05/16/2025 15:30:23", "content": "\"Post processing and predictions smoothing\" what do I need to google in order to find useful ideas for that?
", "votes": null, "replies": [] } ] }, { "id": 3203335, "author_name": "tim6502", "author_url": "", "post_date": "05/16/2025 15:27:24", "content": "I get a slightly better score when promoting orthogonal matrices. Not sure if that still works when the performance is better than mine.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3194889": "For me, \n1. SED works better than CNN\n2. Pseudo Label is useful on some configurations\n3. random crop better than first 5/10 seconds \n4. CE better than BCE", "3194913": "for my experiments\n- CE better than BCE.\n- Pseudo labels improved. \n- random crop without human voice improved.\n\none fold max score 0.80+ with CV 0.97. - will try SED Thanks.\n\n@lihaoweicvch is local CV matching with LB? whats your best local CV/LB?", "3194934": "Not matching，so I didn’t make local validation, CV 0.97, LB 0.85 That’s my very early version", "3195002": "Thank for sharing this. \n\n\nI just did a single experiment with SED. \nPseudo Label works.\nI haven't tried CE, for me FocalBCE works.", "3195037": "Do you use all the dataset to train model instead of cross validation?", "3195041": "> Pseudo Label is useful on some configurations\n\nFor me, it's on some models", "3195087": "In my tests, CE usually outperforms BCE. I'm still trying out Focal BCE now.", "3195330": "Cool. Please share if it worked for you or not.", "3195345": "Yes I use all", "3195387": "Hi @lihaoweicvch \n\nI have a question, when you created pseudo labels for unlabeled_soundscapes, how did you apply CE loss, if no specie was detected? Did you set a threshold to include or exclude those chunks?\n\nIf multiple species are detected in unlabeled_soundscapes, did you normalized it to have the sum of 1 for passing in to CE loss?", "3195397": "Thanks for replying. Do you also use average model method? such as ema, swa, model soup......", "3196417": "I include those chunks which has species, exclude no species trunks but see them as background noise on some configuration.\nI didn't normalized it to have the sum of 1.\nwhat about yours?", "3196418": "not yet, try later", "3196420": "I see, I tried CE earlier in the competition and then dropped the idea because it can produce noisy predictions.\n\nI just started training again with a CE pipeline by normalizing labels to be 1. Let’s see how it works.", "3196483": "Well it scored 0.752 xD", "3196487": "I use loss like this.\n\n```python\ndef soft_cross_entropy(logits, soft_targets):\n log_probs = torch.nn.functional.log_softmax(logits, dim=1)\n loss = -torch.sum(soft_targets * log_probs, dim=1)\n return loss.mean()\n```", "3196522": "salmanahmedtamu I just tried CNN + CE, its predictions really noisy, many false positives, now I'am trying FocalBCE to see if it's better for CNN.", "3196529": "I use FocalLossBCE from your notebook: https://www.kaggle.com/code/salmanahmedtamu/training-0-65-0-66\nMay I ask your BCE and Focal weight? default is 1.0 and 1.0", "3196531": "Yes, both weights are set to 1", "3196651": "Nice work Thanks for sharing this", "3197305": "How to build a SED model?", "3197346": "https://www.kaggle.com/competitions/birdclef-2025/discussion/576765", "3198809": "Have you tried cleaning vocal segments and using FocalBCE? It doesn't work for me.", "3198868": "Yup, that's the solution which scored 0.907 LB.", "3198869": "You need to consider the noisy audios in the train set, specially for the CSA recordings.", "3199561": "does CE + Sigmoid works better than CE + softmax for you?", "3199817": "I didn’t try CE + softmax (infer), I guess there maybe more false positives, I”ll try tmr and report to here. what about yours", "3199928": "I am not able to get more than 0.752 with CE + sigmoid. 🥲", "3199929": "Yeah, it might work, if you have `nocall` as well in your classes.", "3200089": "0.9 drop to 0.88 if using softmax for me", "3200447": "Hi @salmanahmedtamu! What is `nocall` here? Is it a fake class of having no bird calls in the current 5 sec interval made by preprocessing of the training data?", "3201570": "how to deal with length less 5s segments?😭", "3201626": "simply pad it to 5s or repeat that segment to 5s.", "3201655": "i do pad it to 5s its hurts pb", "3201859": "Could you please provide some more detailed information about sed?", "3201870": "nothing worked for me after overfitting to 0.845, it seems to be overfitting very hard 🤔", "3202144": "The following methods worked well for me:\n\n- SED model\n- Removing human voices\n- Pseudo‑labeling\n- Post‑processing (see [notebook](https://www.kaggle.com/code/myso1987/post-processing-with-power-adjustment-for-low-rank))", "3202176": "Hi, I tried Removing human voices too, But got bad rewards from pb score, May you provide mroe detail, thanks.", "3202179": "Is pink noise and log-mel spectrogram really helpful for you?", "3202203": "Did you delete the entire file if it contained human voices?\nIt might be better to remove only the frames that contain human voices within the file.", "3202215": "What’s the performance difference between SED and CNN in your experiments?", "3202406": "Compared to a CNN baseline, the SED model improved the LB score by about 0.03 in my experiments.", "3202450": "Thank you. Is this SED with CE same as @lihaoweicvch mentioned?\nIt's not working for me. 😅", "3202462": "I'm only using FocalLoss(BCE)", "3202489": "Thanks for sharing.", "3202552": "I remove only the frames that contain human voices within the file", "3203071": "The following methods worked for me\n1. Removing human voices\n2. Smaller efficient net architecture like efficientnet b0 ns\n3. Pseudolabelling train soundscapes \n4. Post processing and predictions smoothing", "3203075": "is it a single model?", "3203335": "I get a slightly better score when promoting orthogonal matrices. Not sure if that still works when the performance is better than mine.", "3203337": "\"Post processing and predictions smoothing\" what do I need to google in order to find useful ideas for that?", "3209255": "wow just to confirm: 0.03? or 0.003?", "3209262": "It’s 0.03, but that score comes from the very early stage of my experiments. After that, I used only the SED model.", "3209266": "I see, that's a lot of improvement. I jsut start trying SED, but not too much improvement on our side.", "3210155": "I haven't been able to get a boost from this either, but I'm still working on it.", "3210156": "For the without human voice, did you just sample from regions that the VAD didn't detect as speech?" }, "source": "meta" }