I don't expect you to tell me what you did that worked. I'm more curious about if others are finding ways to use the unlabeled soundscapes that actually improves LB. Actually an anonymous poll feature would be cool.
", "messages": [ { "id": "2840005", "postDate": "05/27/2024 22:02:17", "content": "I don't expect you to tell me what you did that worked. I'm more curious about if others are finding ways to use the unlabeled soundscapes that actually improves LB. Actually an anonymous poll feature would be cool.
", "rawMarkdown": "I don't expect you to tell me what you did that worked. I'm more curious about if others are finding ways to use the unlabeled soundscapes that actually improves LB. Actually an anonymous poll feature would be cool.", "votes": null }, { "id": "2840139", "postDate": "05/28/2024 01:44:46", "content": "I tried to find nocall in the unlabelled soundscape, then add it as noise during training. It didn't help on my first try judging from this morning result (CV didn't change much as well)
", "rawMarkdown": "I tried to find nocall in the unlabelled soundscape, then add it as noise during training. It didn't help on my first try judging from this morning result (CV didn't change much as well)", "votes": null }, { "id": "2840168", "postDate": "05/28/2024 02:06:57", "content": "I did same with you. LB decrease significantly.
", "rawMarkdown": "I did same with you. LB decrease significantly.", "votes": null }, { "id": "2840169", "postDate": "05/28/2024 02:08:03", "content": "\"LB decrease significantly\" the same
", "rawMarkdown": "\"LB decrease significantly\" the same", "votes": null }, { "id": "2840188", "postDate": "05/28/2024 02:34:41", "content": "the same it decrease too much :))
", "rawMarkdown": "the same it decrease too much :))", "votes": null }, { "id": "2841089", "postDate": "05/28/2024 11:53:28", "content": "I preprocess the unlabelled soundscape with low/highpass filters (<100hz, >1000hz) and use that as background noise when training, but LB doesn't improve. Then I removed these filters and train again, same result. I set a random SNR around (0.05-4), p=0.5 and random 5s sampling from the \"noise dataset\" for both training. May be my method is wrong? IDK, my experiment logs are all over the place. 🤔
", "rawMarkdown": "I preprocess the unlabelled soundscape with low/highpass filters (<100hz, >1000hz) and use that as background noise when training, but LB doesn't improve. Then I removed these filters and train again, same result. I set a random SNR around (0.05-4), p=0.5 and random 5s sampling from the \"noise dataset\" for both training. May be my method is wrong? IDK, my experiment logs are all over the place. 🤔", "votes": null }, { "id": "2841090", "postDate": "05/28/2024 11:54:13", "content": "I tried pretraining on labels that my best model generated and that didn't help. I also tried nocall for background noise which didn't help either. I'm pretty confident in my nocall because I've listened to large portion of them and don't hear any obvious birds.
", "rawMarkdown": "I tried pretraining on labels that my best model generated and that didn't help. I also tried nocall for background noise which didn't help either. I'm pretty confident in my nocall because I've listened to large portion of them and don't hear any obvious birds.", "votes": null }, { "id": "2841286", "postDate": "05/28/2024 13:56:50", "content": "I think that the major problem is that some crickets are too similar to birds on spectrograms and when we push model not to learn features from unlabeled nocall we suppress the necessary for some species learning. So it should be taken for granted that we are going to mistake some crickets with birds.
", "rawMarkdown": "I think that the major problem is that some crickets are too similar to birds on spectrograms and when we push model not to learn features from unlabeled nocall we suppress the necessary for some species learning. So it should be taken for granted that we are going to mistake some crickets with birds.", "votes": null }, { "id": "2841319", "postDate": "05/28/2024 14:22:12", "content": "Not for me for some reason it decreased but not by much lb (from 0.65 - 0.65 but bit worse)
", "rawMarkdown": "Not for me for some reason it decreased but not by much lb (from 0.65 - 0.65 but bit worse)", "votes": null }, { "id": "2841321", "postDate": "05/28/2024 14:23:12", "content": "That is probably true and very interesting.
", "rawMarkdown": "That is probably true and very interesting.", "votes": null }, { "id": "2841322", "postDate": "05/28/2024 14:23:17", "content": "I tried using unlabeled as noise during training but it did not help my model it kept it basically the same.
", "rawMarkdown": "I tried using unlabeled as noise during training but it did not help my model it kept it basically the same.", "votes": null }, { "id": "2853803", "postDate": "06/04/2024 01:25:58", "content": "I have used the unlabeled dataset together with a pseudo-labeling strategy. It hasn't improved my top score, but I have noticed the variance between submissions has reduced, so I think this is a small victory. It was a lot of work for quite a small gain though, so I'm still puzzled why the top performing notebooks are so far ahead.
\nWith regards to backgrounds, you might want to consider using the past competitions soundscapes, since they only have one overlapping bird class. I am doing this too, but haven't noticed any score improvement from it. See this post from Stephan Kahl
", "rawMarkdown": "I have used the unlabeled dataset together with a pseudo-labeling strategy. It hasn't improved my top score, but I have noticed the variance between submissions has reduced, so I think this is a small victory. It was a lot of work for quite a small gain though, so I'm still puzzled why the top performing notebooks are so far ahead.\n\nWith regards to backgrounds, you might want to consider using the past competitions soundscapes, since they only have one overlapping bird class. I am doing this too, but haven't noticed any score improvement from it. [See this post](https://www.kaggle.com/competitions/birdclef-2023/discussion/394358#2179605) from Stephan Kahl", "votes": null } ], "comments": [ { "id": 2840139, "author_name": "nyleve", "author_url": "", "post_date": "05/28/2024 01:44:46", "content": "I tried to find nocall in the unlabelled soundscape, then add it as noise during training. It didn't help on my first try judging from this morning result (CV didn't change much as well)
", "votes": null, "replies": [ { "id": 2840168, "author_name": "quan0095", "author_url": "", "post_date": "05/28/2024 02:06:57", "content": "I did same with you. LB decrease significantly.
", "votes": null, "replies": [ { "id": 2840169, "author_name": "yuanzhezhou", "author_url": "", "post_date": "05/28/2024 02:08:03", "content": "\"LB decrease significantly\" the same
", "votes": null, "replies": [ { "id": 2840188, "author_name": "sentrankim", "author_url": "", "post_date": "05/28/2024 02:34:41", "content": "the same it decrease too much :))
", "votes": null, "replies": [ { "id": 2841319, "author_name": "max1mum", "author_url": "", "post_date": "05/28/2024 14:22:12", "content": "Not for me for some reason it decreased but not by much lb (from 0.65 - 0.65 but bit worse)
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 2841090, "author_name": "willrice", "author_url": "", "post_date": "05/28/2024 11:54:13", "content": "I tried pretraining on labels that my best model generated and that didn't help. I also tried nocall for background noise which didn't help either. I'm pretty confident in my nocall because I've listened to large portion of them and don't hear any obvious birds.
", "votes": null, "replies": [ { "id": 2841286, "author_name": "nikitababich", "author_url": "", "post_date": "05/28/2024 13:56:50", "content": "I think that the major problem is that some crickets are too similar to birds on spectrograms and when we push model not to learn features from unlabeled nocall we suppress the necessary for some species learning. So it should be taken for granted that we are going to mistake some crickets with birds.
", "votes": null, "replies": [ { "id": 2841321, "author_name": "max1mum", "author_url": "", "post_date": "05/28/2024 14:23:12", "content": "That is probably true and very interesting.
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 2841089, "author_name": "sakurayuyuko", "author_url": "", "post_date": "05/28/2024 11:53:28", "content": "I preprocess the unlabelled soundscape with low/highpass filters (<100hz, >1000hz) and use that as background noise when training, but LB doesn't improve. Then I removed these filters and train again, same result. I set a random SNR around (0.05-4), p=0.5 and random 5s sampling from the \"noise dataset\" for both training. May be my method is wrong? IDK, my experiment logs are all over the place. 🤔
", "votes": null, "replies": [] }, { "id": 2841322, "author_name": "max1mum", "author_url": "", "post_date": "05/28/2024 14:23:17", "content": "I tried using unlabeled as noise during training but it did not help my model it kept it basically the same.
", "votes": null, "replies": [] }, { "id": 2853803, "author_name": "ollypowell", "author_url": "", "post_date": "06/04/2024 01:25:58", "content": "I have used the unlabeled dataset together with a pseudo-labeling strategy. It hasn't improved my top score, but I have noticed the variance between submissions has reduced, so I think this is a small victory. It was a lot of work for quite a small gain though, so I'm still puzzled why the top performing notebooks are so far ahead.
\nWith regards to backgrounds, you might want to consider using the past competitions soundscapes, since they only have one overlapping bird class. I am doing this too, but haven't noticed any score improvement from it. See this post from Stephan Kahl
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2840005": "I don't expect you to tell me what you did that worked. I'm more curious about if others are finding ways to use the unlabeled soundscapes that actually improves LB. Actually an anonymous poll feature would be cool.", "2840139": "I tried to find nocall in the unlabelled soundscape, then add it as noise during training. It didn't help on my first try judging from this morning result (CV didn't change much as well)", "2840168": "I did same with you. LB decrease significantly.", "2840169": "\"LB decrease significantly\" the same", "2840188": "the same it decrease too much :))", "2841089": "I preprocess the unlabelled soundscape with low/highpass filters (<100hz, >1000hz) and use that as background noise when training, but LB doesn't improve. Then I removed these filters and train again, same result. I set a random SNR around (0.05-4), p=0.5 and random 5s sampling from the \"noise dataset\" for both training. May be my method is wrong? IDK, my experiment logs are all over the place. 🤔", "2841090": "I tried pretraining on labels that my best model generated and that didn't help. I also tried nocall for background noise which didn't help either. I'm pretty confident in my nocall because I've listened to large portion of them and don't hear any obvious birds.", "2841286": "I think that the major problem is that some crickets are too similar to birds on spectrograms and when we push model not to learn features from unlabeled nocall we suppress the necessary for some species learning. So it should be taken for granted that we are going to mistake some crickets with birds.", "2841319": "Not for me for some reason it decreased but not by much lb (from 0.65 - 0.65 but bit worse)", "2841321": "That is probably true and very interesting.", "2841322": "I tried using unlabeled as noise during training but it did not help my model it kept it basically the same.", "2853803": "I have used the unlabeled dataset together with a pseudo-labeling strategy. It hasn't improved my top score, but I have noticed the variance between submissions has reduced, so I think this is a small victory. It was a lot of work for quite a small gain though, so I'm still puzzled why the top performing notebooks are so far ahead.\n\nWith regards to backgrounds, you might want to consider using the past competitions soundscapes, since they only have one overlapping bird class. I am doing this too, but haven't noticed any score improvement from it. [See this post](https://www.kaggle.com/competitions/birdclef-2023/discussion/394358#2179605) from Stephan Kahl" }, "source": "meta" }