The data for this competition is very special in the sense that the record are only partially labeled, and the labeled samples are done at a one-specy-basis, ie even if there is many species on the same [t_min, t_max], you may end up with only one (or some) of them labeled.
\nThis situation must require all of our attention. For instance, with the same ResNeSt50 model , I got a LB score of 0.66 when randomly cropping the audios. And guess what, my score jumped to 0.861 when doing smarter cropping. I'm definitely sure that, even with a simple model, one can win this competition with just a good data cropping / normalization strategy.
\nAnd you, does the cropping strategy impact your performance as much ?
", "messages": [ { "id": "1132688", "postDate": "12/30/2020 16:02:10", "content": "The data for this competition is very special in the sense that the record are only partially labeled, and the labeled samples are done at a one-specy-basis, ie even if there is many species on the same [t_min, t_max], you may end up with only one (or some) of them labeled.
\nThis situation must require all of our attention. For instance, with the same ResNeSt50 model , I got a LB score of 0.66 when randomly cropping the audios. And guess what, my score jumped to 0.861 when doing smarter cropping. I'm definitely sure that, even with a simple model, one can win this competition with just a good data cropping / normalization strategy.
\nAnd you, does the cropping strategy impact your performance as much ?
", "rawMarkdown": "The data for this competition is very special in the sense that the record are only partially labeled, and the labeled samples are done at a one-specy-basis, ie even if there is many species on the same **[t_min, t_max]**, you may end up with only one (or some) of them labeled.\n\nThis situation must require all of our attention. For instance, with the same **ResNeSt50** model , I got a LB score of **0.66** when randomly cropping the audios. And guess what, my score jumped to **0.861** when doing smarter cropping. I'm definitely sure that, even with a simple model, one can win this competition with just a good data cropping / normalization strategy.\n\nAnd you, does the cropping strategy impact your performance as much ?", "votes": null }, { "id": "1132794", "postDate": "12/30/2020 17:34:44", "content": "Amazing thinking!
\nA lot of models in the notebooks neglect this part!!!!
\nThey just simply crop the data according to the fmin fmax tmin tmax..
But how should we crop the data smartly? I have no idea…….Please give a bit inspiration!
", "rawMarkdown": "Amazing thinking!\nA lot of models in the notebooks neglect this part!!!!\nThey just simply crop the data according to the fmin fmax tmin tmax..\n\nBut how should we crop the data smartly? I have no idea.......Please give a bit inspiration!", "votes": null }, { "id": "1137410", "postDate": "01/03/2021 22:41:55", "content": "Hi @kneroma , thanks a lot for sharing! 👀
\nSimilar to @Elliot321, I didn't quite get it yet. Two questions.
\ncrop as in randomly cropping the audios, do you mean crop training audio to small pieces to train, but not directly using the specified time and frequency limit? Ah, I see, it seems a common thing for \"Cornell Birdcall Identification\".randomly cropping the audios. Does it mean when we cut test the 60s test audio into small pieces to classify, we just cut in a whatever convenient manner?@kneroma Do you think you could share some of your insight into a smarter way to crop the audio?
", "rawMarkdown": "kneroma Do you think you could share some of your insight into a smarter way to crop the audio?", "votes": null } ], "comments": [ { "id": 1132794, "author_name": "elliot321", "author_url": "", "post_date": "12/30/2020 17:34:44", "content": "Amazing thinking!
\nA lot of models in the notebooks neglect this part!!!!
\nThey just simply crop the data according to the fmin fmax tmin tmax..
But how should we crop the data smartly? I have no idea…….Please give a bit inspiration!
", "votes": null, "replies": [] }, { "id": 1137410, "author_name": "barnwellguy", "author_url": "", "post_date": "01/03/2021 22:41:55", "content": "Hi @kneroma , thanks a lot for sharing! 👀
\nSimilar to @Elliot321, I didn't quite get it yet. Two questions.
\ncrop as in randomly cropping the audios, do you mean crop training audio to small pieces to train, but not directly using the specified time and frequency limit? Ah, I see, it seems a common thing for \"Cornell Birdcall Identification\".randomly cropping the audios. Does it mean when we cut test the 60s test audio into small pieces to classify, we just cut in a whatever convenient manner?@kneroma Do you think you could share some of your insight into a smarter way to crop the audio?
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1132688": "The data for this competition is very special in the sense that the record are only partially labeled, and the labeled samples are done at a one-specy-basis, ie even if there is many species on the same **[t_min, t_max]**, you may end up with only one (or some) of them labeled.\n\nThis situation must require all of our attention. For instance, with the same **ResNeSt50** model , I got a LB score of **0.66** when randomly cropping the audios. And guess what, my score jumped to **0.861** when doing smarter cropping. I'm definitely sure that, even with a simple model, one can win this competition with just a good data cropping / normalization strategy.\n\nAnd you, does the cropping strategy impact your performance as much ?", "1132794": "Amazing thinking!\nA lot of models in the notebooks neglect this part!!!!\nThey just simply crop the data according to the fmin fmax tmin tmax..\n\nBut how should we crop the data smartly? I have no idea.......Please give a bit inspiration!", "1137410": "Hi @kneroma , thanks a lot for sharing! 👀\n\nSimilar to @Elliot321, I didn't quite get it yet. Two questions.\n- When you say `crop` as in `randomly cropping the audios`, do you mean crop training audio to small pieces to train, but not directly using the specified time and frequency limit? Ah, I see, it seems a common thing for \"Cornell Birdcall Identification\".\n- What does it mean by `randomly cropping the audios`. Does it mean when we cut test the 60s test audio into small pieces to classify, we just cut in a whatever convenient manner?", "1153379": "kneroma Do you think you could share some of your insight into a smarter way to crop the audio?" }, "source": "meta" }