![\"single_clip\"](\"https://i.ibb.co/6gQy9d3/Inferenec-Weighting-drawio-1.png\")
Hi there, I've noticed this phenomenon in the past BirdCLEF competitions and it will probably occur in this competition as well.
\nThe main example of this technique are kneroma's notebooks for training and inference.
\nThe thing I've noticed is:
\nWould this not stretch the spectrograms? I've noticed in some of the published papers, the reason given is that training on 7-second segments 'works better' and is 'more likely to contain bird calls'. I like the idea of replicating previous competition solutions, but I like to understand why they did what they did. In this case, I'm most of the way there…. but I don't understand this inference part? Can anyone (or @kneroma ?) clarify the logic here and why this doesn't result in worse performance?
\nJust as an aside… I think I may handle it by using some sort of overlapping segment system.
\nI took the time to create a diagram illustrating what this inference would look like for a single 5-second test clip given a step size of 2 instead of 1 (mostly because it's easier to draw). Note the weighting is wrong for the 3/5 overlap… duh… obviously this should have a weight of 0.6 not 0.5…
\n
If anyone can offer clarity on my above question or comment on this potential inference approach, it would be greatly appreciated. Thanks in advance!
", "messages": [ { "id": "1723505", "postDate": "03/15/2022 13:29:22", "content": "Hi there, I've noticed this phenomenon in the past BirdCLEF competitions and it will probably occur in this competition as well.
\nThe main example of this technique are kneroma's notebooks for training and inference.
\nThe thing I've noticed is:
\nWould this not stretch the spectrograms? I've noticed in some of the published papers, the reason given is that training on 7-second segments 'works better' and is 'more likely to contain bird calls'. I like the idea of replicating previous competition solutions, but I like to understand why they did what they did. In this case, I'm most of the way there…. but I don't understand this inference part? Can anyone (or @kneroma ?) clarify the logic here and why this doesn't result in worse performance?
\nJust as an aside… I think I may handle it by using some sort of overlapping segment system.
\nI took the time to create a diagram illustrating what this inference would look like for a single 5-second test clip given a step size of 2 instead of 1 (mostly because it's easier to draw). Note the weighting is wrong for the 3/5 overlap… duh… obviously this should have a weight of 0.6 not 0.5…
\n
If anyone can offer clarity on my above question or comment on this potential inference approach, it would be greatly appreciated. Thanks in advance!
", "rawMarkdown": "Hi there, I've noticed this phenomenon in the past BirdCLEF competitions and it will probably occur in this competition as well.\n\nThe main example of this technique are [**kneroma's**](https://www.kaggle.com/kneroma) notebooks for [**training**](https://www.kaggle.com/kneroma/clean-fast-simple-bird-identifier-training-colab) and [**inference**](https://www.kaggle.com/kneroma/clean-fast-simple-bird-identifier-inference).\n\nThe thing I've noticed is:\n* Train is done on 7-second clips (128,281,3) MEL spectrograms. \n* However, we perform inference on 5-second clips (128,201,3) MEL spectrograms.\n\nWould this not stretch the spectrograms? I've noticed in some of the published papers, the reason given is that training on 7-second segments 'works better' and is 'more likely to contain bird calls'. I like the idea of replicating previous competition solutions, but I like to understand why they did what they did. In this case, I'm most of the way there.... but I don't understand this inference part? Can anyone (or @kneroma ?) clarify the logic here and why this doesn't result in worse performance?\n\n---\n\nJust as an aside... I think I may handle it by using some sort of overlapping segment system.\n* i.e. Pad beginning and end of long audio clip w/ 6 seconds of blank audio\n* Split the test audio clip into 7-second clips with a 1-second step\n* Use a weighted average (based on the percentage of overlap between the 7-second and 5-second segment) of all predictions.\n * i.e. when a 7-second segment completely overlaps a 5-second segment, it will have a weight of 1 and those predictions will contribute strongly. When a 7-second segment only overlaps with HALF a 5-second segment, its weight will be 0.5. \n\nI wish I could give you an answer right now but not only i'm tired but also I have to read again these old works to be sure I'm not telling you nonsense ! Good competition and hope I can join before the end since audio competitions are always interesting ones.
", "rawMarkdown": "I wish I could give you an answer right now but not only i'm tired but also I have to read again these old works to be sure I'm not telling you nonsense ! Good competition and hope I can join before the end since audio competitions are always interesting ones.", "votes": null }, { "id": "1723571", "postDate": "03/15/2022 14:29:05", "content": "Looking forward to it! :)
", "rawMarkdown": "Looking forward to it! :)", "votes": null } ], "comments": [ { "id": 1723550, "author_name": "kneroma", "author_url": "", "post_date": "03/15/2022 14:09:24", "content": "I wish I could give you an answer right now but not only i'm tired but also I have to read again these old works to be sure I'm not telling you nonsense ! Good competition and hope I can join before the end since audio competitions are always interesting ones.
", "votes": null, "replies": [ { "id": 1723571, "author_name": "dschettler8845", "author_url": "", "post_date": "03/15/2022 14:29:05", "content": "Looking forward to it! :)
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1723505": "Hi there, I've noticed this phenomenon in the past BirdCLEF competitions and it will probably occur in this competition as well.\n\nThe main example of this technique are [**kneroma's**](https://www.kaggle.com/kneroma) notebooks for [**training**](https://www.kaggle.com/kneroma/clean-fast-simple-bird-identifier-training-colab) and [**inference**](https://www.kaggle.com/kneroma/clean-fast-simple-bird-identifier-inference).\n\nThe thing I've noticed is:\n* Train is done on 7-second clips (128,281,3) MEL spectrograms. \n* However, we perform inference on 5-second clips (128,201,3) MEL spectrograms.\n\nWould this not stretch the spectrograms? I've noticed in some of the published papers, the reason given is that training on 7-second segments 'works better' and is 'more likely to contain bird calls'. I like the idea of replicating previous competition solutions, but I like to understand why they did what they did. In this case, I'm most of the way there.... but I don't understand this inference part? Can anyone (or @kneroma ?) clarify the logic here and why this doesn't result in worse performance?\n\n---\n\nJust as an aside... I think I may handle it by using some sort of overlapping segment system.\n* i.e. Pad beginning and end of long audio clip w/ 6 seconds of blank audio\n* Split the test audio clip into 7-second clips with a 1-second step\n* Use a weighted average (based on the percentage of overlap between the 7-second and 5-second segment) of all predictions.\n * i.e. when a 7-second segment completely overlaps a 5-second segment, it will have a weight of 1 and those predictions will contribute strongly. When a 7-second segment only overlaps with HALF a 5-second segment, its weight will be 0.5. \n\n