\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1557202%2F6e848153a09c907717f27e2b948da5d6%2FScreenshot%20from%202021-01-21%2015-08-10.png?generation=1611263328781944&alt=media\")
I have clipped training data according to the t_min and t_max and trained a model using those clips, score is good, but in test data we don't have any time given to split the audio. What can be the best way from here to make predictions on test data ?
", "messages": [ { "id": "1161375", "postDate": "01/20/2021 14:25:09", "content": "I have clipped training data according to the t_min and t_max and trained a model using those clips, score is good, but in test data we don't have any time given to split the audio. What can be the best way from here to make predictions on test data ?
", "rawMarkdown": "I have clipped training data according to the t_min and t_max and trained a model using those clips, score is good, but in test data we don't have any time given to split the audio. What can be the best way from here to make predictions on test data ?", "votes": null }, { "id": "1161424", "postDate": "01/20/2021 15:08:45", "content": "Obviously you do not have t_min and t_max available in test.
", "rawMarkdown": "Obviously you do not have t_min and t_max available in test.", "votes": null }, { "id": "1161443", "postDate": "01/20/2021 15:14:50", "content": "can you please tell me something about how can i proceed further.
", "rawMarkdown": "can you please tell me something about how can i proceed further.", "votes": null }, { "id": "1161644", "postDate": "01/20/2021 17:02:51", "content": "use a different approach
", "rawMarkdown": "use a different approach", "votes": null }, { "id": "1161680", "postDate": "01/20/2021 17:20:58", "content": "Break into chunks of 1 sec videos? Then merge the results. You would have to do that for training also. Your data processing approach will change but your core logic to identify the bird sounds may not change much
\nI haven't had a deep look at the kernels yet, but I am sure there must be a few which are tackling this issue already
", "rawMarkdown": "Break into chunks of 1 sec videos? Then merge the results. You would have to do that for training also. Your data processing approach will change but your core logic to identify the bird sounds may not change much\n\nI haven't had a deep look at the kernels yet, but I am sure there must be a few which are tackling this issue already", "votes": null }, { "id": "1163688", "postDate": "01/21/2021 21:09:12", "content": "A starting point would be, as many notebooks do, make a sliding window of clips from the 60s audios and feed to the model. Then, you have a series of predictions with which you can take a max or something more sohpisticated.
\n
@barnwellguy good to see you in top 3 :)
", "rawMarkdown": "barnwellguy good to see you in top 3 :)", "votes": null }, { "id": "1165570", "postDate": "01/23/2021 04:33:14", "content": "sorry. I think 1 sec would be too small. 10 sec may be more appropriate. There are some good discussions on this in this forum. You can browse them
", "rawMarkdown": "sorry. I think 1 sec would be too small. 10 sec may be more appropriate. There are some good discussions on this in this forum. You can browse them", "votes": null } ], "comments": [ { "id": 1161424, "author_name": "philippsinger", "author_url": "", "post_date": "01/20/2021 15:08:45", "content": "Obviously you do not have t_min and t_max available in test.
", "votes": null, "replies": [ { "id": 1161443, "author_name": "tusharaggarwal45", "author_url": "", "post_date": "01/20/2021 15:14:50", "content": "can you please tell me something about how can i proceed further.
", "votes": null, "replies": [] }, { "id": 1161644, "author_name": "philippsinger", "author_url": "", "post_date": "01/20/2021 17:02:51", "content": "use a different approach
", "votes": null, "replies": [] } ] }, { "id": 1161680, "author_name": "allohvk", "author_url": "", "post_date": "01/20/2021 17:20:58", "content": "Break into chunks of 1 sec videos? Then merge the results. You would have to do that for training also. Your data processing approach will change but your core logic to identify the bird sounds may not change much
\nI haven't had a deep look at the kernels yet, but I am sure there must be a few which are tackling this issue already
", "votes": null, "replies": [ { "id": 1165570, "author_name": "allohvk", "author_url": "", "post_date": "01/23/2021 04:33:14", "content": "sorry. I think 1 sec would be too small. 10 sec may be more appropriate. There are some good discussions on this in this forum. You can browse them
", "votes": null, "replies": [] } ] }, { "id": 1163688, "author_name": "barnwellguy", "author_url": "", "post_date": "01/21/2021 21:09:12", "content": "A starting point would be, as many notebooks do, make a sliding window of clips from the 60s audios and feed to the model. Then, you have a series of predictions with which you can take a max or something more sohpisticated.
\n
@barnwellguy good to see you in top 3 :)
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1161375": "I have clipped training data according to the t_min and t_max and trained a model using those clips, score is good, but in test data we don't have any time given to split the audio. What can be the best way from here to make predictions on test data ?", "1161424": "Obviously you do not have t_min and t_max available in test.", "1161443": "can you please tell me something about how can i proceed further.", "1161644": "use a different approach", "1161680": "Break into chunks of 1 sec videos? Then merge the results. You would have to do that for training also. Your data processing approach will change but your core logic to identify the bird sounds may not change much\n\nI haven't had a deep look at the kernels yet, but I am sure there must be a few which are tackling this issue already", "1163688": "A starting point would be, as many notebooks do, make a sliding window of clips from the 60s audios and feed to the model. Then, you have a series of predictions with which you can take a max or something more sohpisticated.\n", "1164957": "barnwellguy good to see you in top 3 :)", "1165570": "sorry. I think 1 sec would be too small. 10 sec may be more appropriate. There are some good discussions on this in this forum. You can browse them" }, "source": "meta" }