I've entered this competition framing the problem as a sequence problem, and willing to try a RNN approach. But then I saw several kernels tackling the problem with CNNs trained on MEL Spectogram arrays (which is all new to me). Which makes more sense to you and why?
\n\nHappy coding everyone
", "messages": [ { "id": "913054", "postDate": "07/02/2020 23:27:46", "content": "I've entered this competition framing the problem as a sequence problem, and willing to try a RNN approach. But then I saw several kernels tackling the problem with CNNs trained on MEL Spectogram arrays (which is all new to me). Which makes more sense to you and why?
\n\nHappy coding everyone
", "rawMarkdown": "I've entered this competition framing the problem as a sequence problem, and willing to try a RNN approach. But then I saw several kernels tackling the problem with CNNs trained on MEL Spectogram arrays (which is all new to me). Which makes more sense to you and why?\n\nHappy coding everyone", "votes": null }, { "id": "913116", "postDate": "07/03/2020 02:01:32", "content": "Honestly, just thinking about the problem, RNN makes more sense as sound is more of a sequence data, but looks like using CNNs on the spectrograms, like you said, tends to perform better overall.
", "rawMarkdown": "Honestly, just thinking about the problem, RNN makes more sense as sound is more of a sequence data, but looks like using CNNs on the spectrograms, like you said, tends to perform better overall.", "votes": null }, { "id": "913742", "postDate": "07/03/2020 11:37:05", "content": "Most of the gold medalists in previous years used CNN. And CNN also performs better than RNN in my model.
", "rawMarkdown": "Most of the gold medalists in previous years used CNN. And CNN also performs better than RNN in my model.", "votes": null }, { "id": "913832", "postDate": "07/03/2020 13:08:57", "content": "Combination of both ConvLSTM2D
", "rawMarkdown": "Combination of both ConvLSTM2D", "votes": null }, { "id": "914415", "postDate": "07/03/2020 21:05:20", "content": "bidirectional cuda lstm
", "rawMarkdown": "bidirectional cuda lstm", "votes": null }, { "id": "914444", "postDate": "07/03/2020 22:04:22", "content": "I've been approaching it as an \"identification\" problem but could not come up with anything interesting. Now looking at it from NLP lenses.
", "rawMarkdown": "I've been approaching it as an \"identification\" problem but could not come up with anything interesting. Now looking at it from NLP lenses.", "votes": null }, { "id": "982000", "postDate": "08/23/2020 00:04:13", "content": "maybe using data as images works better because there are a lot of pre-trained models for image, it helps a lot to perform better. But i will not be surprised if anyone could develop a transformer architecture that performs very well too
", "rawMarkdown": "maybe using data as images works better because there are a lot of pre-trained models for image, it helps a lot to perform better. But i will not be surprised if anyone could develop a transformer architecture that performs very well too", "votes": null } ], "comments": [ { "id": 982000, "author_name": "hidekin2011", "author_url": "", "post_date": "08/23/2020 00:04:13", "content": "maybe using data as images works better because there are a lot of pre-trained models for image, it helps a lot to perform better. But i will not be surprised if anyone could develop a transformer architecture that performs very well too
", "votes": null, "replies": [] }, { "id": 913116, "author_name": "jonykarki", "author_url": "", "post_date": "07/03/2020 02:01:32", "content": "Honestly, just thinking about the problem, RNN makes more sense as sound is more of a sequence data, but looks like using CNNs on the spectrograms, like you said, tends to perform better overall.
", "votes": null, "replies": [] }, { "id": 913742, "author_name": "xiaowangiiiii", "author_url": "", "post_date": "07/03/2020 11:37:05", "content": "Most of the gold medalists in previous years used CNN. And CNN also performs better than RNN in my model.
", "votes": null, "replies": [] }, { "id": 913832, "author_name": "mlneo07", "author_url": "", "post_date": "07/03/2020 13:08:57", "content": "Combination of both ConvLSTM2D
", "votes": null, "replies": [] }, { "id": 914415, "author_name": "servietsky", "author_url": "", "post_date": "07/03/2020 21:05:20", "content": "bidirectional cuda lstm
", "votes": null, "replies": [] }, { "id": 914444, "author_name": "ramarlina", "author_url": "", "post_date": "07/03/2020 22:04:22", "content": "I've been approaching it as an \"identification\" problem but could not come up with anything interesting. Now looking at it from NLP lenses.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "913054": "I've entered this competition framing the problem as a sequence problem, and willing to try a RNN approach. But then I saw several kernels tackling the problem with CNNs trained on MEL Spectogram arrays (which is all new to me). Which makes more sense to you and why?\n\nHappy coding everyone", "913116": "Honestly, just thinking about the problem, RNN makes more sense as sound is more of a sequence data, but looks like using CNNs on the spectrograms, like you said, tends to perform better overall.", "913742": "Most of the gold medalists in previous years used CNN. And CNN also performs better than RNN in my model.", "913832": "Combination of both ConvLSTM2D", "914415": "bidirectional cuda lstm", "914444": "I've been approaching it as an \"identification\" problem but could not come up with anything interesting. Now looking at it from NLP lenses.", "982000": "maybe using data as images works better because there are a lot of pre-trained models for image, it helps a lot to perform better. But i will not be surprised if anyone could develop a transformer architecture that performs very well too" }, "source": "meta" }