Since spectrograms are sequential data, it would make sense to apply a RNN along one of the two dimension of the spectrogram, then average over all positions and apply a single dense layer to output a binary score.
\nHowever, the current top scoring notebook instead uses a pre-trained CNN with the 1st, 3rd, and 5th positions as the three \"channels\", and it is working extremely well.
\nI can't wait to see which approach will take the lead here!
", "messages": [ { "id": 1302792, "postDate": "2021-05-11T18:00:26.967Z", "content": "Since spectrograms are sequential data, it would make sense to apply a RNN along one of the two dimension of the spectrogram, then average over all positions and apply a single dense layer to output a binary score.
\nHowever, the current top scoring notebook instead uses a pre-trained CNN with the 1st, 3rd, and 5th positions as the three \"channels\", and it is working extremely well.
\nI can't wait to see which approach will take the lead here!
", "rawMarkdown": "Since spectrograms are sequential data, it would make sense to apply a RNN along one of the two dimension of the spectrogram, then average over all positions and apply a single dense layer to output a binary score.\n\nHowever, the current [top scoring notebook](https://www.kaggle.com/awsaf49/seti-bl-tf-starter-tpu) instead uses a pre-trained CNN with the 1st, 3rd, and 5th positions as the three \"channels\", and it is working extremely well.\n\nI can't wait to see which approach will take the lead here!", "votes": 14 }, { "id": 1302896, "postDate": "2021-05-11T19:12:14.897Z", "content": "Exactly.
\nI am trying to experiment with CNN + RNN pipeline for now.
Interesting. Do you mean apply RNN first or CNN first? Would be nice to have a diagram/pseudocode :)
", "rawMarkdown": "Interesting. Do you mean apply RNN first or CNN first? Would be nice to have a diagram/pseudocode :)" }, { "id": 1303726, "postDate": "2021-05-12T08:01:15.840Z", "content": "Kind of similar to Sliding window problem similar to deep learning pipelines on CT-Scans images.
\nBut it didn't work for me.
Interesting. Thanks for sharing!
", "rawMarkdown": "Interesting. Thanks for sharing!" } ] }, { "id": 1302802, "postDate": "2021-05-11T18:08:31.403Z", "content": "@xhlulu I think it's a very common practice to use CNN even in sequential data. I've seen people using conv1d in signals like ECG or EEG. My intuition was from that. Can't wait to see work on RNN or Transformers
", "rawMarkdown": "@xhlulu I think it's a very common practice to use **CNN** even in sequential data. I've seen people using **conv1d** in signals like **ECG** or **EEG**. My intuition was from that. Can't wait to see work on **RNN** or **Transformers**", "votes": 2, "replies": [ { "id": 1302840, "postDate": "2021-05-11T18:39:47.150Z", "content": "Thanks, that's pretty interesting. Conv1D makes sense for sequential data - I think it might also be a good option here (along with RNN and Conv2D).
", "rawMarkdown": "Thanks, that's pretty interesting. Conv1D makes sense for sequential data - I think it might also be a good option here (along with RNN and Conv2D).", "votes": 2 } ] } ], "comments": [ { "id": 1302896, "author_name": "Salman", "author_url": "", "post_date": "2021-05-11T19:12:14.897000", "content": "Exactly.
\nI am trying to experiment with CNN + RNN pipeline for now.
Interesting. Do you mean apply RNN first or CNN first? Would be nice to have a diagram/pseudocode :)
", "votes": 0, "replies": [] }, { "id": 1303726, "author_name": "Salman", "author_url": "", "post_date": "2021-05-12T08:01:15.840000", "content": "Kind of similar to Sliding window problem similar to deep learning pipelines on CT-Scans images.
\nBut it didn't work for me.
Interesting. Thanks for sharing!
", "votes": 0, "replies": [] } ] }, { "id": 1302802, "author_name": "Awsaf", "author_url": "", "post_date": "2021-05-11T18:08:31.403000", "content": "@xhlulu I think it's a very common practice to use CNN even in sequential data. I've seen people using conv1d in signals like ECG or EEG. My intuition was from that. Can't wait to see work on RNN or Transformers
", "votes": 2, "replies": [ { "id": 1302840, "author_name": "xhlulu", "author_url": "", "post_date": "2021-05-11T18:39:47.150000", "content": "Thanks, that's pretty interesting. Conv1D makes sense for sequential data - I think it might also be a good option here (along with RNN and Conv2D).
", "votes": 2, "replies": [] } ] } ], "raw_markdown_by_id": { "1302792": "Since spectrograms are sequential data, it would make sense to apply a RNN along one of the two dimension of the spectrogram, then average over all positions and apply a single dense layer to output a binary score.\n\nHowever, the current [top scoring notebook](https://www.kaggle.com/awsaf49/seti-bl-tf-starter-tpu) instead uses a pre-trained CNN with the 1st, 3rd, and 5th positions as the three \"channels\", and it is working extremely well.\n\nI can't wait to see which approach will take the lead here!", "1302896": "Exactly. \nI am trying to experiment with CNN + RNN pipeline for now. ", "1302802": "@xhlulu I think it's a very common practice to use **CNN** even in sequential data. I've seen people using **conv1d** in signals like **ECG** or **EEG**. My intuition was from that. Can't wait to see work on **RNN** or **Transformers**" } }