Hello everybody!
\nWe are smoothly moving to the end of this wonderful competition, as a preliminary summary I will give ideas that did not work for me:
Right now my best model is a two stage with 0.32 lb, second stage CV of 0.317.
\nWhat interests me the most now is the distribution between public and private test samples.
\nConsidering that the test sample in general was small, there may be a situation that the private test sample has more samples with a small number of votes, which will automatically mean a shake-up in terms of the density of our results.
\nWhat do you think about it and what ideas of yours didn't work?
\nGood luck to everyone and a peaceful sky above your head!
Hello everybody!
\nWe are smoothly moving to the end of this wonderful competition, as a preliminary summary I will give ideas that did not work for me:
Right now my best model is a two stage with 0.32 lb, second stage CV of 0.317.
\nWhat interests me the most now is the distribution between public and private test samples.
\nConsidering that the test sample in general was small, there may be a situation that the private test sample has more samples with a small number of votes, which will automatically mean a shake-up in terms of the density of our results.
\nWhat do you think about it and what ideas of yours didn't work?
\nGood luck to everyone and a peaceful sky above your head!
what's the cv and lb of the 1st stage ? If you want to reveal.
", "rawMarkdown": "what's the cv and lb of the 1st stage ? If you want to reveal.", "votes": null }, { "id": "2698140", "postDate": "03/15/2024 10:17:31", "content": "At the first stage 0.679-0.680, lb I did not check
", "rawMarkdown": "At the first stage 0.679-0.680, lb I did not check", "votes": null }, { "id": "2698257", "postDate": "03/15/2024 11:28:57", "content": "As you know, high quality model cv is very relevant with LB.
\nSo In my opinion, Competiton host would probably make private dataset(about 1700~1800) with high quality.
\nBut i'm not sure haha
Increase image size actually hurt your cv
", "rawMarkdown": "Increase image size actually hurt your cv", "votes": null }, { "id": "2698582", "postDate": "03/15/2024 15:26:03", "content": "I wrote about it, if there are only samples with more than 10 evaluators, then everything will be fine without strong fluctuations, but if, on the contrary, there are many samples from a sample with a smaller number of evaluators, it will hurt.
", "rawMarkdown": "I wrote about it, if there are only samples with more than 10 evaluators, then everything will be fine without strong fluctuations, but if, on the contrary, there are many samples from a sample with a smaller number of evaluators, it will hurt.", "votes": null }, { "id": "2698595", "postDate": "03/15/2024 15:28:28", "content": "In this competition, the correct preprocessing of the data is important, and the size of the processed images is not so important
", "rawMarkdown": "In this competition, the correct preprocessing of the data is important, and the size of the processed images is not so important", "votes": null }, { "id": "2700170", "postDate": "03/16/2024 11:37:07", "content": "Could you please discribe in high level about 2 stages?
", "rawMarkdown": "Could you please discribe in high level about 2 stages?", "votes": null }, { "id": "2700237", "postDate": "03/16/2024 12:21:24", "content": "Add more spec data from eeg with different kind of converting methods boost my cv and lb😃
", "rawMarkdown": "Add more spec data from eeg with different kind of converting methods boost my cv and lb😃", "votes": null }, { "id": "2700342", "postDate": "03/16/2024 13:29:31", "content": "Yes, it improved my CV, but not lb
", "rawMarkdown": "Yes, it improved my CV, but not lb", "votes": null }, { "id": "2700351", "postDate": "03/16/2024 13:33:12", "content": "Here is the discussion proposed by the author of this idea -
\nhttps://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/477135
Yes, I wrote about it, any changes with the model do not give a tangible result. It is a matter of preprocessing, choice of probabilities and choice of model architectures. I couldn't get WaveNet to run on the TPU, so I used EEGNet.
", "rawMarkdown": "Yes, I wrote about it, any changes with the model do not give a tangible result. It is a matter of preprocessing, choice of probabilities and choice of model architectures. I couldn't get WaveNet to run on the TPU, so I used EEGNet.", "votes": null }, { "id": "2718169", "postDate": "03/27/2024 02:33:59", "content": "Changing model does not improve cv and lb.
\nWith baseline code, here is the result of LB.
\nEfficient Net is fast to train and better performance.
Yes, but you can make an ensemble of these models, it will slightly improve the better model. By the way, in my pipeline, ConvNext does not work at all (shows terrible results)
", "rawMarkdown": "Yes, but you can make an ensemble of these models, it will slightly improve the better model. By the way, in my pipeline, ConvNext does not work at all (shows terrible results)", "votes": null } ], "comments": [ { "id": 2698134, "author_name": "nikhilmishradev", "author_url": "", "post_date": "03/15/2024 10:14:19", "content": "what's the cv and lb of the 1st stage ? If you want to reveal.
", "votes": null, "replies": [ { "id": 2698140, "author_name": "aikhmelnytskyy", "author_url": "", "post_date": "03/15/2024 10:17:31", "content": "At the first stage 0.679-0.680, lb I did not check
", "votes": null, "replies": [] } ] }, { "id": 2698257, "author_name": "seoyunje", "author_url": "", "post_date": "03/15/2024 11:28:57", "content": "As you know, high quality model cv is very relevant with LB.
\nSo In my opinion, Competiton host would probably make private dataset(about 1700~1800) with high quality.
\nBut i'm not sure haha
I wrote about it, if there are only samples with more than 10 evaluators, then everything will be fine without strong fluctuations, but if, on the contrary, there are many samples from a sample with a smaller number of evaluators, it will hurt.
", "votes": null, "replies": [] } ] }, { "id": 2698328, "author_name": "goldenlock", "author_url": "", "post_date": "03/15/2024 12:31:31", "content": "Increase image size actually hurt your cv
", "votes": null, "replies": [ { "id": 2698595, "author_name": "aikhmelnytskyy", "author_url": "", "post_date": "03/15/2024 15:28:28", "content": "In this competition, the correct preprocessing of the data is important, and the size of the processed images is not so important
", "votes": null, "replies": [] } ] }, { "id": 2700170, "author_name": "darkhann", "author_url": "", "post_date": "03/16/2024 11:37:07", "content": "Could you please discribe in high level about 2 stages?
", "votes": null, "replies": [ { "id": 2700351, "author_name": "aikhmelnytskyy", "author_url": "", "post_date": "03/16/2024 13:33:12", "content": "Here is the discussion proposed by the author of this idea -
\nhttps://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/477135
Add more spec data from eeg with different kind of converting methods boost my cv and lb😃
", "votes": null, "replies": [ { "id": 2700342, "author_name": "aikhmelnytskyy", "author_url": "", "post_date": "03/16/2024 13:29:31", "content": "Yes, it improved my CV, but not lb
", "votes": null, "replies": [] } ] }, { "id": 2701047, "author_name": "nartaa", "author_url": "", "post_date": "03/16/2024 20:28:59", "content": "Yes, I wrote about it, any changes with the model do not give a tangible result. It is a matter of preprocessing, choice of probabilities and choice of model architectures. I couldn't get WaveNet to run on the TPU, so I used EEGNet.
", "votes": null, "replies": [] } ] }, { "id": 2718169, "author_name": "clearwaterkzk", "author_url": "", "post_date": "03/27/2024 02:33:59", "content": "Changing model does not improve cv and lb.
\nWith baseline code, here is the result of LB.
\nEfficient Net is fast to train and better performance.
Yes, but you can make an ensemble of these models, it will slightly improve the better model. By the way, in my pipeline, ConvNext does not work at all (shows terrible results)
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "2698044": "Hello everybody!\nWe are smoothly moving to the end of this wonderful competition, as a preliminary summary I will give ideas that did not work for me:\n- MILAttentionLayer - I tried to use MILAttention, but adding this layer just makes lb worse, without much change to CV.\n- Increasing the size of images - this does not affect lb in any way.\n- Using ConvNeXt and Resnet - the models work much worse than the basic one, obviously the problem is somewhere in the architecture itself and my pipeline.\n- Augmentation for eeg - after trying various options the most I could achieve was this improvement in my CV without affecting LB.\n\nRight now my best model is a two stage with 0.32 lb, second stage CV of 0.317.\n\nWhat interests me the most now is the distribution between public and private test samples. \nConsidering that the test sample in general was small, there may be a situation that the private test sample has more samples with a small number of votes, which will automatically mean a shake-up in terms of the density of our results.\nWhat do you think about it and what ideas of yours didn't work?\nGood luck to everyone and a peaceful sky above your head!", "2698134": "what's the cv and lb of the 1st stage ? If you want to reveal.", "2698140": "At the first stage 0.679-0.680, lb I did not check", "2698257": "As you know, high quality model cv is very relevant with LB.\nSo In my opinion, Competiton host would probably make private dataset(about 1700~1800) with high quality.\nBut i'm not sure haha", "2698328": "Increase image size actually hurt your cv", "2698582": "I wrote about it, if there are only samples with more than 10 evaluators, then everything will be fine without strong fluctuations, but if, on the contrary, there are many samples from a sample with a smaller number of evaluators, it will hurt.", "2698595": "In this competition, the correct preprocessing of the data is important, and the size of the processed images is not so important", "2700170": "Could you please discribe in high level about 2 stages?", "2700237": "Add more spec data from eeg with different kind of converting methods boost my cv and lb😃", "2700342": "Yes, it improved my CV, but not lb", "2700351": "Here is the discussion proposed by the author of this idea - \nhttps://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/477135", "2701047": "Image resize and transformations didn't help, but It's inconclusive.\n- Added transformer block to WaveNet didn't help yet.\n- Adding more FC layers to EffecientNet's model head didn't help.\n- Following the official guide at Keras's site on how to train and fine-tune backbone models such as EffecientNet (freezing and unfreezing layers) didn't help.", "2701068": "Yes, I wrote about it, any changes with the model do not give a tangible result. It is a matter of preprocessing, choice of probabilities and choice of model architectures. I couldn't get WaveNet to run on the TPU, so I used EEGNet.", "2718169": "Changing model does not improve cv and lb.\nWith baseline code, here is the result of LB.\nEfficient Net is fast to train and better performance.\n- Efficient Net : 0.40\n- Vit : 0.44\n- ConvNext : 0.44", "2718667": "Yes, but you can make an ensemble of these models, it will slightly improve the better model. By the way, in my pipeline, ConvNext does not work at all (shows terrible results)" }, "source": "meta" }