{ "id": 467863, "title": "*UPDATE* [CV 0.715, LB 0.5] ResNet34-D Baseline", "url": "/competitions/hms-harmful-brain-activity-classification/discussion/467863", "author_name": "", "post_date": "2024-01-14T11:30:03.170757800Z", "votes": 39, "comment_count": 3, "views": 0, "content": "

I applied image classificaition approach for spectrogram files.

\n

I have shared very naive baseline (not using LR Scheduling, Data Augmentation).

\n

I tried several experiments, but I didn't train models well :(

\n

Training Notebook: https://www.kaggle.com/code/ttahara/hms-hbac-resnet34d-baseline-training
\nInference Notebook: https://www.kaggle.com/code/ttahara/hms-hbac-resnet34d-baseline-inference

\n

Update

\n

I added log transform and LR scheduling for comparing with Chris's EfficientNetB2 Starter. It dramatically improved CV and LB score. (CV 0.813, LB; 0.67 -> CV 0.715, LB 0.5)

", "messages": [ { "id": "2601350", "postDate": "01/14/2024 11:30:03", "content": "

I applied image classificaition approach for spectrogram files.

\n

I have shared very naive baseline (not using LR Scheduling, Data Augmentation).

\n

I tried several experiments, but I didn't train models well :(

\n

Training Notebook: https://www.kaggle.com/code/ttahara/hms-hbac-resnet34d-baseline-training
\nInference Notebook: https://www.kaggle.com/code/ttahara/hms-hbac-resnet34d-baseline-inference

\n

Update

\n

I added log transform and LR scheduling for comparing with Chris's EfficientNetB2 Starter. It dramatically improved CV and LB score. (CV 0.813, LB; 0.67 -> CV 0.715, LB 0.5)

", "rawMarkdown": "I applied image classificaition approach for spectrogram files.\n\nI have shared very naive baseline (not using LR Scheduling, Data Augmentation).\n\nI tried several experiments, but I didn't train models well :(\n\nTraining Notebook: https://www.kaggle.com/code/ttahara/hms-hbac-resnet34d-baseline-training\nInference Notebook: https://www.kaggle.com/code/ttahara/hms-hbac-resnet34d-baseline-inference\n\n**Update**\n\nI added **log transform** and **LR scheduling** for comparing with [Chris's EfficientNetB2 Starter](https://www.kaggle.com/code/cdeotte/efficientnetb2-starter-lb-0-57). It dramatically improved CV and LB score. (CV 0.813, LB; 0.67 -> CV 0.715, LB 0.5)", "votes": null }, { "id": "2601425", "postDate": "01/14/2024 12:20:06", "content": "

This is great. Thanks for sharing.

\n

So far GBT (i.e. ML my starter) and CNN (i.e. DL your starter) achieve the same CV 0.82 LB 0.67. It will be interesting to see which performs better in this competition, ML or DL? (Of course the winning solutions will probably be ensemble of both).

", "rawMarkdown": "This is great. Thanks for sharing.\n\nSo far GBT (i.e. ML my starter) and CNN (i.e. DL your starter) achieve the same CV 0.82 LB 0.67. It will be interesting to see which performs better in this competition, ML or DL? (Of course the winning solutions will probably be ensemble of both).", "votes": null }, { "id": "2601433", "postDate": "01/14/2024 12:25:18", "content": "

I'm intersted in that, too 🙂

\n

BTW, my baseline and your starter don't use eegs yet. Next I'll try to use them.

", "rawMarkdown": "I'm intersted in that, too 🙂\n\nBTW, my baseline and your starter don't use eegs yet. Next I'll try to use them.", "votes": null }, { "id": "2601526", "postDate": "01/14/2024 13:42:01", "content": "

So far i have not been able to improve my models using eeg parquets. But they should certainly help because the spectrograms are created from the eegs. See discussion here for tips on how to preprocess the eegs before inputting into DL models.

", "rawMarkdown": "So far i have not been able to improve my models using eeg parquets. But they should certainly help because the spectrograms are created from the eegs. See discussion [here][1] for tips on how to preprocess the eegs before inputting into DL models.\n\n[1]: https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/467877", "votes": null } ], "comments": [ { "id": 2601425, "author_name": "cdeotte", "author_url": "", "post_date": "01/14/2024 12:20:06", "content": "

This is great. Thanks for sharing.

\n

So far GBT (i.e. ML my starter) and CNN (i.e. DL your starter) achieve the same CV 0.82 LB 0.67. It will be interesting to see which performs better in this competition, ML or DL? (Of course the winning solutions will probably be ensemble of both).

", "votes": null, "replies": [ { "id": 2601433, "author_name": "ttahara", "author_url": "", "post_date": "01/14/2024 12:25:18", "content": "

I'm intersted in that, too 🙂

\n

BTW, my baseline and your starter don't use eegs yet. Next I'll try to use them.

", "votes": null, "replies": [ { "id": 2601526, "author_name": "cdeotte", "author_url": "", "post_date": "01/14/2024 13:42:01", "content": "

So far i have not been able to improve my models using eeg parquets. But they should certainly help because the spectrograms are created from the eegs. See discussion here for tips on how to preprocess the eegs before inputting into DL models.

", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2601350": "I applied image classificaition approach for spectrogram files.\n\nI have shared very naive baseline (not using LR Scheduling, Data Augmentation).\n\nI tried several experiments, but I didn't train models well :(\n\nTraining Notebook: https://www.kaggle.com/code/ttahara/hms-hbac-resnet34d-baseline-training\nInference Notebook: https://www.kaggle.com/code/ttahara/hms-hbac-resnet34d-baseline-inference\n\n**Update**\n\nI added **log transform** and **LR scheduling** for comparing with [Chris's EfficientNetB2 Starter](https://www.kaggle.com/code/cdeotte/efficientnetb2-starter-lb-0-57). It dramatically improved CV and LB score. (CV 0.813, LB; 0.67 -> CV 0.715, LB 0.5)", "2601425": "This is great. Thanks for sharing.\n\nSo far GBT (i.e. ML my starter) and CNN (i.e. DL your starter) achieve the same CV 0.82 LB 0.67. It will be interesting to see which performs better in this competition, ML or DL? (Of course the winning solutions will probably be ensemble of both).", "2601433": "I'm intersted in that, too 🙂\n\nBTW, my baseline and your starter don't use eegs yet. Next I'll try to use them.", "2601526": "So far i have not been able to improve my models using eeg parquets. But they should certainly help because the spectrograms are created from the eegs. See discussion [here][1] for tips on how to preprocess the eegs before inputting into DL models.\n\n[1]: https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/467877" }, "source": "meta" }