{ "id": 231950, "title": "EfficientNet + LSTM - LB: 8.89 (with credit to Y. Nakama, duh ;-)", "url": "/competitions/bms-molecular-translation/discussion/231950", "author_name": "", "post_date": "2021-04-11T11:14:29.135054400Z", "votes": 28, "comment_count": 5, "views": 0, "content": "

In the spirit of spring sharing, a combo for getting EfficientNet into your toolkit:

\n", "messages": [ { "id": "1270154", "postDate": "04/11/2021 11:14:29", "content": "

In the spirit of spring sharing, a combo for getting EfficientNet into your toolkit:

\n", "rawMarkdown": "In the spirit of spring sharing, a combo for getting EfficientNet into your toolkit:\n\n- training : https://www.kaggle.com/konradb/model-train-efficientnet\n\n- inference: https://www.kaggle.com/konradb/model-infer-efficientnet/\n\n- weights: https://www.kaggle.com/konradb/fitted-model", "votes": null }, { "id": "1272041", "postDate": "04/13/2021 07:10:03", "content": "

Good work!

", "rawMarkdown": "Good work!", "votes": null }, { "id": "1279070", "postDate": "04/20/2021 15:02:47", "content": "

Hi there.

\n

\n

[EDIT] - Just found out that you used 5 epochs from the inference notebook. Very cool!

\n
\n

Great job by the way!!

", "rawMarkdown": "Hi there.\n\n~~How many epochs did you run your training for? I couldn't deduce it from your training notebook? ~~\n\n**[EDIT] - Just found out that you used 5 epochs from the inference notebook. Very cool!**\n\n---\n\nGreat job by the way!!", "votes": null }, { "id": "1279312", "postDate": "04/20/2021 19:20:22", "content": "

Hi @konradb,

\n

Thank you so much for the share. Can I ask what LB score would this achieve after training for 15+ epochs? I'm trying to see if I should run your solution for 15+ epochs or if I should explore the open source solution Heng shared here: completed transformer starter kit … in pytorch.

\n

Thank you Konrad,

", "rawMarkdown": "Hi @konradb, \n\nThank you so much for the share. Can I ask what LB score would this achieve after training for 15+ epochs? I'm trying to see if I should run your solution for 15+ epochs or if I should explore the open source solution Heng shared here: [completed transformer starter kit ... in pytorch](https://www.kaggle.com/c/bms-molecular-translation/discussion/231190).\n\nThank you Konrad,", "votes": null }, { "id": "1279355", "postDate": "04/20/2021 19:52:29", "content": "

I pushed this combo (architecture + image size) to 12 epochs and LB score ~ 5.9. Didn't go further than that.

", "rawMarkdown": "I pushed this combo (architecture + image size) to 12 epochs and LB score ~ 5.9. Didn't go further than that.", "votes": null }, { "id": "1279369", "postDate": "04/20/2021 20:15:30", "content": "

You can get leaderboard 3.xx with B2 + LSTM with 15 epochs and 224 image size. Takes about 3 hours per epoch on my machine

", "rawMarkdown": "You can get leaderboard 3.xx with B2 + LSTM with 15 epochs and `224` image size. Takes about 3 hours per epoch on my machine", "votes": null } ], "comments": [ { "id": 1272041, "author_name": "zavodrobotov", "author_url": "", "post_date": "04/13/2021 07:10:03", "content": "

Good work!

", "votes": null, "replies": [] }, { "id": 1279070, "author_name": "dschettler8845", "author_url": "", "post_date": "04/20/2021 15:02:47", "content": "

Hi there.

\n

\n

[EDIT] - Just found out that you used 5 epochs from the inference notebook. Very cool!

\n
\n

Great job by the way!!

", "votes": null, "replies": [] }, { "id": 1279312, "author_name": "bopengiowa", "author_url": "", "post_date": "04/20/2021 19:20:22", "content": "

Hi @konradb,

\n

Thank you so much for the share. Can I ask what LB score would this achieve after training for 15+ epochs? I'm trying to see if I should run your solution for 15+ epochs or if I should explore the open source solution Heng shared here: completed transformer starter kit … in pytorch.

\n

Thank you Konrad,

", "votes": null, "replies": [ { "id": 1279355, "author_name": "konradb", "author_url": "", "post_date": "04/20/2021 19:52:29", "content": "

I pushed this combo (architecture + image size) to 12 epochs and LB score ~ 5.9. Didn't go further than that.

", "votes": null, "replies": [] }, { "id": 1279369, "author_name": "tuckerarrants", "author_url": "", "post_date": "04/20/2021 20:15:30", "content": "

You can get leaderboard 3.xx with B2 + LSTM with 15 epochs and 224 image size. Takes about 3 hours per epoch on my machine

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1270154": "In the spirit of spring sharing, a combo for getting EfficientNet into your toolkit:\n\n- training : https://www.kaggle.com/konradb/model-train-efficientnet\n\n- inference: https://www.kaggle.com/konradb/model-infer-efficientnet/\n\n- weights: https://www.kaggle.com/konradb/fitted-model", "1272041": "Good work!", "1279070": "Hi there.\n\n~~How many epochs did you run your training for? I couldn't deduce it from your training notebook? ~~\n\n**[EDIT] - Just found out that you used 5 epochs from the inference notebook. Very cool!**\n\n---\n\nGreat job by the way!!", "1279312": "Hi @konradb, \n\nThank you so much for the share. Can I ask what LB score would this achieve after training for 15+ epochs? I'm trying to see if I should run your solution for 15+ epochs or if I should explore the open source solution Heng shared here: [completed transformer starter kit ... in pytorch](https://www.kaggle.com/c/bms-molecular-translation/discussion/231190).\n\nThank you Konrad,", "1279355": "I pushed this combo (architecture + image size) to 12 epochs and LB score ~ 5.9. Didn't go further than that.", "1279369": "You can get leaderboard 3.xx with B2 + LSTM with 15 epochs and `224` image size. Takes about 3 hours per epoch on my machine" }, "source": "meta" }