{ "id": 136030, "title": "๐Ÿ…Summary of Top Solutions๐Ÿ…", "url": "/competitions/bengaliai-cv19/discussion/136030", "author_name": "Qishen Ha", "post_date": "2020-03-17T05:30:26.109000", "votes": 84, "comment_count": 9, "views": 0, "content": "

To be updated...

\n\n

Main Ideas

\n\n

Basically we have 5 main ideas that works well in this competition as we can see now.

\n\n
    \n
  1. You don't give me unseen graphemes? I'll make it by myself. (1st, 2nd, 7th)
    2. \n
  2. Figuring out unseen graphemes then predict them better. (3rd, 4th, 8th)
    3. \n
  3. Relabel consonant diacritic (5th)
    4. \n
  4. Using exists external data (9th)
    5. \n
  5. Postprocessing (5th, 6th, 13th, 14th)
    6. \n
\n\n

Top Solutions

\n\n", "messages": [ { "id": 776074, "postDate": "2020-03-17T05:30:26.110Z", "content": "

To be updated...

\n\n

Main Ideas

\n\n

Basically we have 5 main ideas that works well in this competition as we can see now.

\n\n
    \n
  1. You don't give me unseen graphemes? I'll make it by myself. (1st, 2nd, 7th)
    2. \n
  2. Figuring out unseen graphemes then predict them better. (3rd, 4th, 8th)
    3. \n
  3. Relabel consonant diacritic (5th)
    4. \n
  4. Using exists external data (9th)
    5. \n
  5. Postprocessing (5th, 6th, 13th, 14th)
    6. \n
\n\n

Top Solutions

\n\n", "rawMarkdown": "To be updated...\n\n\n# Main Ideas\n\nBasically we have 5 main ideas that works well in this competition as we can see now.\n\n\n1. You don't give me unseen graphemes? I'll make it by myself. (1st, 2nd, 7th)\n2. Figuring out unseen graphemes then predict them better. (3rd, 4th, 8th)\n3. Relabel consonant diacritic (5th)\n4. Using exists external data (9th)\n5. Postprocessing (5th, 6th, 13th, 14th)\n\n# Top Solutions\n\n* [1st Place Solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/135984) \n * CycleGAN\n* [2nd Place Solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/135966)\n * Generated grapheme and Improve C=3/6\n* [3rd Place Solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/135982)\n * Arcface detect unseen\n* [4th Place Solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/136982)\n * Arcface detect unseen\n* [5th Place Solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/136129)\n * Relabelled consonant diacritic (Improved C=3/6) & Postprocessing\n* [6th Place Solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/136011)\n * Postprocessing\n* [7th Place Solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/135960)\n * Make unseen grapheme from font.\n* [8th Place Solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/135990)\n * Arcface detect unseen\n* [9th Place Solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/135985)\n * External data /w pesudo label\n* [10th Place Solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/136815)\n * Best single model I've ever see. But why it works??? \n* [12th Place Solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/135998)\n * Many tricks\n* [13th Place Solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/136116)\n * Distribution postprocessing\n* [14th Place Solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/136021)\n * Distribution postprocessing -- REAL MAGIC\n\n\n", "votes": 84 }, { "id": 776703, "postDate": "2020-03-17T15:18:41.827Z", "content": "

The diversity of top solutions in this competition is so interesting๐Ÿ˜ƒ Guess the hosts will be happy with a decent model after combining them

", "rawMarkdown": "The diversity of top solutions in this competition is so interesting๐Ÿ˜ƒ Guess the hosts will be happy with a decent model after combining them", "votes": 5 }, { "id": 776599, "postDate": "2020-03-17T13:54:29.660Z", "content": "

I also made the summary before yours but this one is more elegant; I'll remove mine. ๐Ÿ˜€ \nand Congratulation again. @haqishen ๐Ÿ˜ƒ

", "rawMarkdown": "I also made the summary before yours but this one is more elegant; I'll remove mine. ๐Ÿ˜€ \nand Congratulation again. @haqishen ๐Ÿ˜ƒ ", "votes": 3, "replies": [ { "id": 776608, "postDate": "2020-03-17T14:02:06.630Z", "content": "

I made this idea indexing in case for future for myself, sorry for overlaping yours ๐Ÿ˜‚

", "rawMarkdown": "I made this idea indexing in case for future for myself, sorry for overlaping yours ๐Ÿ˜‚ ", "votes": 3 } ] }, { "id": 778187, "postDate": "2020-03-18T08:02:18.937Z", "content": "

So nice summary! thanks for sharing.

", "rawMarkdown": "So nice summary! thanks for sharing.", "votes": 1, "replies": [ { "id": 778460, "postDate": "2020-03-18T13:13:00.267Z", "content": "

Including yours. But I really don't know why your model works ๐Ÿ˜‚

", "rawMarkdown": "Including yours. But I really don't know why your model works ๐Ÿ˜‚ ", "votes": 1 }, { "id": 778466, "postDate": "2020-03-18T13:19:11.800Z", "content": "

Haha! Me too.

\n\n

My model is very simple so it did not overfit to Public LB? I could not imagine such a incredible shakeup ๐Ÿ˜‚

", "rawMarkdown": "Haha! Me too.\n\nMy model is very simple so it did not overfit to Public LB? I could not imagine such a incredible shakeup ๐Ÿ˜‚ ", "votes": 1 } ] }, { "id": 778281, "postDate": "2020-03-18T10:00:14.963Z", "content": "

Looks like this thread really needs updating ... not doing justice to the amazingly creative solutions like 5th place multilabel classification for consonants.

", "rawMarkdown": "Looks like this thread really needs updating ... not doing justice to the amazingly creative solutions like 5th place multilabel classification for consonants.", "votes": 2, "replies": [ { "id": 778465, "postDate": "2020-03-18T13:18:43.630Z", "content": "

Thanks for your kindly advice as I actually cannot follow everything by myself.\nIf you have any other advice please let me know ;)

", "rawMarkdown": "Thanks for your kindly advice as I actually cannot follow everything by myself.\nIf you have any other advice please let me know ;)" } ] }, { "id": 777720, "postDate": "2020-03-17T21:54:43.937Z", "rawMarkdown": "", "votes": 1, "isDeleted": true } ], "comments": [ { "id": 776703, "author_name": "Nicholas Lyu", "author_url": "", "post_date": "2020-03-17T15:18:41.827000", "content": "

The diversity of top solutions in this competition is so interesting๐Ÿ˜ƒ Guess the hosts will be happy with a decent model after combining them

", "votes": 5, "replies": [] }, { "id": 776599, "author_name": "Innat", "author_url": "", "post_date": "2020-03-17T13:54:29.660000", "content": "

I also made the summary before yours but this one is more elegant; I'll remove mine. ๐Ÿ˜€ \nand Congratulation again. @haqishen ๐Ÿ˜ƒ

", "votes": 3, "replies": [ { "id": 776608, "author_name": "Qishen Ha", "author_url": "", "post_date": "2020-03-17T14:02:06.630000", "content": "

I made this idea indexing in case for future for myself, sorry for overlaping yours ๐Ÿ˜‚

", "votes": 3, "replies": [] } ] }, { "id": 778187, "author_name": "Tawara", "author_url": "", "post_date": "2020-03-18T08:02:18.937000", "content": "

So nice summary! thanks for sharing.

", "votes": 1, "replies": [ { "id": 778460, "author_name": "Qishen Ha", "author_url": "", "post_date": "2020-03-18T13:13:00.267000", "content": "

Including yours. But I really don't know why your model works ๐Ÿ˜‚

", "votes": 1, "replies": [] }, { "id": 778466, "author_name": "Tawara", "author_url": "", "post_date": "2020-03-18T13:19:11.800000", "content": "

Haha! Me too.

\n\n

My model is very simple so it did not overfit to Public LB? I could not imagine such a incredible shakeup ๐Ÿ˜‚

", "votes": 1, "replies": [] } ] }, { "id": 778281, "author_name": "Dipam Chakraborty", "author_url": "", "post_date": "2020-03-18T10:00:14.963000", "content": "

Looks like this thread really needs updating ... not doing justice to the amazingly creative solutions like 5th place multilabel classification for consonants.

", "votes": 2, "replies": [ { "id": 778465, "author_name": "Qishen Ha", "author_url": "", "post_date": "2020-03-18T13:18:43.630000", "content": "

Thanks for your kindly advice as I actually cannot follow everything by myself.\nIf you have any other advice please let me know ;)

", "votes": 0, "replies": [] } ] }, { "id": 777720, "author_name": "", "author_url": "", "post_date": "2020-03-17T21:54:43.937000", "content": "", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "776074": "To be updated...\n\n\n# Main Ideas\n\nBasically we have 5 main ideas that works well in this competition as we can see now.\n\n\n1. You don't give me unseen graphemes? I'll make it by myself. (1st, 2nd, 7th)\n2. Figuring out unseen graphemes then predict them better. (3rd, 4th, 8th)\n3. Relabel consonant diacritic (5th)\n4. Using exists external data (9th)\n5. Postprocessing (5th, 6th, 13th, 14th)\n\n# Top Solutions\n\n* [1st Place Solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/135984) \n * CycleGAN\n* [2nd Place Solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/135966)\n * Generated grapheme and Improve C=3/6\n* [3rd Place Solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/135982)\n * Arcface detect unseen\n* [4th Place Solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/136982)\n * Arcface detect unseen\n* [5th Place Solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/136129)\n * Relabelled consonant diacritic (Improved C=3/6) & Postprocessing\n* [6th Place Solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/136011)\n * Postprocessing\n* [7th Place Solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/135960)\n * Make unseen grapheme from font.\n* [8th Place Solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/135990)\n * Arcface detect unseen\n* [9th Place Solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/135985)\n * External data /w pesudo label\n* [10th Place Solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/136815)\n * Best single model I've ever see. But why it works??? \n* [12th Place Solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/135998)\n * Many tricks\n* [13th Place Solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/136116)\n * Distribution postprocessing\n* [14th Place Solution](https://www.kaggle.com/c/bengaliai-cv19/discussion/136021)\n * Distribution postprocessing -- REAL MAGIC\n\n\n", "776703": "The diversity of top solutions in this competition is so interesting๐Ÿ˜ƒ Guess the hosts will be happy with a decent model after combining them", "776599": "I also made the summary before yours but this one is more elegant; I'll remove mine. ๐Ÿ˜€ \nand Congratulation again. @haqishen ๐Ÿ˜ƒ ", "778187": "So nice summary! thanks for sharing.", "778281": "Looks like this thread really needs updating ... not doing justice to the amazingly creative solutions like 5th place multilabel classification for consonants.", "777720": "" } }