{ "id": 136027, "title": "what to do after competition?", "url": "/competitions/bengaliai-cv19/discussion/136027", "author_name": "hengck23", "post_date": "2020-03-17T05:11:04.946000", "votes": 79, "comment_count": 21, "views": 0, "content": "
    \n

  1. write a summary of your method and make a post. regardless of your rank, you can share your thoughts, experiences, results. *writing solution and sharing is part of data science and also part of kaggle!* it helps you to organize your thoughts as well.

    \n\n
    2. \n
\n\n
\n\n
    \n
  1. implement others solution. from the solution i read so far, this this what you can do:\n
    2. \n
\n\n

proceed these in steps. it is easy to come up with ideas, but the devils is in the details. if you want to learn magic, try their solutions , *till you get the same results*

\n\n

don't just run their opensource, try implement yourselves!

\n\n

you still can form a slack team inviting kagglers to spread out the massive work of repeating several top solutions, etc. you can even hold video conference to discuss results, etc

\n\n
\n\n
    \n
  1. apply what you have learned\n
    2. \n
\n\n

good luck!

", "messages": [ { "id": 776047, "postDate": "2020-03-17T05:11:04.947Z", "content": "
    \n

  1. write a summary of your method and make a post. regardless of your rank, you can share your thoughts, experiences, results. *writing solution and sharing is part of data science and also part of kaggle!* it helps you to organize your thoughts as well.

    \n\n
    2. \n
\n\n
\n\n
    \n
  1. implement others solution. from the solution i read so far, this this what you can do:\n
    2. \n
\n\n

proceed these in steps. it is easy to come up with ideas, but the devils is in the details. if you want to learn magic, try their solutions , *till you get the same results*

\n\n

don't just run their opensource, try implement yourselves!

\n\n

you still can form a slack team inviting kagglers to spread out the massive work of repeating several top solutions, etc. you can even hold video conference to discuss results, etc

\n\n
\n\n
    \n
  1. apply what you have learned\n
    2. \n
\n\n

good luck!

", "rawMarkdown": "1. write a summary of your method and make a post. regardless of your rank, you can share your thoughts, experiences, results. ****writing solution and sharing is part of data science and also part of kaggle!**** it helps you to organize your thoughts as well.\n\n- share what works and what doesn't\n- share code for people to reproduce your results (maybe others can help you ctach unknown bugs or suggest improvement)\n- training logs! this is important. it lets others compare your results against theirs for loss curve, etc\n- don't forget to share others like graphs, data split, notes, etc ...\n\n---\n\n2. implement others solution. from the solution i read so far, this this what you can do:\n- baseline : efficient net +augment (3 days)\n- using arcface for seen and unseen classification (3 days)\n- using GAN (3 days)\n\nproceed these in steps. it is easy to come up with ideas, but the devils is in the details. if you want to learn magic, try their solutions , ****till you get the same results****\n\ndon't just run their opensource, try implement yourselves!\n\nyou still can form a slack team inviting kagglers to spread out the massive work of repeating several top solutions, etc. you can even hold video conference to discuss results, etc\n\n---\n\n3. apply what you have learned\n- see if you make a post submission that is better than the first place solution\n- try these magic in other challenges\n- write a paper and put in arvix or sent to organizer\n\ngood luck!", "votes": 79 }, { "id": 776693, "postDate": "2020-03-17T15:05:58.447Z", "content": "

i am reimplementing some of the method. you can check my work at\nhttps://drive.google.com/drive/folders/1WdmK14BRPOBfIlaPnrEJABEtKepkWwmc

\n\n

there are quite a few i want to try, e.g. arcface, gan, etc

\n\n
    \n
  1. i start off with the 8th-place solution from qishen hai first.\n
    2. \n
\n\n

it is in early stages. i will update as it goes

", "rawMarkdown": "i am reimplementing some of the method. you can check my work at\nhttps://drive.google.com/drive/folders/1WdmK14BRPOBfIlaPnrEJABEtKepkWwmc\n\nthere are quite a few i want to try, e.g. arcface, gan, etc\n\n1. i start off with the 8th-place solution from qishen hai first.\n- implement arcface metric learning\n- try tSNE projection on embedding, etc\n\nit is in early stages. i will update as it goes\n", "votes": 5, "replies": [ { "id": 777509, "postDate": "2020-03-17T18:22:30.753Z", "content": "

@hengck23, thanks for this thoughtful post partner in drop:-)

\n\n

If you have the time I would like to team up with you in reproducing some of the top solutions as part of knowledge enhancement. I must warn you that I am kind of a beginner in pytorch. Having said that, in past competition, I have tried out a streamed line versions of your shared implementations on Colab. I do not have a GPU machine and I am abroad at the moment which makes it harder for me to use my cloud Floyd account. Having said all that, let me know if you are open to post competition colaboration.

", "rawMarkdown": "@hengck23, thanks for this thoughtful post partner in drop:-) \n\nIf you have the time I would like to team up with you in reproducing some of the top solutions as part of knowledge enhancement. I must warn you that I am kind of a beginner in pytorch. Having said that, in past competition, I have tried out a streamed line versions of your shared implementations on Colab. I do not have a GPU machine and I am abroad at the moment which makes it harder for me to use my cloud Floyd account. Having said all that, let me know if you are open to post competition colaboration." }, { "id": 778116, "postDate": "2020-03-18T06:56:26.527Z", "content": "

@sheriytm

\n\n

hi, you can start a slack team for that for discussion. basically i will put and re-implementation and experiments at google drive.

\n\n

my implement uses same framework, so it would be easier for you to use

", "rawMarkdown": "@sheriytm \n\nhi, you can start a slack team for that for discussion. basically i will put and re-implementation and experiments at google drive. \n\nmy implement uses same framework, so it would be easier for you to use" }, { "id": 778413, "postDate": "2020-03-18T12:20:18.690Z", "content": "

yes, arcface is a magic. here is the reimplementtion report\n\"\"

", "rawMarkdown": "yes, arcface is a magic. here is the reimplementtion report\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F113660%2F3b8a8557049813c0d3c62668b55b50a8%2FSelection_101.png?generation=1584533980782520&alt=media)\n", "votes": 4 }, { "id": 778431, "postDate": "2020-03-18T12:47:29.940Z", "content": "

@hengck23 How is your validation split? Is arcface still useful even if there are no unseen graphemes in validation set?

", "rawMarkdown": "@hengck23 How is your validation split? Is arcface still useful even if there are no unseen graphemes in validation set?", "votes": 1 }, { "id": 778444, "postDate": "2020-03-18T12:59:31.770Z", "content": "

\"Is arcface still useful even if there are no unseen graphemes\"

\n\n

i haven't test that yet. currently, it is shown that for the 1295 grapheme classes, arcface ha significant advantage

", "rawMarkdown": "\"Is arcface still useful even if there are no unseen graphemes\"\n\ni haven't test that yet. currently, it is shown that for the 1295 grapheme classes, arcface ha significant advantage" }, { "id": 778453, "postDate": "2020-03-18T13:07:16.760Z", "content": "

I wonder if using arcface as an additional loss would be useful for all future multilabel competitions, for example https://www.kaggle.com/c/iwildcam-2020-fgvc7 and https://www.kaggle.com/c/herbarium-2020-fgvc7 . It makes sense to use it here because you can use its output as a threshold to choose \"seen vs unseen model\". Maybe it acts as a really effective regularization technique that helps generalize.

", "rawMarkdown": "I wonder if using arcface as an additional loss would be useful for all future multilabel competitions, for example https://www.kaggle.com/c/iwildcam-2020-fgvc7 and https://www.kaggle.com/c/herbarium-2020-fgvc7 . It makes sense to use it here because you can use its output as a threshold to choose \"seen vs unseen model\". Maybe it acts as a really effective regularization technique that helps generalize." }, { "id": 778732, "postDate": "2020-03-18T17:20:39.583Z", "content": "

it is worth trying.

\n\n

i will update my graph later.\ncurrently my results showed all loss is \"much better\" using either one alone

", "rawMarkdown": "it is worth trying.\n\ni will update my graph later.\ncurrently my results showed all loss is \"much better\" using either one alone" } ] }, { "id": 776085, "postDate": "2020-03-17T05:41:41.733Z", "content": "

I started to write code for this competition last Friday with the help of Abhishek Thakur's youtube tutorial. I learnt few new things through the video and modified few things in last 2/3 days to improve the result.\nAnd today when I see the 1st/3rd place solutions, they are just amazing. Planning to reproduce the results in upcoming days.\nThanks for sharing the \"to dos after competition\".

", "rawMarkdown": "I started to write code for this competition last Friday with the help of Abhishek Thakur's youtube tutorial. I learnt few new things through the video and modified few things in last 2/3 days to improve the result.\nAnd today when I see the 1st/3rd place solutions, they are just amazing. Planning to reproduce the results in upcoming days.\nThanks for sharing the \"to dos after competition\".", "votes": 2 }, { "id": 776429, "postDate": "2020-03-17T11:41:46.143Z", "content": "

Also, I believe the techniques used in this competition will also help advance SOTA for other Indian languages, most of which are similar to Bengali. There are 15+ of them but I have no idea about datasets :P . As always, thanks Heng, yours and Qishen's insights were invaluable to myself and others.

", "rawMarkdown": "Also, I believe the techniques used in this competition will also help advance SOTA for other Indian languages, most of which are similar to Bengali. There are 15+ of them but I have no idea about datasets :P . As always, thanks Heng, yours and Qishen's insights were invaluable to myself and others.", "votes": 2 }, { "id": 781277, "postDate": "2020-03-21T05:16:30.380Z", "content": "

Thanks a lot for the suggestions. I will start working on implementing the top solutions soon. I've already published my codes for this competition here. Will keep updating.

", "rawMarkdown": "Thanks a lot for the suggestions. I will start working on implementing the top solutions soon. I've already published my codes for this competition [here](https://github.com/tahsin314/Bengali.ai-Handwritten-Character-Recognition-Challenge). Will keep updating. " }, { "id": 779311, "postDate": "2020-03-19T07:41:32.813Z", "content": "

Is it too much to expect the codes of top solutions?

", "rawMarkdown": "Is it too much to expect the codes of top solutions?" }, { "id": 778826, "postDate": "2020-03-18T19:17:53.393Z", "content": "

For how long is this competition open?

", "rawMarkdown": "For how long is this competition open?" }, { "id": 777863, "postDate": "2020-03-18T01:56:50.087Z", "content": "

Thanks, I could read many top solutions for this competition :)

", "rawMarkdown": "Thanks, I could read many top solutions for this competition :)" }, { "id": 777601, "postDate": "2020-03-17T20:06:05Z", "content": "

I hope I can help if it is possible, I am not here for money...

", "rawMarkdown": "I hope I can help if it is possible, I am not here for money..." }, { "id": 776862, "postDate": "2020-03-17T17:25:44.883Z", "content": "

Thanks for the info! Always great to have a guide for how to help out in the community!

", "rawMarkdown": "Thanks for the info! Always great to have a guide for how to help out in the community!" }, { "id": 778647, "postDate": "2020-03-18T16:00:26.110Z", "rawMarkdown": "", "isDeleted": true }, { "id": 776398, "postDate": "2020-03-17T11:08:54.923Z", "content": "

Wow, thanks!

", "rawMarkdown": "Wow, thanks!", "votes": 1 }, { "id": 776323, "postDate": "2020-03-17T10:00:11.157Z", "content": "

Thank a lot, frog brother @hengck23

", "rawMarkdown": "Thank a lot, frog brother @hengck23 ", "votes": 1 }, { "id": 779101, "postDate": "2020-03-19T02:54:22.743Z", "content": "

Learn a lot from you, thank you, Heng

", "rawMarkdown": "Learn a lot from you, thank you, Heng" }, { "id": 776215, "postDate": "2020-03-17T07:56:41.407Z", "content": "

Thanks @hengck23, will definetely do

", "rawMarkdown": "Thanks @hengck23, will definetely do" } ], "comments": [ { "id": 776693, "author_name": "hengck23", "author_url": "", "post_date": "2020-03-17T15:05:58.447000", "content": "

i am reimplementing some of the method. you can check my work at\nhttps://drive.google.com/drive/folders/1WdmK14BRPOBfIlaPnrEJABEtKepkWwmc

\n\n

there are quite a few i want to try, e.g. arcface, gan, etc

\n\n
    \n
  1. i start off with the 8th-place solution from qishen hai first.\n
    2. \n
\n\n

it is in early stages. i will update as it goes

", "votes": 5, "replies": [ { "id": 777509, "author_name": "YaGana Sheriff-Hussaini", "author_url": "", "post_date": "2020-03-17T18:22:30.753000", "content": "

@hengck23, thanks for this thoughtful post partner in drop:-)

\n\n

If you have the time I would like to team up with you in reproducing some of the top solutions as part of knowledge enhancement. I must warn you that I am kind of a beginner in pytorch. Having said that, in past competition, I have tried out a streamed line versions of your shared implementations on Colab. I do not have a GPU machine and I am abroad at the moment which makes it harder for me to use my cloud Floyd account. Having said all that, let me know if you are open to post competition colaboration.

", "votes": 0, "replies": [] }, { "id": 778116, "author_name": "hengck23", "author_url": "", "post_date": "2020-03-18T06:56:26.527000", "content": "

@sheriytm

\n\n

hi, you can start a slack team for that for discussion. basically i will put and re-implementation and experiments at google drive.

\n\n

my implement uses same framework, so it would be easier for you to use

", "votes": 0, "replies": [] }, { "id": 778413, "author_name": "hengck23", "author_url": "", "post_date": "2020-03-18T12:20:18.690000", "content": "

yes, arcface is a magic. here is the reimplementtion report\n\"\"

", "votes": 4, "replies": [] }, { "id": 778431, "author_name": "CoreyJamesLevinson", "author_url": "", "post_date": "2020-03-18T12:47:29.940000", "content": "

@hengck23 How is your validation split? Is arcface still useful even if there are no unseen graphemes in validation set?

", "votes": 1, "replies": [] }, { "id": 778444, "author_name": "hengck23", "author_url": "", "post_date": "2020-03-18T12:59:31.770000", "content": "

\"Is arcface still useful even if there are no unseen graphemes\"

\n\n

i haven't test that yet. currently, it is shown that for the 1295 grapheme classes, arcface ha significant advantage

", "votes": 0, "replies": [] }, { "id": 778453, "author_name": "CoreyJamesLevinson", "author_url": "", "post_date": "2020-03-18T13:07:16.760000", "content": "

I wonder if using arcface as an additional loss would be useful for all future multilabel competitions, for example https://www.kaggle.com/c/iwildcam-2020-fgvc7 and https://www.kaggle.com/c/herbarium-2020-fgvc7 . It makes sense to use it here because you can use its output as a threshold to choose \"seen vs unseen model\". Maybe it acts as a really effective regularization technique that helps generalize.

", "votes": 0, "replies": [] }, { "id": 778732, "author_name": "hengck23", "author_url": "", "post_date": "2020-03-18T17:20:39.583000", "content": "

it is worth trying.

\n\n

i will update my graph later.\ncurrently my results showed all loss is \"much better\" using either one alone

", "votes": 0, "replies": [] } ] }, { "id": 776085, "author_name": "jabertuhin", "author_url": "", "post_date": "2020-03-17T05:41:41.733000", "content": "

I started to write code for this competition last Friday with the help of Abhishek Thakur's youtube tutorial. I learnt few new things through the video and modified few things in last 2/3 days to improve the result.\nAnd today when I see the 1st/3rd place solutions, they are just amazing. Planning to reproduce the results in upcoming days.\nThanks for sharing the \"to dos after competition\".

", "votes": 2, "replies": [] }, { "id": 776429, "author_name": "Utsav Nandi", "author_url": "", "post_date": "2020-03-17T11:41:46.143000", "content": "

Also, I believe the techniques used in this competition will also help advance SOTA for other Indian languages, most of which are similar to Bengali. There are 15+ of them but I have no idea about datasets :P . As always, thanks Heng, yours and Qishen's insights were invaluable to myself and others.

", "votes": 2, "replies": [] }, { "id": 781277, "author_name": "Tahsin Mostafiz", "author_url": "", "post_date": "2020-03-21T05:16:30.380000", "content": "

Thanks a lot for the suggestions. I will start working on implementing the top solutions soon. I've already published my codes for this competition here. Will keep updating.

", "votes": 0, "replies": [] }, { "id": 779311, "author_name": "Tian", "author_url": "", "post_date": "2020-03-19T07:41:32.813000", "content": "

Is it too much to expect the codes of top solutions?

", "votes": 0, "replies": [] }, { "id": 778826, "author_name": "GAURAV B MACHAIAH", "author_url": "", "post_date": "2020-03-18T19:17:53.393000", "content": "

For how long is this competition open?

", "votes": 0, "replies": [] }, { "id": 777863, "author_name": "corochann", "author_url": "", "post_date": "2020-03-18T01:56:50.087000", "content": "

Thanks, I could read many top solutions for this competition :)

", "votes": 0, "replies": [] }, { "id": 777601, "author_name": "u8prototype", "author_url": "", "post_date": "2020-03-17T20:06:05", "content": "

I hope I can help if it is possible, I am not here for money...

", "votes": 0, "replies": [] }, { "id": 776862, "author_name": "Tyler Tsang", "author_url": "", "post_date": "2020-03-17T17:25:44.883000", "content": "

Thanks for the info! Always great to have a guide for how to help out in the community!

", "votes": 0, "replies": [] }, { "id": 778647, "author_name": "", "author_url": "", "post_date": "2020-03-18T16:00:26.110000", "content": "", "votes": 0, "replies": [] }, { "id": 776398, "author_name": "Ang Li", "author_url": "", "post_date": "2020-03-17T11:08:54.923000", "content": "

Wow, thanks!

", "votes": 1, "replies": [] }, { "id": 776323, "author_name": "Mohammad Azam Khan", "author_url": "", "post_date": "2020-03-17T10:00:11.157000", "content": "

Thank a lot, frog brother @hengck23

", "votes": 1, "replies": [] }, { "id": 779101, "author_name": "liuze", "author_url": "", "post_date": "2020-03-19T02:54:22.743000", "content": "

Learn a lot from you, thank you, Heng

", "votes": 0, "replies": [] }, { "id": 776215, "author_name": "Kurian Benoy", "author_url": "", "post_date": "2020-03-17T07:56:41.407000", "content": "

Thanks @hengck23, will definetely do

", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "776047": "1. write a summary of your method and make a post. regardless of your rank, you can share your thoughts, experiences, results. ****writing solution and sharing is part of data science and also part of kaggle!**** it helps you to organize your thoughts as well.\n\n- share what works and what doesn't\n- share code for people to reproduce your results (maybe others can help you ctach unknown bugs or suggest improvement)\n- training logs! this is important. it lets others compare your results against theirs for loss curve, etc\n- don't forget to share others like graphs, data split, notes, etc ...\n\n---\n\n2. implement others solution. from the solution i read so far, this this what you can do:\n- baseline : efficient net +augment (3 days)\n- using arcface for seen and unseen classification (3 days)\n- using GAN (3 days)\n\nproceed these in steps. it is easy to come up with ideas, but the devils is in the details. if you want to learn magic, try their solutions , ****till you get the same results****\n\ndon't just run their opensource, try implement yourselves!\n\nyou still can form a slack team inviting kagglers to spread out the massive work of repeating several top solutions, etc. you can even hold video conference to discuss results, etc\n\n---\n\n3. apply what you have learned\n- see if you make a post submission that is better than the first place solution\n- try these magic in other challenges\n- write a paper and put in arvix or sent to organizer\n\ngood luck!", "776693": "i am reimplementing some of the method. you can check my work at\nhttps://drive.google.com/drive/folders/1WdmK14BRPOBfIlaPnrEJABEtKepkWwmc\n\nthere are quite a few i want to try, e.g. arcface, gan, etc\n\n1. i start off with the 8th-place solution from qishen hai first.\n- implement arcface metric learning\n- try tSNE projection on embedding, etc\n\nit is in early stages. i will update as it goes\n", "776085": "I started to write code for this competition last Friday with the help of Abhishek Thakur's youtube tutorial. I learnt few new things through the video and modified few things in last 2/3 days to improve the result.\nAnd today when I see the 1st/3rd place solutions, they are just amazing. Planning to reproduce the results in upcoming days.\nThanks for sharing the \"to dos after competition\".", "776429": "Also, I believe the techniques used in this competition will also help advance SOTA for other Indian languages, most of which are similar to Bengali. There are 15+ of them but I have no idea about datasets :P . As always, thanks Heng, yours and Qishen's insights were invaluable to myself and others.", "781277": "Thanks a lot for the suggestions. I will start working on implementing the top solutions soon. I've already published my codes for this competition [here](https://github.com/tahsin314/Bengali.ai-Handwritten-Character-Recognition-Challenge). Will keep updating. ", "779311": "Is it too much to expect the codes of top solutions?", "778826": "For how long is this competition open?", "777863": "Thanks, I could read many top solutions for this competition :)", "777601": "I hope I can help if it is possible, I am not here for money...", "776862": "Thanks for the info! Always great to have a guide for how to help out in the community!", "778647": "", "776398": "Wow, thanks!", "776323": "Thank a lot, frog brother @hengck23 ", "779101": "Learn a lot from you, thank you, Heng", "776215": "Thanks @hengck23, will definetely do" } }