{ "id": 301014, "title": "LB above 0.7! Wow!", "url": "/competitions/tensorflow-great-barrier-reef/discussion/301014", "author_name": "", "post_date": "2022-01-15T12:59:30.932417700Z", "votes": 42, "comment_count": 21, "views": 0, "content": "

Congratulations Team Hydrogen ( @philippsinger and @ybabakhin )…. first above 0.7! Massive! Outstanding! 👍👍👍😍😍👍👍

\n

I just saw how many effort you put in experimentation and …. eventually …. big jump. One question: how did you feel when you saw the result?

", "messages": [ { "id": "1650953", "postDate": "01/15/2022 12:59:30", "content": "

Congratulations Team Hydrogen ( @philippsinger and @ybabakhin )…. first above 0.7! Massive! Outstanding! 👍👍👍😍😍👍👍

\n

I just saw how many effort you put in experimentation and …. eventually …. big jump. One question: how did you feel when you saw the result?

", "rawMarkdown": "Congratulations Team Hydrogen ( @philippsinger and @ybabakhin ).... first above 0.7! Massive! Outstanding! 👍👍👍😍😍👍👍\n\nI just saw how many effort you put in experimentation and .... eventually .... big jump. One question: how did you feel when you saw the result?", "votes": null }, { "id": "1651004", "postDate": "01/15/2022 13:34:08", "content": "
\n

One question: how did you feel when you saw the result?

\n
\n

\"\"

", "rawMarkdown": "> One question: how did you feel when you saw the result?\n\n![](https://i.imgur.com/EWu31sr.png)", "votes": null }, { "id": "1651048", "postDate": "01/15/2022 13:55:14", "content": "

Somebody down voted my post … I understand … switching on silent mode in progress 🤣😂

", "rawMarkdown": "Somebody down voted my post … I understand … switching on silent mode in progress 🤣😂", "votes": null }, { "id": "1651102", "postDate": "01/15/2022 14:22:23", "content": "

Just ignore , some people will always downvote irrespective of content …

", "rawMarkdown": "Just ignore , some people will always downvote irrespective of content ...", "votes": null }, { "id": "1651153", "postDate": "01/15/2022 15:25:27", "content": "

Some people downvote every off topic content so don't get surprised. (I wasn't the downvoter)

", "rawMarkdown": "Some people downvote every off topic content so don't get surprised. (I wasn't the downvoter)", "votes": null }, { "id": "1651379", "postDate": "01/15/2022 17:56:18", "content": "

Yes, you know … haters do not understand this post (maybe this is too abstract question) … I asked important question … for this competition … If they search deeper into discussion forum maybe they will find answer (tip f2 competition score).

", "rawMarkdown": "Yes, you know … haters do not understand this post (maybe this is too abstract question) … I asked important question … for this competition … If they search deeper into discussion forum maybe they will find answer (tip f2 competition score).", "votes": null }, { "id": "1651392", "postDate": "01/15/2022 18:06:54", "content": "

But this post is connected with competition. Our 15 days long TOP1 I would describe very similar. We had totally different fillings about our model but it magically worked on LB. And please do not answer “because test is different”. We checked many, many possibilities (we develop score debugger to explain this phenomena) as a result we dramatically raised TP and lowered FP (on many folds configurations) but competition f2 score was not correlated with our local score. Not correlated is delicate word …

\n

Yes I know Team Hydrogen is absolutely better then we are (more experienced guys) … but … I am sure that they had locally “better” models as well :) (top 2 GMs and over 200 experiments). Maybe I am wrong … I do not know. If yes jus please fore give me.

", "rawMarkdown": "But this post is connected with competition. Our 15 days long TOP1 I would describe very similar. We had totally different fillings about our model but it magically worked on LB. And please do not answer “because test is different”. We checked many, many possibilities (we develop score debugger to explain this phenomena) as a result we dramatically raised TP and lowered FP (on many folds configurations) but competition f2 score was not correlated with our local score. Not correlated is delicate word … \n\nYes I know Team Hydrogen is absolutely better then we are (more experienced guys) … but … I am sure that they had locally “better” models as well :) (top 2 GMs and over 200 experiments). Maybe I am wrong … I do not know. If yes jus please fore give me.", "votes": null }, { "id": "1651435", "postDate": "01/15/2022 18:36:42", "content": "

It is just one side , I am sure you have plenty of killer model locally which completely embarrassed you in LB .

", "rawMarkdown": "It is just one side , I am sure you have plenty of killer model locally which completely embarrassed you in LB .", "votes": null }, { "id": "1651447", "postDate": "01/15/2022 18:45:35", "content": "

Yes! That’s right.
\nWe used many techniques to increase f2. Locally perfect (on many fold configuration - and on many video scenes - as a consequence we found many leaky splits). But LB was rather unpredictable… However we have many models above 0.65 … but with no correlation…

", "rawMarkdown": "Yes! That’s right. \nWe used many techniques to increase f2. Locally perfect (on many fold configuration - and on many video scenes - as a consequence we found many leaky splits). But LB was rather unpredictable… However we have many models above 0.65 … but with no correlation…", "votes": null }, { "id": "1651514", "postDate": "01/15/2022 20:16:27", "content": "

single metric may be unfair to measure model robustness.
\nmetric with weighted combination of Precision, Recall and F2 ?

", "rawMarkdown": "single metric may be unfair to measure model robustness. \nmetric with weighted combination of Precision, Recall and F2 ?", "votes": null }, { "id": "1651543", "postDate": "01/15/2022 20:50:34", "content": "

@dragonzhang F2 is a combination of precision and recall :) . Host wants F2 not F1 so he described what relation of precision and recall he favours

", "rawMarkdown": "dragonzhang F2 is a combination of precision and recall :) . Host wants F2 not F1 so he described what relation of precision and recall he favours", "votes": null }, { "id": "1652113", "postDate": "01/16/2022 10:56:31", "content": "

wow，amazing

", "rawMarkdown": "wow，amazing", "votes": null }, { "id": "1652226", "postDate": "01/16/2022 12:43:19", "content": "

for reality application, Recall is good for precaution cure than F1.

\n

for model performance measurement, weighted sum of F2, F1, R etc, better than a single biased F2.

", "rawMarkdown": "for reality application, Recall is good for precaution cure than F1.\n\nfor model performance measurement, weighted sum of F2, F1, R etc, better than a single biased F2.", "votes": null }, { "id": "1652234", "postDate": "01/16/2022 12:51:56", "content": "

https://www.youtube.com/watch?v=W5meQnGACGo

\n

Enjoy!

\n

And btw if you choose sub with f1 score best metric you will overshoot f2 peak ALWAYS. you can even make spreadsheet with prec recall values and corresponding f1 f2 score you will see f2 is lagged back.

", "rawMarkdown": "https://www.youtube.com/watch?v=W5meQnGACGo\n\nEnjoy!\n\nAnd btw if you choose sub with f1 score best metric you will overshoot f2 peak ALWAYS. you can even make spreadsheet with prec recall values and corresponding f1 f2 score you will see f2 is lagged back.", "votes": null }, { "id": "1652392", "postDate": "01/16/2022 15:07:31", "content": "

There was recently some high-level sharing of tips, so wouldn't be surprised if jump is due to combining their existing work with that in some way.

", "rawMarkdown": "There was recently some high-level sharing of tips, so wouldn't be surprised if jump is due to combining their existing work with that in some way.", "votes": null }, { "id": "1654846", "postDate": "01/18/2022 21:35:20", "content": "

@philippsinger you are really smart guys :) I am really impressed. I do not tell exactly why but …. you must have a lot of fun reading topics about inference on 10.000px images 😂😂😂

", "rawMarkdown": "philippsinger you are really smart guys :) I am really impressed. I do not tell exactly why but .... you must have a lot of fun reading topics about inference on 10.000px images 😂😂😂", "votes": null }, { "id": "1656383", "postDate": "01/19/2022 10:05:43", "content": "

really amazing congrats

", "rawMarkdown": "really amazing congrats", "votes": null }, { "id": "1656426", "postDate": "01/19/2022 10:48:52", "content": "

outrunned.. :)

\n

\"\"

", "rawMarkdown": "outrunned.. :)\n\n![](https://i.imgur.com/EWu31sr.png)", "votes": null }, { "id": "1656446", "postDate": "01/19/2022 10:54:07", "content": "

(copy my meme from below here)

", "rawMarkdown": "(copy my meme from below here)", "votes": null }, { "id": "1656460", "postDate": "01/19/2022 10:59:47", "content": "

there you go ;)

", "rawMarkdown": "there you go ;)", "votes": null }, { "id": "1658600", "postDate": "01/21/2022 06:38:01", "content": "

Congratulations!

", "rawMarkdown": "Congratulations!", "votes": null }, { "id": "1660932", "postDate": "01/23/2022 05:26:52", "content": "

Wow. Congratulations!!!

", "rawMarkdown": "Wow. Congratulations!!!", "votes": null } ], "comments": [ { "id": 1651004, "author_name": "philippsinger", "author_url": "", "post_date": "01/15/2022 13:34:08", "content": "
\n

One question: how did you feel when you saw the result?

\n
\n

\"\"

", "votes": null, "replies": [] }, { "id": 1651048, "author_name": "remekkinas", "author_url": "", "post_date": "01/15/2022 13:55:14", "content": "

Somebody down voted my post … I understand … switching on silent mode in progress 🤣😂

", "votes": null, "replies": [ { "id": 1651102, "author_name": "phoenix9032", "author_url": "", "post_date": "01/15/2022 14:22:23", "content": "

Just ignore , some people will always downvote irrespective of content …

", "votes": null, "replies": [] }, { "id": 1651153, "author_name": "gunesevitan", "author_url": "", "post_date": "01/15/2022 15:25:27", "content": "

Some people downvote every off topic content so don't get surprised. (I wasn't the downvoter)

", "votes": null, "replies": [] }, { "id": 1651379, "author_name": "remekkinas", "author_url": "", "post_date": "01/15/2022 17:56:18", "content": "

Yes, you know … haters do not understand this post (maybe this is too abstract question) … I asked important question … for this competition … If they search deeper into discussion forum maybe they will find answer (tip f2 competition score).

", "votes": null, "replies": [] }, { "id": 1651392, "author_name": "remekkinas", "author_url": "", "post_date": "01/15/2022 18:06:54", "content": "

But this post is connected with competition. Our 15 days long TOP1 I would describe very similar. We had totally different fillings about our model but it magically worked on LB. And please do not answer “because test is different”. We checked many, many possibilities (we develop score debugger to explain this phenomena) as a result we dramatically raised TP and lowered FP (on many folds configurations) but competition f2 score was not correlated with our local score. Not correlated is delicate word …

\n

Yes I know Team Hydrogen is absolutely better then we are (more experienced guys) … but … I am sure that they had locally “better” models as well :) (top 2 GMs and over 200 experiments). Maybe I am wrong … I do not know. If yes jus please fore give me.

", "votes": null, "replies": [] }, { "id": 1651435, "author_name": "phoenix9032", "author_url": "", "post_date": "01/15/2022 18:36:42", "content": "

It is just one side , I am sure you have plenty of killer model locally which completely embarrassed you in LB .

", "votes": null, "replies": [] }, { "id": 1651447, "author_name": "remekkinas", "author_url": "", "post_date": "01/15/2022 18:45:35", "content": "

Yes! That’s right.
\nWe used many techniques to increase f2. Locally perfect (on many fold configuration - and on many video scenes - as a consequence we found many leaky splits). But LB was rather unpredictable… However we have many models above 0.65 … but with no correlation…

", "votes": null, "replies": [] }, { "id": 1651514, "author_name": "dragonzhang", "author_url": "", "post_date": "01/15/2022 20:16:27", "content": "

single metric may be unfair to measure model robustness.
\nmetric with weighted combination of Precision, Recall and F2 ?

", "votes": null, "replies": [] }, { "id": 1651543, "author_name": "lukaszborecki", "author_url": "", "post_date": "01/15/2022 20:50:34", "content": "

@dragonzhang F2 is a combination of precision and recall :) . Host wants F2 not F1 so he described what relation of precision and recall he favours

", "votes": null, "replies": [] }, { "id": 1652226, "author_name": "dragonzhang", "author_url": "", "post_date": "01/16/2022 12:43:19", "content": "

for reality application, Recall is good for precaution cure than F1.

\n

for model performance measurement, weighted sum of F2, F1, R etc, better than a single biased F2.

", "votes": null, "replies": [] }, { "id": 1652234, "author_name": "lukaszborecki", "author_url": "", "post_date": "01/16/2022 12:51:56", "content": "

https://www.youtube.com/watch?v=W5meQnGACGo

\n

Enjoy!

\n

And btw if you choose sub with f1 score best metric you will overshoot f2 peak ALWAYS. you can even make spreadsheet with prec recall values and corresponding f1 f2 score you will see f2 is lagged back.

", "votes": null, "replies": [] } ] }, { "id": 1652113, "author_name": "", "author_url": "", "post_date": "01/16/2022 10:56:31", "content": "

wow，amazing

", "votes": null, "replies": [] }, { "id": 1652392, "author_name": "jacob34", "author_url": "", "post_date": "01/16/2022 15:07:31", "content": "

There was recently some high-level sharing of tips, so wouldn't be surprised if jump is due to combining their existing work with that in some way.

", "votes": null, "replies": [] }, { "id": 1654846, "author_name": "remekkinas", "author_url": "", "post_date": "01/18/2022 21:35:20", "content": "

@philippsinger you are really smart guys :) I am really impressed. I do not tell exactly why but …. you must have a lot of fun reading topics about inference on 10.000px images 😂😂😂

", "votes": null, "replies": [] }, { "id": 1656383, "author_name": "", "author_url": "", "post_date": "01/19/2022 10:05:43", "content": "

really amazing congrats

", "votes": null, "replies": [] }, { "id": 1656426, "author_name": "lukaszborecki", "author_url": "", "post_date": "01/19/2022 10:48:52", "content": "

outrunned.. :)

\n

\"\"

", "votes": null, "replies": [ { "id": 1656446, "author_name": "philippsinger", "author_url": "", "post_date": "01/19/2022 10:54:07", "content": "

(copy my meme from below here)

", "votes": null, "replies": [] }, { "id": 1656460, "author_name": "lukaszborecki", "author_url": "", "post_date": "01/19/2022 10:59:47", "content": "

there you go ;)

", "votes": null, "replies": [] } ] }, { "id": 1658600, "author_name": "dfced6", "author_url": "", "post_date": "01/21/2022 06:38:01", "content": "

Congratulations!

", "votes": null, "replies": [] }, { "id": 1660932, "author_name": "eugeneryu", "author_url": "", "post_date": "01/23/2022 05:26:52", "content": "

Wow. Congratulations!!!

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1650953": "Congratulations Team Hydrogen ( @philippsinger and @ybabakhin ).... first above 0.7! Massive! Outstanding! 👍👍👍😍😍👍👍\n\nI just saw how many effort you put in experimentation and .... eventually .... big jump. One question: how did you feel when you saw the result?", "1651004": "> One question: how did you feel when you saw the result?\n\n![](https://i.imgur.com/EWu31sr.png)", "1651048": "Somebody down voted my post … I understand … switching on silent mode in progress 🤣😂", "1651102": "Just ignore , some people will always downvote irrespective of content ...", "1651153": "Some people downvote every off topic content so don't get surprised. (I wasn't the downvoter)", "1651379": "Yes, you know … haters do not understand this post (maybe this is too abstract question) … I asked important question … for this competition … If they search deeper into discussion forum maybe they will find answer (tip f2 competition score).", "1651392": "But this post is connected with competition. Our 15 days long TOP1 I would describe very similar. We had totally different fillings about our model but it magically worked on LB. And please do not answer “because test is different”. We checked many, many possibilities (we develop score debugger to explain this phenomena) as a result we dramatically raised TP and lowered FP (on many folds configurations) but competition f2 score was not correlated with our local score. Not correlated is delicate word … \n\nYes I know Team Hydrogen is absolutely better then we are (more experienced guys) … but … I am sure that they had locally “better” models as well :) (top 2 GMs and over 200 experiments). Maybe I am wrong … I do not know. If yes jus please fore give me.", "1651435": "It is just one side , I am sure you have plenty of killer model locally which completely embarrassed you in LB .", "1651447": "Yes! That’s right. \nWe used many techniques to increase f2. Locally perfect (on many fold configuration - and on many video scenes - as a consequence we found many leaky splits). But LB was rather unpredictable… However we have many models above 0.65 … but with no correlation…", "1651514": "single metric may be unfair to measure model robustness. \nmetric with weighted combination of Precision, Recall and F2 ?", "1651543": "dragonzhang F2 is a combination of precision and recall :) . Host wants F2 not F1 so he described what relation of precision and recall he favours", "1652113": "wow，amazing", "1652226": "for reality application, Recall is good for precaution cure than F1.\n\nfor model performance measurement, weighted sum of F2, F1, R etc, better than a single biased F2.", "1652234": "https://www.youtube.com/watch?v=W5meQnGACGo\n\nEnjoy!\n\nAnd btw if you choose sub with f1 score best metric you will overshoot f2 peak ALWAYS. you can even make spreadsheet with prec recall values and corresponding f1 f2 score you will see f2 is lagged back.", "1652392": "There was recently some high-level sharing of tips, so wouldn't be surprised if jump is due to combining their existing work with that in some way.", "1654846": "philippsinger you are really smart guys :) I am really impressed. I do not tell exactly why but .... you must have a lot of fun reading topics about inference on 10.000px images 😂😂😂", "1656383": "really amazing congrats", "1656426": "outrunned.. :)\n\n![](https://i.imgur.com/EWu31sr.png)", "1656446": "(copy my meme from below here)", "1656460": "there you go ;)", "1658600": "Congratulations!", "1660932": "Wow. Congratulations!!!" }, "source": "meta" }