![\"\"](\"https://i0.wp.com/picsdownloadz.com/wp-content/uploads/2015/07/teamwork-funny-pics.jpg?fit=400%2C344&ssl=1\")
1 Million Dollars!!!😲
\n\nAt first, We try to put the same value.\nAccording to the public kernel, the score is as follows.
\n\nHappy Kaggling✨
", "messages": [ { "id": "693638", "postDate": "12/12/2019 15:50:30", "content": "1 Million Dollars!!!😲
\n\nAt first, We try to put the same value.\nAccording to the public kernel, the score is as follows.
\n\nHappy Kaggling✨
", "rawMarkdown": "1 Million Dollars!!!😲\n\nAt first, We try to put the same value.\nAccording to the public kernel, the score is as follows.\n\n- ALL **0** ---> **17.26938** Thanks @grapestone5321 \n- ALL **0.1** ---> **1.20397** Thanks @artgor\n- ALL **0.4** ---> **0.71355** @mashlyn\n- ALL **0.5** ---> **0.69314** Thanks @mmmarchetti @gpreda @paulorzp \n- ALL **0.51** ---> **0.69334** Thanks @timesler\n- ALL **0.6** ---> **0.71355** Thanks @minhtam \n- ALL **0.65** ---> **0.74030** Thanks @marcovasquez \n\nHappy Kaggling✨", "votes": null }, { "id": "693649", "postDate": "12/12/2019 16:07:48", "content": "Good job, you've done what I want, I'll try other value tomorrow😂 @mashlyn
", "rawMarkdown": "Good job, you've done what I want, I'll try other value tomorrow😂 @mashlyn", "votes": null }, { "id": "693668", "postDate": "12/12/2019 16:54:34", "content": "Is this the first Discussion thread inviting to collaborative Public Leaderboard probing?
\n\nJust joking.
\n\n
0.7 ---> 0.7834 from my kernel
", "rawMarkdown": "**0.7** ---> **0.7834** from my [kernel](https://www.kaggle.com/techytushar/frame-size-and-video-length-eda)", "votes": null }, { "id": "693934", "postDate": "12/13/2019 01:02:44", "content": "0.505 -> 0.69319 @mashlyn
", "rawMarkdown": "0.505 -> 0.69319 @mashlyn", "votes": null }, { "id": "693937", "postDate": "12/13/2019 01:08:17", "content": "0.501 -> 0.69314, actually 0.5 is the best. @mashlyn
", "rawMarkdown": "0.501 -> 0.69314, actually 0.5 is the best. @mashlyn", "votes": null }, { "id": "693949", "postDate": "12/13/2019 01:26:55", "content": "I'm not sure if I feel happy or sad to know that 0.69314 is ln(2). \nAlso, as a practice problem, you can show that you always get LogLoss = ln(2) when picking 0.5 for all preds. This result does not depend on the labels.
", "rawMarkdown": "I'm not sure if I feel happy or sad to know that 0.69314 is ln(2). \nAlso, as a practice problem, you can show that you always get LogLoss = ln(2) when picking 0.5 for all preds. This result does not depend on the labels.", "votes": null }, { "id": "693953", "postDate": "12/13/2019 01:40:59", "content": "I'm also thinking that since 0.6 and 0.4 give the same result, this implies count(real)= count(fake)\n(I proved that to myself on the back of an envelope)
", "rawMarkdown": "I'm also thinking that since 0.6 and 0.4 give the same result, this implies **count(real)= count(fake)**\n(I proved that to myself on the back of an envelope)", "votes": null }, { "id": "693962", "postDate": "12/13/2019 02:00:29", "content": "Finally, given my conclusion that count(real)= count(fake), I found that if you choose a value p (same value for all entries), then LogLoss = ln(1/sqrt(p(1-p))
\n\nSeems to work for a few values from above I've tried.
\n\nObviously, when p=0, we don't get infinite, which shows that an epsilon = 1e-15 is added.
\n\nFinally, the derivative of the LogLoss (when p is the same for all entries and count(real)= count(fake)) has the same sign as (2p-1), and therefore LogLoss(p) has a minimum at p=0.5.
\n\nConclusion: we got all the info we can get from this type of probing, i.e. that count(real)= count(fake)
", "rawMarkdown": "Finally, given my conclusion that **count(real)= count(fake)**, I found that if you choose a value p (same value for all entries), then LogLoss = ln(1/sqrt(p(1-p))\n\nSeems to work for a few values from above I've tried.\n\nObviously, when p=0, we don't get infinite, which shows that an epsilon = 1e-15 is added.\n\nFinally, the derivative of the LogLoss (when p is the same for all entries and count(real)= count(fake)) has the same sign as (2p-1), and therefore LogLoss(p) has a minimum at p=0.5.\n\n\n\n**Conclusion:** we got all the info we can get from this type of probing, i.e. that **count(real)= count(fake)**", "votes": null }, { "id": "693964", "postDate": "12/13/2019 02:05:14", "content": "Don't waste your subs, all the values are predictable at this point.
", "rawMarkdown": "Don't waste your subs, all the values are predictable at this point.", "votes": null }, { "id": "694007", "postDate": "12/13/2019 03:12:24", "content": "Thanks for your sharing, you give a deeper understand of loss function👍 @sdoria
", "rawMarkdown": "Thanks for your sharing, you give a deeper understand of loss function👍 @sdoria", "votes": null }, { "id": "738023", "postDate": "02/06/2020 04:01:04", "content": "All 1 ---> 17.26978 Why? Is that exactly 2000 REALs and 2000 FAKEs?
", "rawMarkdown": "All 1 ---> 17.26978 Why? Is that exactly 2000 REALs and 2000 FAKEs?", "votes": null }, { "id": "741322", "postDate": "02/10/2020 13:32:24", "content": "Keep in mind, it's only the publib LB probing, the distribution of the private LB can be completely different.
", "rawMarkdown": "Keep in mind, it's only the publib LB probing, the distribution of the private LB can be completely different.", "votes": null }, { "id": "755190", "postDate": "02/24/2020 14:42:29", "content": "Thanks for sharing @mashlyn
", "rawMarkdown": "Thanks for sharing @mashlyn", "votes": null }, { "id": "758212", "postDate": "02/27/2020 14:50:20", "content": "It looks like that yes
", "rawMarkdown": "It looks like that yes", "votes": null } ], "comments": [ { "id": 693649, "author_name": "diegojohnson", "author_url": "", "post_date": "12/12/2019 16:07:48", "content": "Good job, you've done what I want, I'll try other value tomorrow😂 @mashlyn
", "votes": null, "replies": [ { "id": 693964, "author_name": "sdoria", "author_url": "", "post_date": "12/13/2019 02:05:14", "content": "Don't waste your subs, all the values are predictable at this point.
", "votes": null, "replies": [] } ] }, { "id": 693668, "author_name": "gpreda", "author_url": "", "post_date": "12/12/2019 16:54:34", "content": "Is this the first Discussion thread inviting to collaborative Public Leaderboard probing?
\n\nJust joking.
\n\n
0.7 ---> 0.7834 from my kernel
", "votes": null, "replies": [] }, { "id": 693934, "author_name": "diegojohnson", "author_url": "", "post_date": "12/13/2019 01:02:44", "content": "0.505 -> 0.69319 @mashlyn
", "votes": null, "replies": [] }, { "id": 693937, "author_name": "diegojohnson", "author_url": "", "post_date": "12/13/2019 01:08:17", "content": "0.501 -> 0.69314, actually 0.5 is the best. @mashlyn
", "votes": null, "replies": [] }, { "id": 693949, "author_name": "sdoria", "author_url": "", "post_date": "12/13/2019 01:26:55", "content": "I'm not sure if I feel happy or sad to know that 0.69314 is ln(2). \nAlso, as a practice problem, you can show that you always get LogLoss = ln(2) when picking 0.5 for all preds. This result does not depend on the labels.
", "votes": null, "replies": [ { "id": 693953, "author_name": "sdoria", "author_url": "", "post_date": "12/13/2019 01:40:59", "content": "I'm also thinking that since 0.6 and 0.4 give the same result, this implies count(real)= count(fake)\n(I proved that to myself on the back of an envelope)
", "votes": null, "replies": [] }, { "id": 693962, "author_name": "sdoria", "author_url": "", "post_date": "12/13/2019 02:00:29", "content": "Finally, given my conclusion that count(real)= count(fake), I found that if you choose a value p (same value for all entries), then LogLoss = ln(1/sqrt(p(1-p))
\n\nSeems to work for a few values from above I've tried.
\n\nObviously, when p=0, we don't get infinite, which shows that an epsilon = 1e-15 is added.
\n\nFinally, the derivative of the LogLoss (when p is the same for all entries and count(real)= count(fake)) has the same sign as (2p-1), and therefore LogLoss(p) has a minimum at p=0.5.
\n\nConclusion: we got all the info we can get from this type of probing, i.e. that count(real)= count(fake)
", "votes": null, "replies": [] }, { "id": 694007, "author_name": "diegojohnson", "author_url": "", "post_date": "12/13/2019 03:12:24", "content": "Thanks for your sharing, you give a deeper understand of loss function👍 @sdoria
", "votes": null, "replies": [] }, { "id": 741322, "author_name": "dmytropoplavskiy", "author_url": "", "post_date": "02/10/2020 13:32:24", "content": "Keep in mind, it's only the publib LB probing, the distribution of the private LB can be completely different.
", "votes": null, "replies": [] } ] }, { "id": 738023, "author_name": "wuliaokaola", "author_url": "", "post_date": "02/06/2020 04:01:04", "content": "All 1 ---> 17.26978 Why? Is that exactly 2000 REALs and 2000 FAKEs?
", "votes": null, "replies": [ { "id": 758212, "author_name": "itamargr", "author_url": "", "post_date": "02/27/2020 14:50:20", "content": "It looks like that yes
", "votes": null, "replies": [] } ] }, { "id": 755190, "author_name": "dataraj", "author_url": "", "post_date": "02/24/2020 14:42:29", "content": "Thanks for sharing @mashlyn
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "693638": "1 Million Dollars!!!😲\n\nAt first, We try to put the same value.\nAccording to the public kernel, the score is as follows.\n\n- ALL **0** ---> **17.26938** Thanks @grapestone5321 \n- ALL **0.1** ---> **1.20397** Thanks @artgor\n- ALL **0.4** ---> **0.71355** @mashlyn\n- ALL **0.5** ---> **0.69314** Thanks @mmmarchetti @gpreda @paulorzp \n- ALL **0.51** ---> **0.69334** Thanks @timesler\n- ALL **0.6** ---> **0.71355** Thanks @minhtam \n- ALL **0.65** ---> **0.74030** Thanks @marcovasquez \n\nHappy Kaggling✨", "693649": "Good job, you've done what I want, I'll try other value tomorrow😂 @mashlyn", "693668": "Is this the first Discussion thread inviting to collaborative Public Leaderboard probing? \n\nJust joking.\n\n\n", "693864": "**0.7** ---> **0.7834** from my [kernel](https://www.kaggle.com/techytushar/frame-size-and-video-length-eda)", "693934": "0.505 -> 0.69319 @mashlyn", "693937": "0.501 -> 0.69314, actually 0.5 is the best. @mashlyn", "693949": "I'm not sure if I feel happy or sad to know that 0.69314 is ln(2). \nAlso, as a practice problem, you can show that you always get LogLoss = ln(2) when picking 0.5 for all preds. This result does not depend on the labels.", "693953": "I'm also thinking that since 0.6 and 0.4 give the same result, this implies **count(real)= count(fake)**\n(I proved that to myself on the back of an envelope)", "693962": "Finally, given my conclusion that **count(real)= count(fake)**, I found that if you choose a value p (same value for all entries), then LogLoss = ln(1/sqrt(p(1-p))\n\nSeems to work for a few values from above I've tried.\n\nObviously, when p=0, we don't get infinite, which shows that an epsilon = 1e-15 is added.\n\nFinally, the derivative of the LogLoss (when p is the same for all entries and count(real)= count(fake)) has the same sign as (2p-1), and therefore LogLoss(p) has a minimum at p=0.5.\n\n\n\n**Conclusion:** we got all the info we can get from this type of probing, i.e. that **count(real)= count(fake)**", "693964": "Don't waste your subs, all the values are predictable at this point.", "694007": "Thanks for your sharing, you give a deeper understand of loss function👍 @sdoria", "738023": "All 1 ---> 17.26978 Why? Is that exactly 2000 REALs and 2000 FAKEs?", "741322": "Keep in mind, it's only the publib LB probing, the distribution of the private LB can be completely different.", "755190": "Thanks for sharing @mashlyn", "758212": "It looks like that yes" }, "source": "meta" }