{ "id": 174897, "title": "The best single model CV/LB score?", "url": "/competitions/siim-isic-melanoma-classification/discussion/174897", "author_name": "gao-hongnan", "post_date": "2020-08-16T04:25:32.127000", "votes": 5, "comment_count": 37, "views": 0, "content": "

For me, the best single model (5-fold) scored a
\nCV: 0.9570
\nLB: 0.9590

\n

Not sure if this is good for a single model (no meta blend), what are you guys best score for single model only? Hopefully, this kind of post is not frowned upon!

\n

oops, I just realized there is already a post on it. Click here for the link

", "messages": [ { "id": 971914, "postDate": "2020-08-16T04:25:32.127Z", "content": "

For me, the best single model (5-fold) scored a
\nCV: 0.9570
\nLB: 0.9590

\n

Not sure if this is good for a single model (no meta blend), what are you guys best score for single model only? Hopefully, this kind of post is not frowned upon!

\n

oops, I just realized there is already a post on it. Click here for the link

", "rawMarkdown": "For me, the best single model (5-fold) scored a \nCV: 0.9570\nLB: 0.9590\n\nNot sure if this is good for a single model (no meta blend), what are you guys best score for single model only? Hopefully, this kind of post is not frowned upon!\n\noops, I just realized there is already a post on it. [Click here for the link](https://www.kaggle.com/c/siim-isic-melanoma-classification/discussion/156027)", "votes": 5 }, { "id": 972232, "postDate": "2020-08-16T11:09:39.120Z", "content": "

0.9483 CV -- 0.9467 LB (Only 2020 data in validation)
\nBut I'm only using 3-folds in all my models.

", "rawMarkdown": "0.9483 CV -- 0.9467 LB (Only 2020 data in validation)\nBut I'm only using 3-folds in all my models. ", "votes": 3, "replies": [ { "id": 972538, "postDate": "2020-08-16T16:12:03.627Z", "content": "

Great! You have strong single model!
\nSeems a bit strange to me that your CV is so close to LB, which quite rare as many others reported.

", "rawMarkdown": "Great! You have strong single model! \nSeems a bit strange to me that your CV is so close to LB, which quite rare as many others reported." }, { "id": 972569, "postDate": "2020-08-16T16:38:54.373Z", "content": "

That's not my single model :p that's the ensemble
\nMy single is 0.934CV -- 0.949LB

", "rawMarkdown": "That's not my single model :p that's the ensemble\nMy single is 0.934CV -- 0.949LB", "votes": 1 }, { "id": 972574, "postDate": "2020-08-16T16:42:11.563Z", "content": "

@seif95
\nstill strong one! Goodluck for the Private LB 🤘

", "rawMarkdown": "@seif95 \nstill strong one! Goodluck for the Private LB 🤘" }, { "id": 972921, "postDate": "2020-08-17T01:23:26.450Z", "content": "

Strong model. Great job!

", "rawMarkdown": "Strong model. Great job!", "votes": 1, "replies": [ { "id": 972927, "postDate": "2020-08-17T01:28:52.147Z", "content": "

I got a CV0.94 LB 0.9503 from your notebook :) not sure if its good since there's a 0.01 gap in between, but thats the best so far using chris's notebook

", "rawMarkdown": "I got a CV0.94 LB 0.9503 from your notebook :) not sure if its good since there's a 0.01 gap in between, but thats the best so far using chris's notebook" }, { "id": 972932, "postDate": "2020-08-17T01:31:26.200Z", "content": "

Anything CV 0.940+ is great. Good job!

", "rawMarkdown": "Anything CV 0.940+ is great. Good job!", "votes": 1 }, { "id": 972936, "postDate": "2020-08-17T01:33:36.983Z", "content": "

@cdeotte Saw your post on your first failure in your first comp, what an inspiration! I hope to be like you one day… 10 months into ML field and already addicted to it!

", "rawMarkdown": "@cdeotte Saw your post on your first failure in your first comp, what an inspiration! I hope to be like you one day... 10 months into ML field and already addicted to it!" } ] } ] }, { "id": 973468, "postDate": "2020-08-17T10:13:39.993Z", "content": "

With effnet b5 I got :
\nCV 0.942 LB 0.9578

", "rawMarkdown": "With effnet b5 I got :\nCV 0.942 LB 0.9578", "votes": 1 }, { "id": 973090, "postDate": "2020-08-17T05:16:08.830Z", "content": "

Best Single model CV : 95.2 and LB 94.48 with 3 folds

", "rawMarkdown": "Best Single model CV : 95.2 and LB 94.48 with 3 folds", "votes": 1 }, { "id": 972084, "postDate": "2020-08-16T08:10:54.827Z", "content": "

Wow! That's a great score. Are you using any external data other than the ones given by Chris?

", "rawMarkdown": "Wow! That's a great score. Are you using any external data other than the ones given by Chris?", "votes": 1, "replies": [ { "id": 972117, "postDate": "2020-08-16T08:48:33.413Z", "content": "

2019+2020 images just these 2

", "rawMarkdown": "2019+2020 images just these 2", "votes": 1 }, { "id": 972160, "postDate": "2020-08-16T09:29:41.493Z", "content": "

Thanks!!!!

", "rawMarkdown": "Thanks!!!!", "votes": 1 }, { "id": 972161, "postDate": "2020-08-16T09:29:41.517Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 972004, "postDate": "2020-08-16T06:47:50.943Z", "content": "

I think your folds are very different from mine. If you used triple stratified k-fold. It's impossible to get >.95 single model. Thus the comparison is pointless.

", "rawMarkdown": "I think your folds are very different from mine. If you used triple stratified k-fold. It's impossible to get >.95 single model. Thus the comparison is pointless. ", "votes": 1, "replies": [ { "id": 972012, "postDate": "2020-08-16T06:52:45.337Z", "content": "

@veryrobustperson correct me if i'm wrong, @chris tfrecords are already triple stratified, therefore I think that performing a KFold on his tfrecords (stratified) is also triple stratified.

", "rawMarkdown": "@veryrobustperson correct me if i'm wrong, @chris tfrecords are already triple stratified, therefore I think that performing a KFold on his tfrecords (stratified) is also triple stratified. ", "votes": 1 }, { "id": 972016, "postDate": "2020-08-16T06:59:35.800Z", "content": "

Wow i'd say that's an amazing model then. I wish you the best when the private leaderboard reveals. I am more than happy to learn from you. I tried many many techniques yet still can't break 940 CV and your CV is near 960.

", "rawMarkdown": "Wow i'd say that's an amazing model then. I wish you the best when the private leaderboard reveals. I am more than happy to learn from you. I tried many many techniques yet still can't break 940 CV and your CV is near 960.", "votes": 1, "replies": [ { "id": 972017, "postDate": "2020-08-16T07:02:08.857Z", "content": "

@veryrobustperson We learn from each other and I feel you are already better than me :) I may be wrong and overfitting is an issue as well. But I did tune some parameters that are different from Chris's original notebook…

", "rawMarkdown": "@veryrobustperson We learn from each other and I feel you are already better than me :) I may be wrong and overfitting is an issue as well. But I did tune some parameters that are different from Chris's original notebook...", "votes": 1 }, { "id": 972019, "postDate": "2020-08-16T07:05:26.273Z", "content": "

I'd be a little skeptical. If you may open-source your code after the competition, I am more than happy to learn!

", "rawMarkdown": "I'd be a little skeptical. If you may open-source your code after the competition, I am more than happy to learn! ", "votes": 1 }, { "id": 972320, "postDate": "2020-08-16T13:13:33.663Z", "rawMarkdown": "", "votes": 1, "isDeleted": true } ] } ] }, { "id": 971965, "postDate": "2020-08-16T06:03:39.463Z", "content": "

@reighns, a thread for the same topic is here - https://www.kaggle.com/c/siim-isic-melanoma-classification/discussion/156027, no need for a new one.

", "rawMarkdown": "@reighns, a thread for the same topic is here - https://www.kaggle.com/c/siim-isic-melanoma-classification/discussion/156027, no need for a new one.", "votes": 1, "replies": [ { "id": 971981, "postDate": "2020-08-16T06:21:47.817Z", "content": "

@sheriytm Sorry I totally missed that, my bad; but since the comp is ending soon and that topic wasn't on top of the list of discussions, let me just link the thread to my link so more people can see :) Thanks again for pointing it out!

", "rawMarkdown": "@sheriytm Sorry I totally missed that, my bad; but since the comp is ending soon and that topic wasn't on top of the list of discussions, let me just link the thread to my link so more people can see :) Thanks again for pointing it out!" }, { "id": 972514, "postDate": "2020-08-16T15:53:37.620Z", "content": "

No worries @reighns. It is a good practice to search the discussion section for the topic you have in mind before posting. People get really annoyed with duplicate topics and downvote. I learnt it the hard way when I was a NOOB on Kaggle. Since down voting a not so obvious topic does not tell the poster why, I just comment whenever possible.

", "rawMarkdown": "No worries @reighns. It is a good practice to search the discussion section for the topic you have in mind before posting. People get really annoyed with duplicate topics and downvote. I learnt it the hard way when I was a NOOB on Kaggle. Since down voting a not so obvious topic does not tell the poster why, I just comment whenever possible." } ] }, { "id": 973044, "postDate": "2020-08-17T04:20:35.690Z", "content": "

Single Models: (Best CV: 0.941-0.942) There are 3 models in this range, LB varies from 0.945 to 0.951.
\nEnsembles: (Simple Mean of best (8) Models: 0.9512 CV and 0.9513 LB ; Stack of 40 models: 0.9529 CV and 0.9548 LB)

", "rawMarkdown": "Single Models: (Best CV: 0.941-0.942) There are 3 models in this range, LB varies from 0.945 to 0.951.\nEnsembles: (Simple Mean of best (8) Models: 0.9512 CV and 0.9513 LB ; Stack of 40 models: 0.9529 CV and 0.9548 LB)", "votes": 2, "replies": [ { "id": 973045, "postDate": "2020-08-17T04:22:45.177Z", "content": "

Fantastic!

", "rawMarkdown": "Fantastic!", "votes": 2 } ] }, { "id": 971975, "postDate": "2020-08-16T06:16:46.797Z", "content": "

that's the crazy CV score without no ensemble
\nMine CV:0.9390 LB : 0.9488(Single Model)

", "rawMarkdown": "that's the crazy CV score without no ensemble\nMine CV:0.9390 LB : 0.9488(Single Model)\n", "votes": 2, "replies": [ { "id": 971978, "postDate": "2020-08-16T06:20:06.487Z", "content": "

@deepkim Yea, but its 5 folds :) yours also?

", "rawMarkdown": "@deepkim Yea, but its 5 folds :) yours also?", "votes": 1 }, { "id": 972021, "postDate": "2020-08-16T07:07:01.520Z", "content": "

Right ! Stratified 5 fold just on isic2020

", "rawMarkdown": "Right ! Stratified 5 fold just on isic2020 ", "votes": 1 } ] }, { "id": 973166, "postDate": "2020-08-17T06:34:05.943Z", "content": "

CV with TTA = 0.921
\nLB = 0.9464
\nadd meta + 1 model
\nLB = 0.9592

\n

That was my last submission, GL to everyone

", "rawMarkdown": "CV with TTA = 0.921\nLB = 0.9464\nadd meta + 1 model \nLB = 0.9592\n\nThat was my last submission, GL to everyone", "replies": [ { "id": 973202, "postDate": "2020-08-17T07:01:19.867Z", "content": "

How do you add meta?

", "rawMarkdown": "How do you add meta? " }, { "id": 973205, "postDate": "2020-08-17T07:05:52.570Z", "content": "

weight*meta_model['target'] + …

", "rawMarkdown": "weight*meta_model['target'] + ...", "votes": 2 } ] }, { "id": 972295, "postDate": "2020-08-16T12:42:19.223Z", "content": "

cv 0.931, lb 0.9417, it seems like many of my single model get a relatively good lb score (maximum 0.9488), but cv not.
\nAnd when I try to do some ensemble based on cv, lb not increase much or even decrease.

", "rawMarkdown": "cv 0.931, lb 0.9417, it seems like many of my single model get a relatively good lb score (maximum 0.9488), but cv not.\nAnd when I try to do some ensemble based on cv, lb not increase much or even decrease.\n", "replies": [ { "id": 972369, "postDate": "2020-08-16T13:51:13.490Z", "content": "

Ensembling in the right way will increase your cv score but will reduce your lb score in this competition. But you will have close lb and cv scores and this is a good sign for me. That means your work generalize much better than 0.960LB - 0.920 CV

", "rawMarkdown": "Ensembling in the right way will increase your cv score but will reduce your lb score in this competition. But you will have close lb and cv scores and this is a good sign for me. That means your work generalize **much** better than 0.960LB - 0.920 CV", "votes": 1 }, { "id": 972920, "postDate": "2020-08-17T01:23:05.103Z", "content": "

That's weird that ensemble based on CV doesn't increase LB. Are you using the same folds for all your models before CV ensemble? For me, ensemble CV increases LB.

", "rawMarkdown": "That's weird that ensemble based on CV doesn't increase LB. Are you using the same folds for all your models before CV ensemble? For me, ensemble CV increases LB.", "votes": 1 }, { "id": 973401, "postDate": "2020-08-17T09:28:57.273Z", "content": "

Yep I am using the same folds for all my models before CV ensemble.

", "rawMarkdown": "Yep I am using the same folds for all my models before CV ensemble." } ] }, { "id": 972083, "postDate": "2020-08-16T08:10:54.810Z", "rawMarkdown": "", "isDeleted": true }, { "id": 971946, "postDate": "2020-08-16T05:38:09.753Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 972232, "author_name": "Seifeddine Fezzani", "author_url": "", "post_date": "2020-08-16T11:09:39.120000", "content": "

0.9483 CV -- 0.9467 LB (Only 2020 data in validation)
\nBut I'm only using 3-folds in all my models.

", "votes": 3, "replies": [ { "id": 972538, "author_name": "Bayartsogt Yadamsuren", "author_url": "", "post_date": "2020-08-16T16:12:03.627000", "content": "

Great! You have strong single model!
\nSeems a bit strange to me that your CV is so close to LB, which quite rare as many others reported.

", "votes": 0, "replies": [] }, { "id": 972569, "author_name": "Seifeddine Fezzani", "author_url": "", "post_date": "2020-08-16T16:38:54.373000", "content": "

That's not my single model :p that's the ensemble
\nMy single is 0.934CV -- 0.949LB

", "votes": 1, "replies": [] }, { "id": 972574, "author_name": "Bayartsogt Yadamsuren", "author_url": "", "post_date": "2020-08-16T16:42:11.563000", "content": "

@seif95
\nstill strong one! Goodluck for the Private LB 🤘

", "votes": 0, "replies": [] }, { "id": 972921, "author_name": "Chris Deotte", "author_url": "", "post_date": "2020-08-17T01:23:26.450000", "content": "

Strong model. Great job!

", "votes": 1, "replies": [ { "id": 972927, "author_name": "gao-hongnan", "author_url": "", "post_date": "2020-08-17T01:28:52.147000", "content": "

I got a CV0.94 LB 0.9503 from your notebook :) not sure if its good since there's a 0.01 gap in between, but thats the best so far using chris's notebook

", "votes": 0, "replies": [] }, { "id": 972932, "author_name": "Chris Deotte", "author_url": "", "post_date": "2020-08-17T01:31:26.200000", "content": "

Anything CV 0.940+ is great. Good job!

", "votes": 1, "replies": [] }, { "id": 972936, "author_name": "gao-hongnan", "author_url": "", "post_date": "2020-08-17T01:33:36.983000", "content": "

@cdeotte Saw your post on your first failure in your first comp, what an inspiration! I hope to be like you one day… 10 months into ML field and already addicted to it!

", "votes": 0, "replies": [] } ] } ] }, { "id": 973468, "author_name": "Changyi", "author_url": "", "post_date": "2020-08-17T10:13:39.993000", "content": "

With effnet b5 I got :
\nCV 0.942 LB 0.9578

", "votes": 1, "replies": [] }, { "id": 973090, "author_name": "ashok", "author_url": "", "post_date": "2020-08-17T05:16:08.830000", "content": "

Best Single model CV : 95.2 and LB 94.48 with 3 folds

", "votes": 1, "replies": [] }, { "id": 972084, "author_name": "Jose", "author_url": "", "post_date": "2020-08-16T08:10:54.827000", "content": "

Wow! That's a great score. Are you using any external data other than the ones given by Chris?

", "votes": 1, "replies": [ { "id": 972117, "author_name": "gao-hongnan", "author_url": "", "post_date": "2020-08-16T08:48:33.413000", "content": "

2019+2020 images just these 2

", "votes": 1, "replies": [] }, { "id": 972160, "author_name": "Jose", "author_url": "", "post_date": "2020-08-16T09:29:41.493000", "content": "

Thanks!!!!

", "votes": 1, "replies": [] }, { "id": 972161, "author_name": "", "author_url": "", "post_date": "2020-08-16T09:29:41.517000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 972004, "author_name": "cat", "author_url": "", "post_date": "2020-08-16T06:47:50.943000", "content": "

I think your folds are very different from mine. If you used triple stratified k-fold. It's impossible to get >.95 single model. Thus the comparison is pointless.

", "votes": 1, "replies": [ { "id": 972012, "author_name": "gao-hongnan", "author_url": "", "post_date": "2020-08-16T06:52:45.337000", "content": "

@veryrobustperson correct me if i'm wrong, @chris tfrecords are already triple stratified, therefore I think that performing a KFold on his tfrecords (stratified) is also triple stratified.

", "votes": 1, "replies": [] }, { "id": 972016, "author_name": "cat", "author_url": "", "post_date": "2020-08-16T06:59:35.800000", "content": "

Wow i'd say that's an amazing model then. I wish you the best when the private leaderboard reveals. I am more than happy to learn from you. I tried many many techniques yet still can't break 940 CV and your CV is near 960.

", "votes": 1, "replies": [ { "id": 972017, "author_name": "gao-hongnan", "author_url": "", "post_date": "2020-08-16T07:02:08.857000", "content": "

@veryrobustperson We learn from each other and I feel you are already better than me :) I may be wrong and overfitting is an issue as well. But I did tune some parameters that are different from Chris's original notebook…

", "votes": 1, "replies": [] }, { "id": 972019, "author_name": "cat", "author_url": "", "post_date": "2020-08-16T07:05:26.273000", "content": "

I'd be a little skeptical. If you may open-source your code after the competition, I am more than happy to learn!

", "votes": 1, "replies": [] }, { "id": 972320, "author_name": "", "author_url": "", "post_date": "2020-08-16T13:13:33.663000", "content": "", "votes": 1, "replies": [] } ] } ] }, { "id": 971965, "author_name": "YaGana Sheriff-Hussaini", "author_url": "", "post_date": "2020-08-16T06:03:39.463000", "content": "

@reighns, a thread for the same topic is here - https://www.kaggle.com/c/siim-isic-melanoma-classification/discussion/156027, no need for a new one.

", "votes": 1, "replies": [ { "id": 971981, "author_name": "gao-hongnan", "author_url": "", "post_date": "2020-08-16T06:21:47.817000", "content": "

@sheriytm Sorry I totally missed that, my bad; but since the comp is ending soon and that topic wasn't on top of the list of discussions, let me just link the thread to my link so more people can see :) Thanks again for pointing it out!

", "votes": 0, "replies": [] }, { "id": 972514, "author_name": "YaGana Sheriff-Hussaini", "author_url": "", "post_date": "2020-08-16T15:53:37.620000", "content": "

No worries @reighns. It is a good practice to search the discussion section for the topic you have in mind before posting. People get really annoyed with duplicate topics and downvote. I learnt it the hard way when I was a NOOB on Kaggle. Since down voting a not so obvious topic does not tell the poster why, I just comment whenever possible.

", "votes": 0, "replies": [] } ] }, { "id": 973044, "author_name": "Kumar Shubham", "author_url": "", "post_date": "2020-08-17T04:20:35.690000", "content": "

Single Models: (Best CV: 0.941-0.942) There are 3 models in this range, LB varies from 0.945 to 0.951.
\nEnsembles: (Simple Mean of best (8) Models: 0.9512 CV and 0.9513 LB ; Stack of 40 models: 0.9529 CV and 0.9548 LB)

", "votes": 2, "replies": [ { "id": 973045, "author_name": "Chris Deotte", "author_url": "", "post_date": "2020-08-17T04:22:45.177000", "content": "

Fantastic!

", "votes": 2, "replies": [] } ] }, { "id": 971975, "author_name": "kaggler", "author_url": "", "post_date": "2020-08-16T06:16:46.797000", "content": "

that's the crazy CV score without no ensemble
\nMine CV:0.9390 LB : 0.9488(Single Model)

", "votes": 2, "replies": [ { "id": 971978, "author_name": "gao-hongnan", "author_url": "", "post_date": "2020-08-16T06:20:06.487000", "content": "

@deepkim Yea, but its 5 folds :) yours also?

", "votes": 1, "replies": [] }, { "id": 972021, "author_name": "kaggler", "author_url": "", "post_date": "2020-08-16T07:07:01.520000", "content": "

Right ! Stratified 5 fold just on isic2020

", "votes": 1, "replies": [] } ] }, { "id": 973166, "author_name": "Iurii Uspangaliev", "author_url": "", "post_date": "2020-08-17T06:34:05.943000", "content": "

CV with TTA = 0.921
\nLB = 0.9464
\nadd meta + 1 model
\nLB = 0.9592

\n

That was my last submission, GL to everyone

", "votes": 0, "replies": [ { "id": 973202, "author_name": "Dieter", "author_url": "", "post_date": "2020-08-17T07:01:19.867000", "content": "

How do you add meta?

", "votes": 0, "replies": [] }, { "id": 973205, "author_name": "Iurii Uspangaliev", "author_url": "", "post_date": "2020-08-17T07:05:52.570000", "content": "

weight*meta_model['target'] + …

", "votes": 2, "replies": [] } ] }, { "id": 972295, "author_name": "ElaborateJia", "author_url": "", "post_date": "2020-08-16T12:42:19.223000", "content": "

cv 0.931, lb 0.9417, it seems like many of my single model get a relatively good lb score (maximum 0.9488), but cv not.
\nAnd when I try to do some ensemble based on cv, lb not increase much or even decrease.

", "votes": 0, "replies": [ { "id": 972369, "author_name": "Seifeddine Fezzani", "author_url": "", "post_date": "2020-08-16T13:51:13.490000", "content": "

Ensembling in the right way will increase your cv score but will reduce your lb score in this competition. But you will have close lb and cv scores and this is a good sign for me. That means your work generalize much better than 0.960LB - 0.920 CV

", "votes": 1, "replies": [] }, { "id": 972920, "author_name": "Chris Deotte", "author_url": "", "post_date": "2020-08-17T01:23:05.103000", "content": "

That's weird that ensemble based on CV doesn't increase LB. Are you using the same folds for all your models before CV ensemble? For me, ensemble CV increases LB.

", "votes": 1, "replies": [] }, { "id": 973401, "author_name": "Seifeddine Fezzani", "author_url": "", "post_date": "2020-08-17T09:28:57.273000", "content": "

Yep I am using the same folds for all my models before CV ensemble.

", "votes": 0, "replies": [] } ] }, { "id": 972083, "author_name": "", "author_url": "", "post_date": "2020-08-16T08:10:54.810000", "content": "", "votes": 0, "replies": [] }, { "id": 971946, "author_name": "", "author_url": "", "post_date": "2020-08-16T05:38:09.753000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "971914": "For me, the best single model (5-fold) scored a \nCV: 0.9570\nLB: 0.9590\n\nNot sure if this is good for a single model (no meta blend), what are you guys best score for single model only? Hopefully, this kind of post is not frowned upon!\n\noops, I just realized there is already a post on it. [Click here for the link](https://www.kaggle.com/c/siim-isic-melanoma-classification/discussion/156027)", "972232": "0.9483 CV -- 0.9467 LB (Only 2020 data in validation)\nBut I'm only using 3-folds in all my models. ", "973468": "With effnet b5 I got :\nCV 0.942 LB 0.9578", "973090": "Best Single model CV : 95.2 and LB 94.48 with 3 folds", "972084": "Wow! That's a great score. Are you using any external data other than the ones given by Chris?", "972004": "I think your folds are very different from mine. If you used triple stratified k-fold. It's impossible to get >.95 single model. Thus the comparison is pointless. ", "971965": "@reighns, a thread for the same topic is here - https://www.kaggle.com/c/siim-isic-melanoma-classification/discussion/156027, no need for a new one.", "973044": "Single Models: (Best CV: 0.941-0.942) There are 3 models in this range, LB varies from 0.945 to 0.951.\nEnsembles: (Simple Mean of best (8) Models: 0.9512 CV and 0.9513 LB ; Stack of 40 models: 0.9529 CV and 0.9548 LB)", "971975": "that's the crazy CV score without no ensemble\nMine CV:0.9390 LB : 0.9488(Single Model)\n", "973166": "CV with TTA = 0.921\nLB = 0.9464\nadd meta + 1 model \nLB = 0.9592\n\nThat was my last submission, GL to everyone", "972295": "cv 0.931, lb 0.9417, it seems like many of my single model get a relatively good lb score (maximum 0.9488), but cv not.\nAnd when I try to do some ensemble based on cv, lb not increase much or even decrease.\n", "972083": "", "971946": "" } }