\nSingle fold CV-LB has a gap around 0.1 and all 5-fold submission has a gap around 0.04
\nFolds stratified on
species\n
![\"\"](\"https://i.ibb.co/30vR2hT/Screenshot-2022-03-01-at-1-46-13-PM.png\")
Framework : Tensorflow
\nModel / Size : Efficientnet5, 512 Size
\nSplit : Single fold
\nCV : 0.639
\nLB : 0.618
Model / Size : 768 Size
\nSplit : Single fold
\nCV : 0.673
\nLB : 0.651
Single Fold, Tensorflow
\nMODEL / SIZE : EfficientNet5, 768 size
\nCV : 0.744
\nLB : 0.724
MODEL / SIZE : EfficientNet5, 768 size
\nSingle Fold, Tensorflow
\nCV : 0.756
\nLB : 0.732
MODEL / SIZE : EfficientNet5, 768 size
\nSingle Fold, Tensorflow
\nCV : 0.827
\nLB : 0.795
\nI will keep uploading my scores here
Framework : Tensorflow
\nModel / Size : Efficientnet5, 512 Size
\nSplit : Single fold
\nCV : 0.639
\nLB : 0.618
Model / Size : 768 Size
\nSplit : Single fold
\nCV : 0.673
\nLB : 0.651
Single Fold, Tensorflow
\nMODEL / SIZE : EfficientNet5, 768 size
\nCV : 0.744
\nLB : 0.724
MODEL / SIZE : EfficientNet5, 768 size
\nSingle Fold, Tensorflow
\nCV : 0.756
\nLB : 0.732
MODEL / SIZE : EfficientNet5, 768 size
\nSingle Fold, Tensorflow
\nCV : 0.827
\nLB : 0.795
\nI will keep uploading my scores here
deja vu dear @deepkim but nice to see you here and thanks for creating this topic. I will follow this one.
", "rawMarkdown": "**deja vu** dear @deepkim but nice to see you here and thanks for creating this topic. I will follow this one.", "votes": null }, { "id": "1697152", "postDate": "02/19/2022 12:07:36", "content": "@deepkim I had pointed out in your previous post that I had started a similar thread.
\nYou've deleted the older post and restarted a new thread when I suggested a similar thread exists 🤔
", "rawMarkdown": "deepkim I had pointed out in your previous post that I had started a similar thread.\n\nYou've deleted the older post and restarted a new thread when I suggested a similar thread [exists](https://www.kaggle.com/c/happy-whale-and-dolphin/discussion/308588#1697063) 🤔", "votes": null }, { "id": "1697162", "postDate": "02/19/2022 12:16:00", "content": "I checked your thread but I think your post is about cv strategy
\nso I uploaded this thread again
I had created it with the intention to discuss both, from the thread:
\n\n", "rawMarkdown": "I had created it with the intention to discuss both, from the thread:\n\n> Also, what are your CV/LB scores, if you're okay to share them :)", "votes": null }, { "id": "1697169", "postDate": "02/19/2022 12:21:18", "content": "Also, what are your CV/LB scores, if you're okay to share them :)
\n
then I am going to upload my cv/lb here
\nwhat's the problem?
\nI just want to share my result. the name of your post is about cv strategy. so i didn't look it in detail.
\nI checked your thread again and i saw \"Also, what are your CV/LB scores, if you're okay to share them\"
\nso i decide to just upload my result
Nice to see you! and welcome again!
\nIt seems like someone posted already CV vs LB, so I will upload my score.
Ok. But these topics are different for me. Yours is LB/local score check. Second one is about Cv strategy. For me two different topics which can exist simultaneously
", "rawMarkdown": "Ok. But these topics are different for me. Yours is LB/local score check. Second one is about Cv strategy. For me two different topics which can exist simultaneously", "votes": null }, { "id": "1697188", "postDate": "02/19/2022 12:42:08", "content": "\n\nso i didn't look it in detail.
\n
Thanks for clarifying.
\nNo problems-I got confused by why you'd delete and repost after I pointed out :)
\nGood Luck :)
", "rawMarkdown": "> so i didn't look it in detail.\n\nThanks for clarifying. \n\nNo problems-I got confused by why you'd delete and repost after I pointed out :)\n\nGood Luck :)", "votes": null }, { "id": "1697228", "postDate": "02/19/2022 13:15:06", "content": "Hello, COTS buddies✋ I've learned a lot from you guys👍
", "rawMarkdown": "Hello, COTS buddies✋ I've learned a lot from you guys👍", "votes": null }, { "id": "1697231", "postDate": "02/19/2022 13:16:19", "content": "Nice to see you again!
", "rawMarkdown": "Nice to see you again!", "votes": null }, { "id": "1697721", "postDate": "02/19/2022 19:51:16", "content": "@deepkim @remekkinas how are you splitting the data? and are you calculating CV after the model gets trained? I cant get the validation work while the model trains
", "rawMarkdown": "deepkim @remekkinas how are you splitting the data? and are you calculating CV after the model gets trained? I cant get the validation work while the model trains", "votes": null }, { "id": "1697808", "postDate": "02/19/2022 21:11:32", "content": "We will let you know soon. We just started day ago … and today managed to submit first time. We need some time to come to first serious conclusions.
", "rawMarkdown": "We will let you know soon. We just started day ago … and today managed to submit first time. We need some time to come to first serious conclusions.", "votes": null }, { "id": "1697922", "postDate": "02/20/2022 00:55:31", "content": "F: pytorch
\nS: 5 fold
CV: 0.76
\nLB: 0.719
Single model, 0 fold.
", "rawMarkdown": "F: pytorch\nS: 5 fold \n\nCV: 0.76\nLB: 0.719\n\nSingle model, 0 fold.", "votes": null }, { "id": "1698008", "postDate": "02/20/2022 04:13:54", "content": "Framework : Tensorflow\nModel / Size : Efficientnet5, 512 Size\nSplit : Single holdout\nCV : 0.639\nLB : 0.618\nHi, is this a single model or k-folds ensemble？
", "rawMarkdown": "```\nFramework : Tensorflow\nModel / Size : Efficientnet5, 512 Size\nSplit : Single holdout\nCV : 0.639\nLB : 0.618\n```\nHi, is this a single model or k-folds ensemble？", "votes": null }, { "id": "1698063", "postDate": "02/20/2022 05:43:43", "content": "@mrinath it's just random selection. still not logical.
", "rawMarkdown": "mrinath it's just random selection. still not logical.", "votes": null }, { "id": "1698097", "postDate": "02/20/2022 06:18:43", "content": "single model. not k-folds
", "rawMarkdown": "single model. not k-folds", "votes": null }, { "id": "1698098", "postDate": "02/20/2022 06:20:16", "content": "so high! good luck!
", "rawMarkdown": "so high! good luck!", "votes": null }, { "id": "1698099", "postDate": "02/20/2022 06:20:34", "content": "That' cool, Thanks for reply !
", "rawMarkdown": "That' cool, Thanks for reply !", "votes": null }, { "id": "1698828", "postDate": "02/20/2022 17:39:48", "content": "Seems like Cots group moved here :D
\nF: Torch (for now I am trying to design the training and inference notebooks to equalize TF public kernels)
\nS: 5 folds
\nCV: 402
\nLB: 0.378
Pytorch
\nEfficientNetB6, 768x768, single fold
\nCV: 0.775, LB: 0.721
I don't know,my eff_nets are training very slowly even with image size = 384, do you mind saying how much time each epoch is taking?
", "rawMarkdown": "I don't know,my eff_nets are training very slowly even with image size = 384, do you mind saying how much time each epoch is taking?", "votes": null }, { "id": "1699340", "postDate": "02/21/2022 05:49:59", "content": "By using DDP + amp autocast, I can fit a batch size of 16 on 2x 3090. 1 epoch took 23 minutes
", "rawMarkdown": "By using DDP + amp autocast, I can fit a batch size of 16 on 2x 3090. 1 epoch took 23 minutes", "votes": null }, { "id": "1699508", "postDate": "02/21/2022 08:21:45", "content": "May I ask , what is your batch size ？ what is your GPU/TPU ram?
", "rawMarkdown": "May I ask , what is your batch size ？ what is your GPU/TPU ram?", "votes": null }, { "id": "1699604", "postDate": "02/21/2022 10:00:34", "content": "Hi，if it is convinient, can you tell what is the epoch number you set to reach this score?
", "rawMarkdown": "Hi，if it is convinient, can you tell what is the epoch number you set to reach this score?", "votes": null }, { "id": "1699874", "postDate": "02/21/2022 13:51:10", "content": "@rainfalllove I trained for 20 epochs
", "rawMarkdown": "rainfalllove I trained for 20 epochs", "votes": null }, { "id": "1700557", "postDate": "02/22/2022 05:09:57", "content": "per gpu bs is 8 , 2x3090
", "rawMarkdown": "per gpu bs is 8 , 2x3090", "votes": null }, { "id": "1700743", "postDate": "02/22/2022 07:52:07", "content": "thx for your reply.
", "rawMarkdown": "thx for your reply.", "votes": null }, { "id": "1701053", "postDate": "02/22/2022 13:48:15", "content": "did 16 batch use all the GPU memory? @andy2709
", "rawMarkdown": "did 16 batch use all the GPU memory? @andy2709", "votes": null }, { "id": "1701829", "postDate": "02/23/2022 05:04:42", "content": "@vladvdv good luck!
", "rawMarkdown": "vladvdv good luck!", "votes": null }, { "id": "1701878", "postDate": "02/23/2022 06:16:08", "content": "@andreaschandra yes, that was the max bsize I could set for b6 size 768
", "rawMarkdown": "andreaschandra yes, that was the max bsize I could set for b6 size 768", "votes": null }, { "id": "1705426", "postDate": "02/26/2022 13:31:39", "content": "I don't know why but I have huge gap between my CV and LB. My single fold validation is 0.713 and its submission scores 0.625.
", "rawMarkdown": "I don't know why but I have huge gap between my CV and LB. My single fold validation is 0.713 and its submission scores 0.625.", "votes": null }, { "id": "1705431", "postDate": "02/26/2022 13:39:57", "content": "I think Maybe it depends on the validation strategy or a threshold that decides a new individual
", "rawMarkdown": "I think Maybe it depends on the validation strategy or a threshold that decides a new individual", "votes": null }, { "id": "1705438", "postDate": "02/26/2022 13:44:43", "content": "@aerdem4 I have a similar gap too, I have tried stratified kfold on both individual_id and species but the gap is always around 0.08 to 0.09. I don't get how others have very less gap 👀
I have same kind of correlation - eg.
\nSingle model.
CV->LB
\n28.02.2022
\n0,76993509->0,706
\n0,78205229->0,716
\n0,785635449->0,722
\n0,784678947->0,725
\n0,78744856->0,723
01.03.2022
\n0,805614345->0,746
\n0,804673721->0,748
03.03.2022
\n0,811917826->0,761
Single fold:
\nCV : 0.749
\nLB : 0.710
also a huge gap.
", "rawMarkdown": "Single fold:\nCV : 0.749\nLB : 0.710\n\nalso a huge gap.", "votes": null }, { "id": "1705868", "postDate": "02/26/2022 22:44:43", "content": "As several people say, I also have some CV / LB gaps around 0.06 ~ 0.08
\nPytorch single fold:
\nCV: 0.750
\nLB: 0.678
CV: 0.740
\nLB: 0.667
@deepkim still you use TF/ArcFace?
", "rawMarkdown": "deepkim still you use TF/ArcFace?", "votes": null }, { "id": "1707169", "postDate": "02/28/2022 07:19:55", "content": "[Update]
\nSingle fold:
\nCV : 0.848
\nLB : 0.762
gap from 0.039 to 0.08.
", "rawMarkdown": "[Update]\nSingle fold:\nCV : 0.848\nLB : 0.762\n\ngap from 0.039 to 0.08.", "votes": null }, { "id": "1707313", "postDate": "02/28/2022 10:06:35", "content": "So high! I wonder if it's ok to reach such a high score just with kaggle platform?
", "rawMarkdown": "So high! I wonder if it's ok to reach such a high score just with kaggle platform?", "votes": null }, { "id": "1707340", "postDate": "02/28/2022 10:57:08", "content": "Yes，I just use kaggle kernel.
", "rawMarkdown": "Yes，I just use kaggle kernel.", "votes": null }, { "id": "1707343", "postDate": "02/28/2022 11:03:55", "content": "@remekkinas I use TF but ArcFace can be an option.
", "rawMarkdown": "remekkinas I use TF but ArcFace can be an option.", "votes": null }, { "id": "1707357", "postDate": "02/28/2022 11:33:32", "content": "Hi @librauee, I have varying gaps between CV-LB too, may I ask how are you splitting the data?
", "rawMarkdown": "Hi @librauee, I have varying gaps between CV-LB too, may I ask how are you splitting the data?", "votes": null }, { "id": "1707424", "postDate": "02/28/2022 13:06:31", "content": "Just using StratifiedKFold by species, I think there might be a better way to do this.
", "rawMarkdown": "Just using StratifiedKFold by species, I think there might be a better way to do this.", "votes": null }, { "id": "1707762", "postDate": "02/28/2022 18:40:17", "content": "Folds: 5
\nCV: 0.7875
\nLB: 0.741
PyTorch
\nEfficientNetB5, 512
\n5 folds
\nCV: 0.747
\nLB: 0.731
[UPDATE]
\nSingle fold CV-LB has a gap around 0.1 and all 5-fold submission has a gap around 0.04
\nFolds stratified on species
\n
Nice CV, good job
", "rawMarkdown": "Nice CV, good job", "votes": null }, { "id": "1708812", "postDate": "03/01/2022 18:09:49", "content": "@thanhns nice score! This is with ArcFace/GeM?
", "rawMarkdown": "thanhns nice score! This is with ArcFace/GeM?", "votes": null }, { "id": "1708854", "postDate": "03/01/2022 18:38:03", "content": "@andy2709 Did you use a dataset cropped by detic?
", "rawMarkdown": "andy2709 Did you use a dataset cropped by detic?", "votes": null }, { "id": "1709127", "postDate": "03/02/2022 01:53:50", "content": "Single fold or total fold?
", "rawMarkdown": "Single fold or total fold?", "votes": null }, { "id": "1709313", "postDate": "03/02/2022 05:13:58", "content": "Single Fold(4 fold)
\nTF
\nCV: 0.794
\nLB: 0.745
same fold&same model:
\npytroch
\nCV: 0.717
\nLB:0.636
\nwhy?😳
I have the same issue that drives me crazy.
", "rawMarkdown": "I have the same issue that drives me crazy.", "votes": null }, { "id": "1710373", "postDate": "03/03/2022 02:00:40", "content": "it reminds me of the google ventilator challenge. Anw, I am using Pytorch
", "rawMarkdown": "it reminds me of the google ventilator challenge. Anw, I am using Pytorch", "votes": null }, { "id": "1710700", "postDate": "03/03/2022 08:50:24", "content": "what is your valid loss?
", "rawMarkdown": "what is your valid loss?", "votes": null }, { "id": "1710701", "postDate": "03/03/2022 08:51:38", "content": "Framework : Pytorch
\nModel / Size : Efficientnet7, 768 Size
\nSplit : 5 folds
\nCV : 0.731
\nLB : 0.724
Awesome!!! So high cv with single fold.👍
", "rawMarkdown": "Awesome!!! So high cv with single fold.👍", "votes": null }, { "id": "1711041", "postDate": "03/03/2022 15:31:55", "content": "Wow, great CV-LB!
", "rawMarkdown": "Wow, great CV-LB!", "votes": null }, { "id": "1711670", "postDate": "03/04/2022 07:33:03", "content": "zhe ge li hai le.
", "rawMarkdown": "zhe ge li hai le.", "votes": null }, { "id": "1711685", "postDate": "03/04/2022 07:54:43", "content": "Pytorch is some how have very low score compared to TF :'(
", "rawMarkdown": "Pytorch is some how have very low score compared to TF :'(", "votes": null }, { "id": "1712624", "postDate": "03/05/2022 05:49:41", "content": "Single Fold, Tensorflow
\nCV : 0.848
\nLB : 0.749
\nso huge gap!
holdout, pytorch
\nCV: 0.705
\nLB: 0.712
CV: 0.741
\nLB: 0.741
CV: 0.756
\nLB: 0.749
Sorry for the late reply.
\n@remekkinas I use PyTorch/ArcFace without GeM.
\n@chihantsai Total of 5 folds.
", "rawMarkdown": "Sorry for the late reply.\n\n@remekkinas I use PyTorch/ArcFace without GeM.\n\n@chihantsai Total of 5 folds.", "votes": null }, { "id": "1713434", "postDate": "03/06/2022 02:53:37", "content": "5 Fold Split
\nPytorch
\nFold 0 Single Model
\nB5-512
\nCV: 0.679
\nLB: 0.706
No fold, Pytorch
\nEffnet-B0 - 380x380
\nLB: 0.454
I'm confusing between single fold and single model. anyone can describe me?
", "rawMarkdown": "I'm confusing between single fold and single model. anyone can describe me?", "votes": null }, { "id": "1719191", "postDate": "03/11/2022 14:58:06", "content": "Mostly used for same meaning, 1 model file.
", "rawMarkdown": "Mostly used for same meaning, 1 model file.", "votes": null }, { "id": "1719198", "postDate": "03/11/2022 15:02:54", "content": "In your meaning are if I train with each 5 fold split. and I use the model from one of 5 fold right?
", "rawMarkdown": "In your meaning are if I train with each 5 fold split. and I use the model from one of 5 fold right?", "votes": null }, { "id": "1719388", "postDate": "03/11/2022 17:45:53", "content": "Yes, that's right
", "rawMarkdown": "Yes, that's right", "votes": null }, { "id": "1723093", "postDate": "03/15/2022 05:51:39", "content": "Single Fold (TF)
\nCV - 0.769763
\nLB - 0.730 😂 (Am I doing sommething wrong ? 🥴)
I remember that I was able to reproduce TF scores using PyTorch (in google ventilator challenge), let's see if this is the case here
", "rawMarkdown": "I remember that I was able to reproduce TF scores using PyTorch (in google ventilator challenge), let's see if this is the case here", "votes": null }, { "id": "1727554", "postDate": "03/18/2022 03:52:41", "content": "Initially I used detic boxes but I switched to something else 😄
", "rawMarkdown": "Initially I used detic boxes but I switched to something else 😄", "votes": null }, { "id": "1727737", "postDate": "03/18/2022 08:17:11", "content": "@deepkim nice jump in score ….. from CV: 0.756 to 0.827 ….
", "rawMarkdown": "deepkim nice jump in score ..... from CV: 0.756 to 0.827 ....", "votes": null }, { "id": "1727764", "postDate": "03/18/2022 09:04:24", "content": "What are these guys I did not understand anything? Could someone explain what CV and LB scores are?
", "rawMarkdown": "What are these guys I did not understand anything? Could someone explain what CV and LB scores are?", "votes": null }, { "id": "1727768", "postDate": "03/18/2022 09:10:34", "content": "No problem i will try explain.
\nCV is your local Cross Validated score (validated on folds).
\nLB is just score you got after submitting the same prediction to Public Leaderboard (~25% out of test dataset).
\nand then we will have Private Leaderboard (validated on 100% of test data) - when competition finish.
", "rawMarkdown": "No problem i will try explain.\n\nCV is your local Cross Validated score (validated on folds).\n\nLB is just score you got after submitting the same prediction to Public Leaderboard (~25% out of test dataset).\n\nand then we will have Private Leaderboard (validated on 100% of test data) - when competition finish.", "votes": null }, { "id": "1727839", "postDate": "03/18/2022 11:23:58", "content": "I would add one correction what Remek mentioned: private LB is on the remaining 76%. Qouting from
\nthe LB:
\n", "rawMarkdown": "I would add one correction what Remek mentioned: private LB is on the remaining 76%. Qouting from \nthe LB:\n\n> The final results will be based on the other 76%, so the final standings may be different.", "votes": null }, { "id": "1727968", "postDate": "03/18/2022 13:36:31", "content": "The final results will be based on the other 76%, so the final standings may be different.
\n
Hi GM @init27 I saw your live youtube video when you reached GM - fantastic! Absolutely fantastic emotions and talk! You are such positive person 👍👍👍👍
", "rawMarkdown": "Hi GM @init27 I saw your live youtube video when you reached GM - fantastic! Absolutely fantastic emotions and talk! You are such positive person 👍👍👍👍", "votes": null }, { "id": "1727984", "postDate": "03/18/2022 13:55:46", "content": "Single Fold (TF)
\nModel: EfficientNetB5
\nImage Size: 512
\nCV - 0.827 (Much higher than lb? Using same CV as in public EffNet notebooks)
\nLB - 0.730
Maybe CV depends on your validation setting.
", "rawMarkdown": "Maybe CV depends on your validation setting.", "votes": null }, { "id": "1728213", "postDate": "03/18/2022 17:12:00", "content": "What do you mean by that? Do you mean the fold I'm using or something else?
", "rawMarkdown": "What do you mean by that? Do you mean the fold I'm using or something else?", "votes": null }, { "id": "1728219", "postDate": "03/18/2022 17:15:35", "content": "hi, can you tell how many epochs you use if you dont mind, i only got 0.7 lb score with single fold1, about 30epoch.
", "rawMarkdown": "hi, can you tell how many epochs you use if you dont mind, i only got 0.7 lb score with single fold1, about 30epoch.", "votes": null }, { "id": "1728223", "postDate": "03/18/2022 17:20:47", "content": "Only 24 epochs.
", "rawMarkdown": "Only 24 epochs.", "votes": null }, { "id": "1728227", "postDate": "03/18/2022 17:26:07", "content": "thats really amazing
", "rawMarkdown": "thats really amazing", "votes": null }, { "id": "1728294", "postDate": "03/18/2022 18:35:25", "content": "The CV is all about calculating, my cv assumes only 9% of individual ids while calculating, that is why I calculate my CVs to be barely .7 and able to reach .745 LB…, in a normal split there is about 20% new ids which are easy to guess, that is why your CV is 827 but the LB currently have about 12-13% new ids
", "rawMarkdown": "The CV is all about calculating, my cv assumes only 9% of individual ids while calculating, that is why I calculate my CVs to be barely .7 and able to reach .745 LB..., in a normal split there is about 20% new ids which are easy to guess, that is why your CV is 827 but the LB currently have about 12-13% new ids", "votes": null }, { "id": "1728462", "postDate": "03/18/2022 23:42:35", "content": "Oh ok. Thanks for the explanation.
", "rawMarkdown": "Oh ok. Thanks for the explanation.", "votes": null }, { "id": "1728558", "postDate": "03/19/2022 03:22:53", "content": "I will never get used to being called a GM 😂
\nThanks so much, Remek! I know I have a lot to learn, but it was still a very emotional moment for me.
\nBtw, I have been learning a lot from your tutorials in the whale competition already! 🙏
I think I had discovered your channel as well, which is mostly in Polish language? Do you plan on making English content in the future?
", "rawMarkdown": "I will never get used to being called a GM 😂\n\nThanks so much, Remek! I know I have a lot to learn, but it was still a very emotional moment for me. \nBtw, I have been learning a lot from your tutorials in the whale competition already! 🙏\n\nI think I had discovered your channel as well, which is mostly in Polish language? Do you plan on making English content in the future?", "votes": null }, { "id": "1729487", "postDate": "03/20/2022 06:16:06", "content": "Using b5 can reach such a high score !!
", "rawMarkdown": "Using b5 can reach such a high score !!", "votes": null }, { "id": "1729554", "postDate": "03/20/2022 07:50:39", "content": "Thank you! I am looking for something :) which gave me even tiny jump :) I feel that my ideas … ran out … :)
", "rawMarkdown": "Thank you! I am looking for something :) which gave me even tiny jump :) I feel that my ideas … ran out … :)", "votes": null }, { "id": "1729558", "postDate": "03/20/2022 07:54:02", "content": "I feel the same - I still feel like Junior Data Scientists :)
\nYes, my channel is in Polish. I am not able to explain hard topic in English unfortunately (or this is my excuse:)). I have to make videos regularly- now I have problem with this.
\nEmotions are good. You are nice and positive person. I like your channel.
", "rawMarkdown": "I feel the same - I still feel like Junior Data Scientists :) \n\nYes, my channel is in Polish. I am not able to explain hard topic in English unfortunately (or this is my excuse:)). I have to make videos regularly- now I have problem with this. \n\nEmotions are good. You are nice and positive person. I like your channel.", "votes": null }, { "id": "1730545", "postDate": "03/21/2022 12:03:29", "content": "PyTorch
\n5 fold split
\nFold 0 single model
\nImage size 384
EfficientNet B7
\nCV: 0.841
\nLB: 0.749
ConvNeXt Base
\nCV: 0.825
\nLB: 0.741
Really considering teaming up with you…, especially cuz most of your code is public right now, I can see a lot of places to improve, and you look like someone who can execute tasks when guided (We are definitely going to get a huge boost if we team up). Contact me with your discord id (you can send me a mail), I will add you for further communication.
", "rawMarkdown": "Really considering teaming up with you..., especially cuz most of your code is public right now, I can see a lot of places to improve, and you look like someone who can execute tasks when guided (We are definitely going to get a huge boost if we team up). Contact me with your discord id (you can send me a mail), I will add you for further communication.", "votes": null }, { "id": "1730646", "postDate": "03/21/2022 13:51:04", "content": "@remekkinas for me, Scoring a high cv is quite easy. but LB is really tricky.
", "rawMarkdown": "remekkinas for me, Scoring a high cv is quite easy. but LB is really tricky.", "votes": null }, { "id": "1730661", "postDate": "03/21/2022 14:02:36", "content": "Sent you a mail
", "rawMarkdown": "Sent you a mail", "votes": null }, { "id": "1730665", "postDate": "03/21/2022 14:05:55", "content": "Wouldn't you want to trust your CV at the end of the day anyways? Ofcourse, it will be bit different if you are using pseudo labeling in your pipeline. As I am increasing the score, I am experiencing the problem of LB vs CV shift, but I believe if we keep improving CV, the shift will catch up in one model or another and when we ensemble multiple models, we will be able to see that score.
", "rawMarkdown": "Wouldn't you want to trust your CV at the end of the day anyways? Ofcourse, it will be bit different if you are using pseudo labeling in your pipeline. As I am increasing the score, I am experiencing the problem of LB vs CV shift, but I believe if we keep improving CV, the shift will catch up in one model or another and when we ensemble multiple models, we will be able to see that score.", "votes": null }, { "id": "1730730", "postDate": "03/21/2022 15:07:23", "content": "the really good score with convnext
", "rawMarkdown": "the really good score with convnext", "votes": null }, { "id": "1731239", "postDate": "03/22/2022 06:22:36", "content": "Hi @harshitsheoran , can you tell the gap between your CV-LB when considering only 9% individual ids ?
", "rawMarkdown": "Hi @harshitsheoran , can you tell the gap between your CV-LB when considering only 9% individual ids ?", "votes": null }, { "id": "1731316", "postDate": "03/22/2022 08:29:08", "content": "-40 points, that is if the CV is .705, I get a LB of .745
", "rawMarkdown": "40 points, that is if the CV is .705, I get a LB of .745", "votes": null }, { "id": "1731696", "postDate": "03/22/2022 15:51:42", "content": "Pytorch
\n5 fold split
\nFold 0 single model
EffNet-B7 (img_size=768)
\nCV: 816
\nLB: 717
ConvNeXt-large (img_size=384)
\nCV: 813
\nLB: 715
Tensorflow
\n5 fold split
\nFold 4
\nEfficientNet-B5
\nImage Size: 768
\nVal Loss: 7.6865
\nCV : 0.8174
\nLB : 0.789
Which dataset do you use? Custom or published here?
", "rawMarkdown": "Which dataset do you use? Custom or published here?", "votes": null }, { "id": "1733442", "postDate": "03/24/2022 09:29:26", "content": "Custom Dataset
", "rawMarkdown": "Custom Dataset", "votes": null }, { "id": "1733499", "postDate": "03/24/2022 10:41:38", "content": "Do you train on Kaggle devices?
", "rawMarkdown": "Do you train on Kaggle devices?", "votes": null }, { "id": "1733876", "postDate": "03/24/2022 17:30:43", "content": "No, we use local machines, those models were trained using 1x 3090
", "rawMarkdown": "No, we use local machines, those models were trained using 1x 3090", "votes": null }, { "id": "1736587", "postDate": "03/27/2022 13:43:03", "content": "Pytorch
\n5-Fold Model
\nModel / Size: EfficientNet-B5, 384 size
CV: 0.755 (I calculate CV with only 9% individual ids in my valid set but I train (like everyone does) with about 20% individual ids in valid_set)
\nOne-Fold-LB: 0.775!
\nAll-Fold-LB: 0.798!
Looking forward to scores with bigger image size and better models!
\nAlso, would like TF-TPU pipeline team to join us…
deja vu dear @deepkim but nice to see you here and thanks for creating this topic. I will follow this one.
", "votes": null, "replies": [ { "id": 1697171, "author_name": "deepkim", "author_url": "", "post_date": "02/19/2022 12:26:17", "content": "Nice to see you! and welcome again!
\nIt seems like someone posted already CV vs LB, so I will upload my score.
Ok. But these topics are different for me. Yours is LB/local score check. Second one is about Cv strategy. For me two different topics which can exist simultaneously
", "votes": null, "replies": [] }, { "id": 1697228, "author_name": "cys1102", "author_url": "", "post_date": "02/19/2022 13:15:06", "content": "Hello, COTS buddies✋ I've learned a lot from you guys👍
", "votes": null, "replies": [] }, { "id": 1697231, "author_name": "deepkim", "author_url": "", "post_date": "02/19/2022 13:16:19", "content": "Nice to see you again!
", "votes": null, "replies": [] }, { "id": 1697721, "author_name": "mrinath", "author_url": "", "post_date": "02/19/2022 19:51:16", "content": "@deepkim @remekkinas how are you splitting the data? and are you calculating CV after the model gets trained? I cant get the validation work while the model trains
", "votes": null, "replies": [] }, { "id": 1697808, "author_name": "remekkinas", "author_url": "", "post_date": "02/19/2022 21:11:32", "content": "We will let you know soon. We just started day ago … and today managed to submit first time. We need some time to come to first serious conclusions.
", "votes": null, "replies": [] }, { "id": 1698063, "author_name": "deepkim", "author_url": "", "post_date": "02/20/2022 05:43:43", "content": "@mrinath it's just random selection. still not logical.
", "votes": null, "replies": [] }, { "id": 1698828, "author_name": "vladvdv", "author_url": "", "post_date": "02/20/2022 17:39:48", "content": "Seems like Cots group moved here :D
\nF: Torch (for now I am trying to design the training and inference notebooks to equalize TF public kernels)
\nS: 5 folds
\nCV: 402
\nLB: 0.378
@vladvdv good luck!
", "votes": null, "replies": [] } ] }, { "id": 1697152, "author_name": "init27", "author_url": "", "post_date": "02/19/2022 12:07:36", "content": "@deepkim I had pointed out in your previous post that I had started a similar thread.
\nYou've deleted the older post and restarted a new thread when I suggested a similar thread exists 🤔
", "votes": null, "replies": [ { "id": 1697162, "author_name": "deepkim", "author_url": "", "post_date": "02/19/2022 12:16:00", "content": "I checked your thread but I think your post is about cv strategy
\nso I uploaded this thread again
I had created it with the intention to discuss both, from the thread:
\n\n", "votes": null, "replies": [] }, { "id": 1697169, "author_name": "deepkim", "author_url": "", "post_date": "02/19/2022 12:21:18", "content": "Also, what are your CV/LB scores, if you're okay to share them :)
\n
then I am going to upload my cv/lb here
\nwhat's the problem?
\nI just want to share my result. the name of your post is about cv strategy. so i didn't look it in detail.
\nI checked your thread again and i saw \"Also, what are your CV/LB scores, if you're okay to share them\"
\nso i decide to just upload my result
\n\nso i didn't look it in detail.
\n
Thanks for clarifying.
\nNo problems-I got confused by why you'd delete and repost after I pointed out :)
\nGood Luck :)
", "votes": null, "replies": [] } ] }, { "id": 1697922, "author_name": "biglafe", "author_url": "", "post_date": "02/20/2022 00:55:31", "content": "F: pytorch
\nS: 5 fold
CV: 0.76
\nLB: 0.719
Single model, 0 fold.
", "votes": null, "replies": [ { "id": 1698098, "author_name": "deepkim", "author_url": "", "post_date": "02/20/2022 06:20:16", "content": "so high! good luck!
", "votes": null, "replies": [] }, { "id": 1699508, "author_name": "dragonzhang", "author_url": "", "post_date": "02/21/2022 08:21:45", "content": "May I ask , what is your batch size ？ what is your GPU/TPU ram?
", "votes": null, "replies": [] }, { "id": 1700557, "author_name": "biglafe", "author_url": "", "post_date": "02/22/2022 05:09:57", "content": "per gpu bs is 8 , 2x3090
", "votes": null, "replies": [] }, { "id": 1700743, "author_name": "dragonzhang", "author_url": "", "post_date": "02/22/2022 07:52:07", "content": "thx for your reply.
", "votes": null, "replies": [] } ] }, { "id": 1698008, "author_name": "rainfalllove", "author_url": "", "post_date": "02/20/2022 04:13:54", "content": "Framework : Tensorflow\nModel / Size : Efficientnet5, 512 Size\nSplit : Single holdout\nCV : 0.639\nLB : 0.618\nHi, is this a single model or k-folds ensemble？
", "votes": null, "replies": [ { "id": 1698097, "author_name": "deepkim", "author_url": "", "post_date": "02/20/2022 06:18:43", "content": "single model. not k-folds
", "votes": null, "replies": [] }, { "id": 1698099, "author_name": "rainfalllove", "author_url": "", "post_date": "02/20/2022 06:20:34", "content": "That' cool, Thanks for reply !
", "votes": null, "replies": [] } ] }, { "id": 1699139, "author_name": "andy2709", "author_url": "", "post_date": "02/21/2022 00:21:55", "content": "Pytorch
\nEfficientNetB6, 768x768, single fold
\nCV: 0.775, LB: 0.721
I don't know,my eff_nets are training very slowly even with image size = 384, do you mind saying how much time each epoch is taking?
", "votes": null, "replies": [] }, { "id": 1699340, "author_name": "andy2709", "author_url": "", "post_date": "02/21/2022 05:49:59", "content": "By using DDP + amp autocast, I can fit a batch size of 16 on 2x 3090. 1 epoch took 23 minutes
", "votes": null, "replies": [] }, { "id": 1699604, "author_name": "rainfalllove", "author_url": "", "post_date": "02/21/2022 10:00:34", "content": "Hi，if it is convinient, can you tell what is the epoch number you set to reach this score?
", "votes": null, "replies": [] }, { "id": 1699874, "author_name": "andy2709", "author_url": "", "post_date": "02/21/2022 13:51:10", "content": "@rainfalllove I trained for 20 epochs
", "votes": null, "replies": [] }, { "id": 1701053, "author_name": "andreaschandra", "author_url": "", "post_date": "02/22/2022 13:48:15", "content": "did 16 batch use all the GPU memory? @andy2709
", "votes": null, "replies": [] }, { "id": 1701878, "author_name": "andy2709", "author_url": "", "post_date": "02/23/2022 06:16:08", "content": "@andreaschandra yes, that was the max bsize I could set for b6 size 768
", "votes": null, "replies": [] }, { "id": 1708854, "author_name": "gwanghan", "author_url": "", "post_date": "03/01/2022 18:38:03", "content": "@andy2709 Did you use a dataset cropped by detic?
", "votes": null, "replies": [] }, { "id": 1727554, "author_name": "andy2709", "author_url": "", "post_date": "03/18/2022 03:52:41", "content": "Initially I used detic boxes but I switched to something else 😄
", "votes": null, "replies": [] } ] }, { "id": 1705426, "author_name": "aerdem4", "author_url": "", "post_date": "02/26/2022 13:31:39", "content": "I don't know why but I have huge gap between my CV and LB. My single fold validation is 0.713 and its submission scores 0.625.
", "votes": null, "replies": [ { "id": 1705431, "author_name": "deepkim", "author_url": "", "post_date": "02/26/2022 13:39:57", "content": "I think Maybe it depends on the validation strategy or a threshold that decides a new individual
", "votes": null, "replies": [] }, { "id": 1705438, "author_name": "atharvaingle", "author_url": "", "post_date": "02/26/2022 13:44:43", "content": "@aerdem4 I have a similar gap too, I have tried stratified kfold on both individual_id and species but the gap is always around 0.08 to 0.09. I don't get how others have very less gap 👀
I have same kind of correlation - eg.
\nSingle model.
CV->LB
\n28.02.2022
\n0,76993509->0,706
\n0,78205229->0,716
\n0,785635449->0,722
\n0,784678947->0,725
\n0,78744856->0,723
01.03.2022
\n0,805614345->0,746
\n0,804673721->0,748
03.03.2022
\n0,811917826->0,761
[UPDATE]
\nSingle fold CV-LB has a gap around 0.1 and all 5-fold submission has a gap around 0.04
\nFolds stratified on species
\n
Single fold:
\nCV : 0.749
\nLB : 0.710
also a huge gap.
", "votes": null, "replies": [ { "id": 1707169, "author_name": "librauee", "author_url": "", "post_date": "02/28/2022 07:19:55", "content": "[Update]
\nSingle fold:
\nCV : 0.848
\nLB : 0.762
gap from 0.039 to 0.08.
", "votes": null, "replies": [] }, { "id": 1707313, "author_name": "rainfalllove", "author_url": "", "post_date": "02/28/2022 10:06:35", "content": "So high! I wonder if it's ok to reach such a high score just with kaggle platform?
", "votes": null, "replies": [] }, { "id": 1707340, "author_name": "librauee", "author_url": "", "post_date": "02/28/2022 10:57:08", "content": "Yes，I just use kaggle kernel.
", "votes": null, "replies": [] }, { "id": 1707357, "author_name": "atharvaingle", "author_url": "", "post_date": "02/28/2022 11:33:32", "content": "Hi @librauee, I have varying gaps between CV-LB too, may I ask how are you splitting the data?
", "votes": null, "replies": [] }, { "id": 1707424, "author_name": "librauee", "author_url": "", "post_date": "02/28/2022 13:06:31", "content": "Just using StratifiedKFold by species, I think there might be a better way to do this.
", "votes": null, "replies": [] }, { "id": 1708425, "author_name": "ragnar123", "author_url": "", "post_date": "03/01/2022 12:51:06", "content": "Nice CV, good job
", "votes": null, "replies": [] }, { "id": 1710707, "author_name": "yingpengchen", "author_url": "", "post_date": "03/03/2022 08:57:59", "content": "Awesome!!! So high cv with single fold.👍
", "votes": null, "replies": [] } ] }, { "id": 1705868, "author_name": "yoichi7yamakawa", "author_url": "", "post_date": "02/26/2022 22:44:43", "content": "As several people say, I also have some CV / LB gaps around 0.06 ~ 0.08
\nPytorch single fold:
\nCV: 0.750
\nLB: 0.678
CV: 0.740
\nLB: 0.667
@deepkim still you use TF/ArcFace?
", "votes": null, "replies": [ { "id": 1707343, "author_name": "deepkim", "author_url": "", "post_date": "02/28/2022 11:03:55", "content": "@remekkinas I use TF but ArcFace can be an option.
", "votes": null, "replies": [] } ] }, { "id": 1707762, "author_name": "ragnar123", "author_url": "", "post_date": "02/28/2022 18:40:17", "content": "Folds: 5
\nCV: 0.7875
\nLB: 0.741
PyTorch
\nEfficientNetB5, 512
\n5 folds
\nCV: 0.747
\nLB: 0.731
@thanhns nice score! This is with ArcFace/GeM?
", "votes": null, "replies": [] }, { "id": 1709127, "author_name": "chihantsai", "author_url": "", "post_date": "03/02/2022 01:53:50", "content": "Single fold or total fold?
", "votes": null, "replies": [] }, { "id": 1713004, "author_name": "thanhns", "author_url": "", "post_date": "03/05/2022 15:04:05", "content": "Sorry for the late reply.
\n@remekkinas I use PyTorch/ArcFace without GeM.
\n@chihantsai Total of 5 folds.
", "votes": null, "replies": [] } ] }, { "id": 1709313, "author_name": "kevin1742064161", "author_url": "", "post_date": "03/02/2022 05:13:58", "content": "Single Fold(4 fold)
\nTF
\nCV: 0.794
\nLB: 0.745
same fold&same model:
\npytroch
\nCV: 0.717
\nLB:0.636
\nwhy?😳
I have the same issue that drives me crazy.
", "votes": null, "replies": [] }, { "id": 1710373, "author_name": "namgalielei", "author_url": "", "post_date": "03/03/2022 02:00:40", "content": "it reminds me of the google ventilator challenge. Anw, I am using Pytorch
", "votes": null, "replies": [] }, { "id": 1710700, "author_name": "chihantsai", "author_url": "", "post_date": "03/03/2022 08:50:24", "content": "what is your valid loss?
", "votes": null, "replies": [] }, { "id": 1711685, "author_name": "ptran1203", "author_url": "", "post_date": "03/04/2022 07:54:43", "content": "Pytorch is some how have very low score compared to TF :'(
", "votes": null, "replies": [] }, { "id": 1727552, "author_name": "martynoveduard", "author_url": "", "post_date": "03/18/2022 03:50:42", "content": "I remember that I was able to reproduce TF scores using PyTorch (in google ventilator challenge), let's see if this is the case here
", "votes": null, "replies": [] } ] }, { "id": 1710701, "author_name": "yingpengchen", "author_url": "", "post_date": "03/03/2022 08:51:38", "content": "Framework : Pytorch
\nModel / Size : Efficientnet7, 768 Size
\nSplit : 5 folds
\nCV : 0.731
\nLB : 0.724
Wow, great CV-LB!
", "votes": null, "replies": [] }, { "id": 1711670, "author_name": "tiandaye", "author_url": "", "post_date": "03/04/2022 07:33:03", "content": "zhe ge li hai le.
", "votes": null, "replies": [] } ] }, { "id": 1712624, "author_name": "mrhagchwh", "author_url": "", "post_date": "03/05/2022 05:49:41", "content": "Single Fold, Tensorflow
\nCV : 0.848
\nLB : 0.749
\nso huge gap!
holdout, pytorch
\nCV: 0.705
\nLB: 0.712
CV: 0.741
\nLB: 0.741
CV: 0.756
\nLB: 0.749
5 Fold Split
\nPytorch
\nFold 0 Single Model
\nB5-512
\nCV: 0.679
\nLB: 0.706
No fold, Pytorch
\nEffnet-B0 - 380x380
\nLB: 0.454
I'm confusing between single fold and single model. anyone can describe me?
", "votes": null, "replies": [ { "id": 1719191, "author_name": "harshitsheoran", "author_url": "", "post_date": "03/11/2022 14:58:06", "content": "Mostly used for same meaning, 1 model file.
", "votes": null, "replies": [] }, { "id": 1719198, "author_name": "premsurawut", "author_url": "", "post_date": "03/11/2022 15:02:54", "content": "In your meaning are if I train with each 5 fold split. and I use the model from one of 5 fold right?
", "votes": null, "replies": [] }, { "id": 1719388, "author_name": "harshitsheoran", "author_url": "", "post_date": "03/11/2022 17:45:53", "content": "Yes, that's right
", "votes": null, "replies": [] } ] }, { "id": 1723093, "author_name": "benihime91", "author_url": "", "post_date": "03/15/2022 05:51:39", "content": "Single Fold (TF)
\nCV - 0.769763
\nLB - 0.730 😂 (Am I doing sommething wrong ? 🥴)
@deepkim nice jump in score ….. from CV: 0.756 to 0.827 ….
", "votes": null, "replies": [ { "id": 1729487, "author_name": "biglafe", "author_url": "", "post_date": "03/20/2022 06:16:06", "content": "Using b5 can reach such a high score !!
", "votes": null, "replies": [] }, { "id": 1729554, "author_name": "remekkinas", "author_url": "", "post_date": "03/20/2022 07:50:39", "content": "Thank you! I am looking for something :) which gave me even tiny jump :) I feel that my ideas … ran out … :)
", "votes": null, "replies": [] }, { "id": 1730646, "author_name": "deepkim", "author_url": "", "post_date": "03/21/2022 13:51:04", "content": "@remekkinas for me, Scoring a high cv is quite easy. but LB is really tricky.
", "votes": null, "replies": [] }, { "id": 1730665, "author_name": "harshitsheoran", "author_url": "", "post_date": "03/21/2022 14:05:55", "content": "Wouldn't you want to trust your CV at the end of the day anyways? Ofcourse, it will be bit different if you are using pseudo labeling in your pipeline. As I am increasing the score, I am experiencing the problem of LB vs CV shift, but I believe if we keep improving CV, the shift will catch up in one model or another and when we ensemble multiple models, we will be able to see that score.
", "votes": null, "replies": [] } ] }, { "id": 1727764, "author_name": "nelsonjoseph", "author_url": "", "post_date": "03/18/2022 09:04:24", "content": "What are these guys I did not understand anything? Could someone explain what CV and LB scores are?
", "votes": null, "replies": [ { "id": 1727768, "author_name": "remekkinas", "author_url": "", "post_date": "03/18/2022 09:10:34", "content": "No problem i will try explain.
\nCV is your local Cross Validated score (validated on folds).
\nLB is just score you got after submitting the same prediction to Public Leaderboard (~25% out of test dataset).
\nand then we will have Private Leaderboard (validated on 100% of test data) - when competition finish.
", "votes": null, "replies": [] }, { "id": 1727839, "author_name": "init27", "author_url": "", "post_date": "03/18/2022 11:23:58", "content": "I would add one correction what Remek mentioned: private LB is on the remaining 76%. Qouting from
\nthe LB:
\n", "votes": null, "replies": [] }, { "id": 1727968, "author_name": "remekkinas", "author_url": "", "post_date": "03/18/2022 13:36:31", "content": "The final results will be based on the other 76%, so the final standings may be different.
\n
Hi GM @init27 I saw your live youtube video when you reached GM - fantastic! Absolutely fantastic emotions and talk! You are such positive person 👍👍👍👍
", "votes": null, "replies": [] }, { "id": 1728558, "author_name": "init27", "author_url": "", "post_date": "03/19/2022 03:22:53", "content": "I will never get used to being called a GM 😂
\nThanks so much, Remek! I know I have a lot to learn, but it was still a very emotional moment for me.
\nBtw, I have been learning a lot from your tutorials in the whale competition already! 🙏
I think I had discovered your channel as well, which is mostly in Polish language? Do you plan on making English content in the future?
", "votes": null, "replies": [] }, { "id": 1729558, "author_name": "remekkinas", "author_url": "", "post_date": "03/20/2022 07:54:02", "content": "I feel the same - I still feel like Junior Data Scientists :)
\nYes, my channel is in Polish. I am not able to explain hard topic in English unfortunately (or this is my excuse:)). I have to make videos regularly- now I have problem with this.
\nEmotions are good. You are nice and positive person. I like your channel.
", "votes": null, "replies": [] } ] }, { "id": 1727984, "author_name": "vexxingbanana", "author_url": "", "post_date": "03/18/2022 13:55:46", "content": "Single Fold (TF)
\nModel: EfficientNetB5
\nImage Size: 512
\nCV - 0.827 (Much higher than lb? Using same CV as in public EffNet notebooks)
\nLB - 0.730
Maybe CV depends on your validation setting.
", "votes": null, "replies": [] }, { "id": 1728213, "author_name": "vexxingbanana", "author_url": "", "post_date": "03/18/2022 17:12:00", "content": "What do you mean by that? Do you mean the fold I'm using or something else?
", "votes": null, "replies": [] }, { "id": 1728219, "author_name": "rainfalllove", "author_url": "", "post_date": "03/18/2022 17:15:35", "content": "hi, can you tell how many epochs you use if you dont mind, i only got 0.7 lb score with single fold1, about 30epoch.
", "votes": null, "replies": [] }, { "id": 1728223, "author_name": "vexxingbanana", "author_url": "", "post_date": "03/18/2022 17:20:47", "content": "Only 24 epochs.
", "votes": null, "replies": [] }, { "id": 1728227, "author_name": "rainfalllove", "author_url": "", "post_date": "03/18/2022 17:26:07", "content": "thats really amazing
", "votes": null, "replies": [] }, { "id": 1728294, "author_name": "harshitsheoran", "author_url": "", "post_date": "03/18/2022 18:35:25", "content": "The CV is all about calculating, my cv assumes only 9% of individual ids while calculating, that is why I calculate my CVs to be barely .7 and able to reach .745 LB…, in a normal split there is about 20% new ids which are easy to guess, that is why your CV is 827 but the LB currently have about 12-13% new ids
", "votes": null, "replies": [] }, { "id": 1728462, "author_name": "vexxingbanana", "author_url": "", "post_date": "03/18/2022 23:42:35", "content": "Oh ok. Thanks for the explanation.
", "votes": null, "replies": [] }, { "id": 1731239, "author_name": "atharvaingle", "author_url": "", "post_date": "03/22/2022 06:22:36", "content": "Hi @harshitsheoran , can you tell the gap between your CV-LB when considering only 9% individual ids ?
", "votes": null, "replies": [] }, { "id": 1731316, "author_name": "harshitsheoran", "author_url": "", "post_date": "03/22/2022 08:29:08", "content": "-40 points, that is if the CV is .705, I get a LB of .745
", "votes": null, "replies": [] } ] }, { "id": 1730545, "author_name": "clemchris", "author_url": "", "post_date": "03/21/2022 12:03:29", "content": "PyTorch
\n5 fold split
\nFold 0 single model
\nImage size 384
EfficientNet B7
\nCV: 0.841
\nLB: 0.749
ConvNeXt Base
\nCV: 0.825
\nLB: 0.741
Really considering teaming up with you…, especially cuz most of your code is public right now, I can see a lot of places to improve, and you look like someone who can execute tasks when guided (We are definitely going to get a huge boost if we team up). Contact me with your discord id (you can send me a mail), I will add you for further communication.
", "votes": null, "replies": [] }, { "id": 1730661, "author_name": "clemchris", "author_url": "", "post_date": "03/21/2022 14:02:36", "content": "Sent you a mail
", "votes": null, "replies": [] }, { "id": 1730730, "author_name": "deepkim", "author_url": "", "post_date": "03/21/2022 15:07:23", "content": "the really good score with convnext
", "votes": null, "replies": [] }, { "id": 1733499, "author_name": "zekunn", "author_url": "", "post_date": "03/24/2022 10:41:38", "content": "Do you train on Kaggle devices?
", "votes": null, "replies": [] }, { "id": 1733876, "author_name": "harshitsheoran", "author_url": "", "post_date": "03/24/2022 17:30:43", "content": "No, we use local machines, those models were trained using 1x 3090
", "votes": null, "replies": [] } ] }, { "id": 1731696, "author_name": "jerrykun", "author_url": "", "post_date": "03/22/2022 15:51:42", "content": "Pytorch
\n5 fold split
\nFold 0 single model
EffNet-B7 (img_size=768)
\nCV: 816
\nLB: 717
ConvNeXt-large (img_size=384)
\nCV: 813
\nLB: 715
Tensorflow
\n5 fold split
\nFold 4
\nEfficientNet-B5
\nImage Size: 768
\nVal Loss: 7.6865
\nCV : 0.8174
\nLB : 0.789
Which dataset do you use? Custom or published here?
", "votes": null, "replies": [] }, { "id": 1733442, "author_name": "kevin1742064161", "author_url": "", "post_date": "03/24/2022 09:29:26", "content": "Custom Dataset
", "votes": null, "replies": [] } ] }, { "id": 1736587, "author_name": "harshitsheoran", "author_url": "", "post_date": "03/27/2022 13:43:03", "content": "Pytorch
\n5-Fold Model
\nModel / Size: EfficientNet-B5, 384 size
CV: 0.755 (I calculate CV with only 9% individual ids in my valid set but I train (like everyone does) with about 20% individual ids in valid_set)
\nOne-Fold-LB: 0.775!
\nAll-Fold-LB: 0.798!
Looking forward to scores with bigger image size and better models!
\nAlso, would like TF-TPU pipeline team to join us…