{ "id": 212411, "title": "Any high solo model performance?", "url": "/competitions/cassava-leaf-disease-classification/discussion/212411", "author_name": "Avik", "post_date": "2021-01-18T19:10:19.401000", "votes": 14, "comment_count": 41, "views": 0, "content": "

Hello to people on the top of the LB,
\nAny solo model (with folds/augs/merged data) performing 90.2 or higher, or is it just ensembling taking us forward?

", "messages": [ { "id": 1162195, "postDate": "2021-01-21T03:26:53.477Z", "content": "

ViT 384x384 LB:0.904+ ;(single model with 5 folds tta)
\nEff-b4 512x512 LB:0.902+; (single model with 5 folds tta)
\nensemble LB:0.906+; (ensemble model with 5 folds tta)
\ndetails train notebook.

", "rawMarkdown": "ViT 384x384 LB:0.904+ ;(single model with 5 folds tta)\nEff-b4 512x512 LB:0.902+; (single model with 5 folds tta)\nensemble LB:0.906+; (ensemble model with 5 folds tta)\ndetails train [notebook](https://www.kaggle.com/yingpengchen/pytorch-cldc-train-with-vit/edit/run/51816828).", "votes": 19, "replies": [ { "id": 1162218, "postDate": "2021-01-21T03:40:35.403Z", "content": "

Thanks for generous sharing!
\nI'll give it a try.

", "rawMarkdown": "Thanks for generous sharing! \nI'll give it a try.", "votes": 1 }, { "id": 1162701, "postDate": "2021-01-21T09:30:04.423Z", "content": "

how many tta do you use? light or heavy tta?

", "rawMarkdown": "how many tta do you use? light or heavy tta?", "votes": 2 }, { "id": 1163820, "postDate": "2021-01-22T00:48:19.287Z", "content": "

hi,the vit you use is base or large?Thanks!

", "rawMarkdown": "hi,the vit you use is base or large?Thanks!", "votes": 1 }, { "id": 1163823, "postDate": "2021-01-22T00:50:34.757Z", "content": "

@qixinyan 4 x heavy tta

", "rawMarkdown": "@qixinyan 4 x heavy tta", "votes": 2 }, { "id": 1163824, "postDate": "2021-01-22T00:51:29.990Z", "content": "

@bcwang vit_base_patch16_384

", "rawMarkdown": "@bcwang vit_base_patch16_384", "votes": 2 }, { "id": 1164383, "postDate": "2021-01-22T10:45:33.243Z", "content": "

impressive,what's your cv without tta?

", "rawMarkdown": "impressive,what's your cv without tta?", "votes": 1 }, { "id": 1164562, "postDate": "2021-01-22T12:55:23.547Z", "content": "

@senkin13 5 folds cv about 89.2 without tta in training.

", "rawMarkdown": "@senkin13 5 folds cv about 89.2 without tta in training.", "votes": 2 }, { "id": 1164586, "postDate": "2021-01-22T13:09:49.060Z", "content": "

labelsmooth 0.3 is a good choice?

", "rawMarkdown": "labelsmooth 0.3 is a good choice? " }, { "id": 1164597, "postDate": "2021-01-22T13:19:29.843Z", "content": "

labelsmooth 0.3 is a good choice for my model.

", "rawMarkdown": "labelsmooth 0.3 is a good choice for my model." }, { "id": 1164607, "postDate": "2021-01-22T13:29:27.383Z", "content": "

thanks, i trained the vit using the notebook you mentioned(just training data of this year), but the lb performance is poor. For VIT, resize or crop, which is a good way to test? I always got a label 2 for this one test image…hhh it 's wrong obviously.

", "rawMarkdown": "thanks, i trained the vit using the notebook you mentioned(just training data of this year), but the lb performance is poor. For VIT, resize or crop, which is a good way to test? I always got a label 2 for this one test image...hhh it 's wrong obviously. ", "votes": 1 }, { "id": 1164667, "postDate": "2021-01-22T14:06:22.420Z", "content": "

@qixinyan add 2019 data remove dup data, crop for test and more augs. Good luck!

", "rawMarkdown": "@qixinyan add 2019 data remove dup data, crop for test and more augs. Good luck!" }, { "id": 1167158, "postDate": "2021-01-24T05:15:15.507Z", "content": "

@yingpengchen ,you local cv is 0.892,you can get lb 0.904?I use this vit model and get local cv is 0.894,but I only get lb is 0.898.

", "rawMarkdown": "@yingpengchen ,you local cv is 0.892,you can get lb 0.904?I use this vit model and get local cv is 0.894,but I only get lb is 0.898." }, { "id": 1167160, "postDate": "2021-01-24T05:17:12.277Z", "content": "

@qixinyan I also get this problem.I use vit model and get local cv is 0.894,but I only get the lb is 0.898.

", "rawMarkdown": "@qixinyan I also get this problem.I use vit model and get local cv is 0.894,but I only get the lb is 0.898." }, { "id": 1167183, "postDate": "2021-01-24T05:49:54.523Z", "content": "

@bcwang Yes, your model may overfit and you should try more augs like cutout、fmix、cutmix、mixup based on my train notebook.

", "rawMarkdown": "@bcwang Yes, your model may overfit and you should try more augs like cutout、fmix、cutmix、mixup based on my train notebook. " }, { "id": 1167689, "postDate": "2021-01-24T12:27:55.207Z", "content": "

@yingpengchen Thanks for you reply! I adjust the tta times,and I get the lb score is 0.900.I use the lr is 6e-5,and train 10 epochs with one epoch warmup,and I added the cutmix.

", "rawMarkdown": "@yingpengchen Thanks for you reply! I adjust the tta times,and I get the lb score is 0.900.I use the lr is 6e-5,and train 10 epochs with one epoch warmup,and I added the cutmix." }, { "id": 1173114, "postDate": "2021-01-27T17:23:25.143Z", "content": "

thanks for sharing!

", "rawMarkdown": "thanks for sharing!" }, { "id": 1173208, "postDate": "2021-01-27T18:21:59.143Z", "content": "

@yingpengchen can you please share what augmentations you use for TTA? Augmentations such as randomresizedcrop cause my submission LB score to be inconsistent.

", "rawMarkdown": "@yingpengchen can you please share what augmentations you use for TTA? Augmentations such as randomresizedcrop cause my submission LB score to be inconsistent." } ] }, { "id": 1158770, "postDate": "2021-01-18T19:10:19.400Z", "content": "

Hello to people on the top of the LB,
\nAny solo model (with folds/augs/merged data) performing 90.2 or higher, or is it just ensembling taking us forward?

", "rawMarkdown": "Hello to people on the top of the LB,\nAny solo model (with folds/augs/merged data) performing 90.2 or higher, or is it just ensembling taking us forward?\n", "votes": 14 }, { "id": 1161028, "postDate": "2021-01-20T09:20:20.407Z", "content": "

@junyingsg we are not at the top of LB.
\nHowever, we can achieve 0.903 with a single model and 5 folds.

", "rawMarkdown": "@junyingsg we are not at the top of LB.\nHowever, we can achieve 0.903 with a single model and 5 folds.", "votes": 6, "replies": [ { "id": 1161067, "postDate": "2021-01-20T10:05:33.690Z", "rawMarkdown": "", "votes": -3, "isDeleted": true } ] }, { "id": 1172633, "postDate": "2021-01-27T13:27:20.413Z", "content": "

EfficientNet-B4, 512x512, only 2020 data, standard augs (no mixup, etc.)
\nsingle model, 5 folds
\nCV: 0.898
\nCV (4xTTA): 0.900
\nLB: 0.902
\nLB (4xTTA): 0.903

\n

upd: CV=0.901 LB=0.904

", "rawMarkdown": "EfficientNet-B4, 512x512, only 2020 data, standard augs (no mixup, etc.)\nsingle model, 5 folds\nCV: 0.898\nCV (4xTTA): 0.900\nLB: 0.902\nLB (4xTTA): 0.903\n\nupd: CV=0.901 LB=0.904", "votes": 3, "replies": [ { "id": 1172878, "postDate": "2021-01-27T15:28:37.943Z", "content": "

Can you tell me what does 2020 data mean, I have seen it in many discussion but do not know what it is, thanks

", "rawMarkdown": "Can you tell me what does 2020 data mean, I have seen it in many discussion but do not know what it is, thanks" }, { "id": 1172893, "postDate": "2021-01-27T15:34:29.367Z", "content": "

It's the data provided by this competition. There was a previous Cassava classification competition in 2019, and some people are using that data.

", "rawMarkdown": "It's the data provided by this competition. There was a previous Cassava classification competition in 2019, and some people are using that data." }, { "id": 1173462, "postDate": "2021-01-27T22:57:36.727Z", "content": "

@sparakhin Can you please share what augmentations you use for TTA?

", "rawMarkdown": "@sparakhin Can you please share what augmentations you use for TTA?" }, { "id": 1173573, "postDate": "2021-01-28T01:10:09.327Z", "content": "

Just horizontal/vertical flips

", "rawMarkdown": "Just horizontal/vertical flips" }, { "id": 1173584, "postDate": "2021-01-28T01:34:33.260Z", "content": "

Thank you for your generous answer

", "rawMarkdown": "Thank you for your generous answer" }, { "id": 1173590, "postDate": "2021-01-28T01:39:56.680Z", "content": "

But it depends on the model and training augs. I pick the best TTA combination on cross-validation, and so far it always helps at LB (+0.001).

", "rawMarkdown": "But it depends on the model and training augs. I pick the best TTA combination on cross-validation, and so far it always helps at LB (+0.001)." }, { "id": 1173592, "postDate": "2021-01-28T01:46:20.757Z", "content": "

Thank you for sharing. What do you mean by 4x TTA then if there are only horizontal and vertical flips?

", "rawMarkdown": "Thank you for sharing. What do you mean by 4x TTA then if there are only horizontal and vertical flips?" }, { "id": 1173593, "postDate": "2021-01-28T01:46:54.737Z", "content": "

update: just found an issue in the code, now the best single EffNet-b4 CV/LB is 0.901/0.904 (with TTA)

", "rawMarkdown": "update: just found an issue in the code, now the best single EffNet-b4 CV/LB is 0.901/0.904 (with TTA)" }, { "id": 1173598, "postDate": "2021-01-28T02:04:58.447Z", "content": "
\n

4x TTA then if there are only horizontal and vertical flips

\n
\n

It's a typo, I use 3xTTA for now. So far only flips give me stable improvements on all folds (CV) and LB (but I'm going to experiment with TTA later)

", "rawMarkdown": "> 4x TTA then if there are only horizontal and vertical flips\n\nIt's a typo, I use 3xTTA for now. So far only flips give me stable improvements on all folds (CV) and LB (but I'm going to experiment with TTA later)", "votes": 1 }, { "id": 1173606, "postDate": "2021-01-28T02:10:29.223Z", "content": "

Ok thanks! Your score is amazing for a relatively simpler model compared to other solutions. I wish mine would go up. I am stuck at 90.2 :(

", "rawMarkdown": "Ok thanks! Your score is amazing for a relatively simpler model compared to other solutions. I wish mine would go up. I am stuck at 90.2 :(" } ] }, { "id": 1160569, "postDate": "2021-01-20T02:41:43.153Z", "content": "

CV LB 0.904+ is possible. @serigne posted about one in another thread.

", "rawMarkdown": "CV LB 0.904+ is possible. @serigne posted about one in another [thread][1].\n\n[1]: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/203111#1111578", "votes": 4, "replies": [ { "id": 1161396, "postDate": "2021-01-20T14:47:13.790Z", "content": "

@cdeotte
\nI guess those at the top (including you) have probably higher ;)

", "rawMarkdown": "@cdeotte \nI guess those at the top (including you) have probably higher ;)", "votes": 2 }, { "id": 1162085, "postDate": "2021-01-21T01:15:47.970Z", "content": "

@serigne @cdeotte
\nYES or NO question, is this a GPU related matter? (T_T)

", "rawMarkdown": "@serigne @cdeotte \nYES or NO question, is this a GPU related matter? (T_T)", "votes": 1 }, { "id": 1162475, "postDate": "2021-01-21T07:06:39.043Z", "content": "

If I say yes, you will rent a DGX A100 and win the competition ? ^^

\n

More seriously, I don't think it's just GPU related matter. And others ongoing competitions are more GPU hungry (RANZCR, Hubmap)

", "rawMarkdown": "If I say yes, you will rent a DGX A100 and win the competition ? ^^\n\nMore seriously, I don't think it's just GPU related matter. And others ongoing competitions are more GPU hungry (RANZCR, Hubmap)", "votes": 3 }, { "id": 1163281, "postDate": "2021-01-21T16:04:55.050Z", "content": "

A little bit hint gathering is vital for survival :)

", "rawMarkdown": "A little bit hint gathering is vital for survival :)", "votes": 1 } ] }, { "id": 1160496, "postDate": "2021-01-20T01:12:39.233Z", "content": "

I believe the top of LB are getting at least 0.903 with a single model, 5 folds.

", "rawMarkdown": "I believe the top of LB are getting at least 0.903 with a single model, 5 folds. ", "votes": 2 }, { "id": 1174684, "postDate": "2021-01-28T16:40:50.933Z", "content": "

my single efficient net 5, 512X512 SIZE -> CV 0.8997 LB 0.901

", "rawMarkdown": "my single efficient net 5, 512X512 SIZE -> CV 0.8997 LB 0.901", "replies": [ { "id": 1174691, "postDate": "2021-01-28T16:44:58.063Z", "content": "

Are you using TTA?

", "rawMarkdown": "Are you using TTA?" }, { "id": 1175573, "postDate": "2021-01-29T08:26:17.137Z", "content": "

when inferencing, I use 3x TTA

", "rawMarkdown": "when inferencing, I use 3x TTA" } ] }, { "id": 1176111, "postDate": "2021-01-29T14:07:49.183Z", "content": "

Thank you everyone for helping :D

", "rawMarkdown": "Thank you everyone for helping :D " } ], "comments": [ { "id": 1162195, "author_name": "Roc", "author_url": "", "post_date": "2021-01-21T03:26:53.477000", "content": "

ViT 384x384 LB:0.904+ ;(single model with 5 folds tta)
\nEff-b4 512x512 LB:0.902+; (single model with 5 folds tta)
\nensemble LB:0.906+; (ensemble model with 5 folds tta)
\ndetails train notebook.

", "votes": 19, "replies": [ { "id": 1162218, "author_name": "Tian", "author_url": "", "post_date": "2021-01-21T03:40:35.403000", "content": "

Thanks for generous sharing!
\nI'll give it a try.

", "votes": 1, "replies": [] }, { "id": 1162701, "author_name": "Qixin Yan", "author_url": "", "post_date": "2021-01-21T09:30:04.423000", "content": "

how many tta do you use? light or heavy tta?

", "votes": 2, "replies": [] }, { "id": 1163820, "author_name": "Bcw93", "author_url": "", "post_date": "2021-01-22T00:48:19.287000", "content": "

hi,the vit you use is base or large?Thanks!

", "votes": 1, "replies": [] }, { "id": 1163823, "author_name": "Roc", "author_url": "", "post_date": "2021-01-22T00:50:34.757000", "content": "

@qixinyan 4 x heavy tta

", "votes": 2, "replies": [] }, { "id": 1163824, "author_name": "Roc", "author_url": "", "post_date": "2021-01-22T00:51:29.990000", "content": "

@bcwang vit_base_patch16_384

", "votes": 2, "replies": [] }, { "id": 1164383, "author_name": "senkin13", "author_url": "", "post_date": "2021-01-22T10:45:33.243000", "content": "

impressive,what's your cv without tta?

", "votes": 1, "replies": [] }, { "id": 1164562, "author_name": "Roc", "author_url": "", "post_date": "2021-01-22T12:55:23.547000", "content": "

@senkin13 5 folds cv about 89.2 without tta in training.

", "votes": 2, "replies": [] }, { "id": 1164586, "author_name": "Qixin Yan", "author_url": "", "post_date": "2021-01-22T13:09:49.060000", "content": "

labelsmooth 0.3 is a good choice?

", "votes": 0, "replies": [] }, { "id": 1164597, "author_name": "Roc", "author_url": "", "post_date": "2021-01-22T13:19:29.843000", "content": "

labelsmooth 0.3 is a good choice for my model.

", "votes": 0, "replies": [] }, { "id": 1164607, "author_name": "Qixin Yan", "author_url": "", "post_date": "2021-01-22T13:29:27.383000", "content": "

thanks, i trained the vit using the notebook you mentioned(just training data of this year), but the lb performance is poor. For VIT, resize or crop, which is a good way to test? I always got a label 2 for this one test image…hhh it 's wrong obviously.

", "votes": 1, "replies": [] }, { "id": 1164667, "author_name": "Roc", "author_url": "", "post_date": "2021-01-22T14:06:22.420000", "content": "

@qixinyan add 2019 data remove dup data, crop for test and more augs. Good luck!

", "votes": 0, "replies": [] }, { "id": 1167158, "author_name": "Bcw93", "author_url": "", "post_date": "2021-01-24T05:15:15.507000", "content": "

@yingpengchen ,you local cv is 0.892,you can get lb 0.904?I use this vit model and get local cv is 0.894,but I only get lb is 0.898.

", "votes": 0, "replies": [] }, { "id": 1167160, "author_name": "Bcw93", "author_url": "", "post_date": "2021-01-24T05:17:12.277000", "content": "

@qixinyan I also get this problem.I use vit model and get local cv is 0.894,but I only get the lb is 0.898.

", "votes": 0, "replies": [] }, { "id": 1167183, "author_name": "Roc", "author_url": "", "post_date": "2021-01-24T05:49:54.523000", "content": "

@bcwang Yes, your model may overfit and you should try more augs like cutout、fmix、cutmix、mixup based on my train notebook.

", "votes": 0, "replies": [] }, { "id": 1167689, "author_name": "Bcw93", "author_url": "", "post_date": "2021-01-24T12:27:55.207000", "content": "

@yingpengchen Thanks for you reply! I adjust the tta times,and I get the lb score is 0.900.I use the lr is 6e-5,and train 10 epochs with one epoch warmup,and I added the cutmix.

", "votes": 0, "replies": [] }, { "id": 1173114, "author_name": "DarknessZX", "author_url": "", "post_date": "2021-01-27T17:23:25.143000", "content": "

thanks for sharing!

", "votes": 0, "replies": [] }, { "id": 1173208, "author_name": "ayu055", "author_url": "", "post_date": "2021-01-27T18:21:59.143000", "content": "

@yingpengchen can you please share what augmentations you use for TTA? Augmentations such as randomresizedcrop cause my submission LB score to be inconsistent.

", "votes": 0, "replies": [] } ] }, { "id": 1161028, "author_name": "Hai Nam Nguyen", "author_url": "", "post_date": "2021-01-20T09:20:20.407000", "content": "

@junyingsg we are not at the top of LB.
\nHowever, we can achieve 0.903 with a single model and 5 folds.

", "votes": 6, "replies": [ { "id": 1161067, "author_name": "", "author_url": "", "post_date": "2021-01-20T10:05:33.690000", "content": "", "votes": -3, "replies": [] } ] }, { "id": 1172633, "author_name": "Serhii Parakhin", "author_url": "", "post_date": "2021-01-27T13:27:20.413000", "content": "

EfficientNet-B4, 512x512, only 2020 data, standard augs (no mixup, etc.)
\nsingle model, 5 folds
\nCV: 0.898
\nCV (4xTTA): 0.900
\nLB: 0.902
\nLB (4xTTA): 0.903

\n

upd: CV=0.901 LB=0.904

", "votes": 3, "replies": [ { "id": 1172878, "author_name": "老肥", "author_url": "", "post_date": "2021-01-27T15:28:37.943000", "content": "

Can you tell me what does 2020 data mean, I have seen it in many discussion but do not know what it is, thanks

", "votes": 0, "replies": [] }, { "id": 1172893, "author_name": "Junyi Ng", "author_url": "", "post_date": "2021-01-27T15:34:29.367000", "content": "

It's the data provided by this competition. There was a previous Cassava classification competition in 2019, and some people are using that data.

", "votes": 0, "replies": [] }, { "id": 1173462, "author_name": "ayu055", "author_url": "", "post_date": "2021-01-27T22:57:36.727000", "content": "

@sparakhin Can you please share what augmentations you use for TTA?

", "votes": 0, "replies": [] }, { "id": 1173573, "author_name": "Serhii Parakhin", "author_url": "", "post_date": "2021-01-28T01:10:09.327000", "content": "

Just horizontal/vertical flips

", "votes": 0, "replies": [] }, { "id": 1173584, "author_name": "老肥", "author_url": "", "post_date": "2021-01-28T01:34:33.260000", "content": "

Thank you for your generous answer

", "votes": 0, "replies": [] }, { "id": 1173590, "author_name": "Serhii Parakhin", "author_url": "", "post_date": "2021-01-28T01:39:56.680000", "content": "

But it depends on the model and training augs. I pick the best TTA combination on cross-validation, and so far it always helps at LB (+0.001).

", "votes": 0, "replies": [] }, { "id": 1173592, "author_name": "ayu055", "author_url": "", "post_date": "2021-01-28T01:46:20.757000", "content": "

Thank you for sharing. What do you mean by 4x TTA then if there are only horizontal and vertical flips?

", "votes": 0, "replies": [] }, { "id": 1173593, "author_name": "Serhii Parakhin", "author_url": "", "post_date": "2021-01-28T01:46:54.737000", "content": "

update: just found an issue in the code, now the best single EffNet-b4 CV/LB is 0.901/0.904 (with TTA)

", "votes": 0, "replies": [] }, { "id": 1173598, "author_name": "Serhii Parakhin", "author_url": "", "post_date": "2021-01-28T02:04:58.447000", "content": "
\n

4x TTA then if there are only horizontal and vertical flips

\n
\n

It's a typo, I use 3xTTA for now. So far only flips give me stable improvements on all folds (CV) and LB (but I'm going to experiment with TTA later)

", "votes": 1, "replies": [] }, { "id": 1173606, "author_name": "ayu055", "author_url": "", "post_date": "2021-01-28T02:10:29.223000", "content": "

Ok thanks! Your score is amazing for a relatively simpler model compared to other solutions. I wish mine would go up. I am stuck at 90.2 :(

", "votes": 0, "replies": [] } ] }, { "id": 1160569, "author_name": "Chris Deotte", "author_url": "", "post_date": "2021-01-20T02:41:43.153000", "content": "

CV LB 0.904+ is possible. @serigne posted about one in another thread.

", "votes": 4, "replies": [ { "id": 1161396, "author_name": "Serigne ", "author_url": "", "post_date": "2021-01-20T14:47:13.790000", "content": "

@cdeotte
\nI guess those at the top (including you) have probably higher ;)

", "votes": 2, "replies": [] }, { "id": 1162085, "author_name": "Avik", "author_url": "", "post_date": "2021-01-21T01:15:47.970000", "content": "

@serigne @cdeotte
\nYES or NO question, is this a GPU related matter? (T_T)

", "votes": 1, "replies": [] }, { "id": 1162475, "author_name": "Serigne ", "author_url": "", "post_date": "2021-01-21T07:06:39.043000", "content": "

If I say yes, you will rent a DGX A100 and win the competition ? ^^

\n

More seriously, I don't think it's just GPU related matter. And others ongoing competitions are more GPU hungry (RANZCR, Hubmap)

", "votes": 3, "replies": [] }, { "id": 1163281, "author_name": "Avik", "author_url": "", "post_date": "2021-01-21T16:04:55.050000", "content": "

A little bit hint gathering is vital for survival :)

", "votes": 1, "replies": [] } ] }, { "id": 1160496, "author_name": "Junyi Ng", "author_url": "", "post_date": "2021-01-20T01:12:39.233000", "content": "

I believe the top of LB are getting at least 0.903 with a single model, 5 folds.

", "votes": 2, "replies": [] }, { "id": 1174684, "author_name": "kaggler", "author_url": "", "post_date": "2021-01-28T16:40:50.933000", "content": "

my single efficient net 5, 512X512 SIZE -> CV 0.8997 LB 0.901

", "votes": 0, "replies": [ { "id": 1174691, "author_name": "ayu055", "author_url": "", "post_date": "2021-01-28T16:44:58.063000", "content": "

Are you using TTA?

", "votes": 0, "replies": [] }, { "id": 1175573, "author_name": "kaggler", "author_url": "", "post_date": "2021-01-29T08:26:17.137000", "content": "

when inferencing, I use 3x TTA

", "votes": 0, "replies": [] } ] }, { "id": 1176111, "author_name": "Avik", "author_url": "", "post_date": "2021-01-29T14:07:49.183000", "content": "

Thank you everyone for helping :D

", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1162195": "ViT 384x384 LB:0.904+ ;(single model with 5 folds tta)\nEff-b4 512x512 LB:0.902+; (single model with 5 folds tta)\nensemble LB:0.906+; (ensemble model with 5 folds tta)\ndetails train [notebook](https://www.kaggle.com/yingpengchen/pytorch-cldc-train-with-vit/edit/run/51816828).", "1158770": "Hello to people on the top of the LB,\nAny solo model (with folds/augs/merged data) performing 90.2 or higher, or is it just ensembling taking us forward?\n", "1161028": "@junyingsg we are not at the top of LB.\nHowever, we can achieve 0.903 with a single model and 5 folds.", "1172633": "EfficientNet-B4, 512x512, only 2020 data, standard augs (no mixup, etc.)\nsingle model, 5 folds\nCV: 0.898\nCV (4xTTA): 0.900\nLB: 0.902\nLB (4xTTA): 0.903\n\nupd: CV=0.901 LB=0.904", "1160569": "CV LB 0.904+ is possible. @serigne posted about one in another [thread][1].\n\n[1]: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/203111#1111578", "1160496": "I believe the top of LB are getting at least 0.903 with a single model, 5 folds. ", "1174684": "my single efficient net 5, 512X512 SIZE -> CV 0.8997 LB 0.901", "1176111": "Thank you everyone for helping :D " } }