![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1807054%2F3c77090ae52887d54c4aef2f0ea2db73%2FUntitled.png?generation=1607176663723372&alt=media\")
Yay an image competition!!
\nEff-b0 single fold
\nCV 0.89
\nLB: 0.893
-- updates --
\nEff-b0 single fold + TTA
\nhttps://www.kaggle.com/kyoshioka47/ttach-kaggle
LB: 0.896
\nEff-b2 5-folds
\nLB: 0.901
Yay an image competition!!
\nEff-b0 single fold
\nCV 0.89
\nLB: 0.893
-- updates --
\nEff-b0 single fold + TTA
\nhttps://www.kaggle.com/kyoshioka47/ttach-kaggle
LB: 0.896
\nEff-b2 5-folds
\nLB: 0.901
resnext50_32x4d
\n256x256
\n10 epochs
\nno TTA
\nCV: 0.87372 (https://www.kaggle.com/yasufuminakama/cassava-resnext50-32x4d-starter-training)
\nLB: 0.890 (https://www.kaggle.com/yasufuminakama/cassava-resnext50-32x4d-starter-inference)
seresnext101_32x4d
\n320x320
\n10 epochs
\nCV: 0.887
\nLB(0TTA): 0.894
\nLB(3TTA): 0.898
resnet18, single fold
\nCV: 0.895
\nLB: 0.898
Your augmentations must be powerful. I tried simple augs with tta using ResNet18 and can't cross 0.87 on LB.
", "rawMarkdown": "Your augmentations must be powerful. I tried simple augs with tta using ResNet18 and can't cross 0.87 on LB." }, { "id": 1093454, "postDate": "2020-11-27T18:17:40.647Z", "content": "@phalanx can you tell what Image size you are using ? also are you using TTA ?
", "rawMarkdown": "@phalanx can you tell what Image size you are using ? also are you using TTA ?\n" } ] }, { "id": 1091621, "postDate": "2020-11-26T06:24:20.117Z", "content": "resnext50
\n448*448
\n5folds
\nCV 0.893
\nLB 0.903
efficientnet-b0
\n5 epochs
\nno TTA
\nCV .892
\nLB .893
Good result for this network! What image size did you use?
", "rawMarkdown": "Good result for this network! What image size did you use?" }, { "id": 1094040, "postDate": "2020-11-28T09:05:41.123Z", "content": "I use 512x512
", "rawMarkdown": "I use 512x512", "votes": 1 }, { "id": 1094136, "postDate": "2020-11-28T10:54:21.887Z", "content": "Thanks! I will try larger sizes too
", "rawMarkdown": "Thanks! I will try larger sizes too" } ] }, { "id": 1169365, "postDate": "2021-01-25T13:08:28.847Z", "content": "Ensemble TTA CV 0.903 - TTA lb 0.904
", "rawMarkdown": "Ensemble TTA CV 0.903 - TTA lb 0.904 ", "votes": 1 }, { "id": 1100260, "postDate": "2020-12-03T00:19:09.663Z", "content": "Wow, a sight for sore eyes after the MoA competition! The CV and LB scores are extremely close to each other here.
", "rawMarkdown": "Wow, a sight for sore eyes after the MoA competition! The CV and LB scores are extremely close to each other here.", "votes": 1 }, { "id": 1095029, "postDate": "2020-11-29T07:44:00.177Z", "content": "efficientnet-b3
\nsingle fold
\n20 epochs
no TTA
\nLB .892
\nCV 0.896
with TTA
\nLB 894
efficientnet-b6
\n12 epochs
\nno TTA
\nCV .8910
\nLB .9010
actually, I'm having a very hard time improving the best single fold score.
\nAll improvements come from ensembling models and its getting quite boring..
What ensembling technique are u using?
", "rawMarkdown": "What ensembling technique are u using?\n" }, { "id": 1091040, "postDate": "2020-11-25T18:31:54.770Z", "content": "Same for me, I have similar CV as you and can't seem to increase it
", "rawMarkdown": "Same for me, I have similar CV as you and can't seem to increase it " }, { "id": 1091371, "postDate": "2020-11-26T00:59:03.880Z", "content": "@mrinath I'm just adding up all the predictions. Since argmax is taken, that's fine.
", "rawMarkdown": "@mrinath I'm just adding up all the predictions. Since argmax is taken, that's fine." }, { "id": 1094150, "postDate": "2020-11-28T11:07:42.503Z", "content": "\n\nAll improvements come from ensembling models and its getting quite boring..
\n
I think one could achieve +0.910 with single model before the end.
\nThe competion just begins. There are still almost 3 months to go.
\nEdit 07/12/2020: I made this comment before the LB was updated. The current best LB after update (+0.905) would like be +0.910 before update.
\nMy new prediction for best single model before the end is +0.908
", "rawMarkdown": "> All improvements come from ensembling models and its getting quite boring..\n\nI think one could achieve +0.910 with single model before the end. \n\nThe competion just begins. There are still almost 3 months to go. \n\nEdit 07/12/2020: I made this comment before the LB was updated. The current best LB after update (+0.905) would like be +0.910 before update. \n\nMy new prediction for best single model before the end is +0.908\n", "votes": 2 }, { "id": 1094156, "postDate": "2020-11-28T11:11:47.417Z", "content": "'single model' means only one fold or k-folds?
", "rawMarkdown": "'single model' means only one fold or k-folds?" }, { "id": 1094405, "postDate": "2020-11-28T15:38:22.573Z", "content": "one or k-folds whatever, but most importantly improvements from data (and external data) processing, model architecture and training pipleline.
", "rawMarkdown": "one or k-folds whatever, but most importantly improvements from data (and external data) processing, model architecture and training pipleline. ", "votes": 1 }, { "id": 1094442, "postDate": "2020-11-28T16:06:29.147Z", "content": "It’s kinda weird the scores are all stagnant at 0.90x
", "rawMarkdown": "It’s kinda weird the scores are all stagnant at 0.90x" }, { "id": 1095070, "postDate": "2020-11-29T08:16:46.083Z", "content": "@reighns do you think this is because of some labelling issue where almost everyone's model is failing to predict
", "rawMarkdown": "@reighns do you think this is because of some labelling issue where almost everyone's model is failing to predict", "votes": 2 }, { "id": 1109666, "postDate": "2020-12-12T00:02:35.923Z", "rawMarkdown": "", "isDeleted": true } ] }, { "id": 1096131, "postDate": "2020-11-30T08:31:06.477Z", "content": "EffNet B4
\n5 folds
\nCV 0.901
\nLB 0.905 (No TTA)
Hi buddy, CV seems to be consistent with LB?
", "rawMarkdown": "Hi buddy, CV seems to be consistent with LB?" }, { "id": 1096548, "postDate": "2020-11-30T15:35:05.913Z", "content": "Yeah it seems stable and pretty much aligned with the LB after some modification of my training pipeline
\nGlad to see you here :)
", "rawMarkdown": "Yeah it seems stable and pretty much aligned with the LB after some modification of my training pipeline\n\nGlad to see you here :)" }, { "id": 1167316, "postDate": "2021-01-24T07:12:57.067Z", "content": "@serigne what is img size you using 512 ?
", "rawMarkdown": "@serigne what is img size you using 512 ?" } ] }, { "id": 1169348, "postDate": "2021-01-25T12:59:19.937Z", "content": "seresnext101_32x4d
\n512x512
\n5-folds
\nBest single Fold: 0.905
\nCV: 0.901; LB: 0.902
best single model score .90 and ensemble 90.6 effnet and others
", "rawMarkdown": "best single model score .90 and ensemble 90.6 effnet and others", "replies": [ { "id": 1168324, "postDate": "2021-01-24T21:51:36.577Z", "content": "Could you provide more details on the ensemble?
", "rawMarkdown": "Could you provide more details on the ensemble?" }, { "id": 1168325, "postDate": "2021-01-24T21:58:41.307Z", "content": "@jaideepvalani Are these your CV or LB scores?
", "rawMarkdown": "@jaideepvalani Are these your CV or LB scores?" }, { "id": 1168973, "postDate": "2021-01-25T09:17:41.217Z", "content": "Lbs, cvs are either same or less by .003 to .004
", "rawMarkdown": "Lbs, cvs are either same or less by .003 to .004 ", "votes": 1 }, { "id": 1171524, "postDate": "2021-01-26T22:44:20.093Z", "content": "@jaideepvalani Can you tell the type of ensembling are you using? Is it simple averaging or something complex?
", "rawMarkdown": "@jaideepvalani Can you tell the type of ensembling are you using? Is it simple averaging or something complex?" } ] }, { "id": 1100462, "postDate": "2020-12-03T05:16:05.217Z", "content": "vit cv 90+ lb87+
", "rawMarkdown": "vit cv 90+ lb87+" }, { "id": 1095244, "postDate": "2020-11-29T11:54:46.663Z", "content": "resnext50
\n5-Folds
\nno TTA
CV 0.893
\nLB 0.902
resnext50
\n5-Folds
\nno TTA
CV 0.895
\nLB 0.903
Has anyone experimented with sizes? bigger the sizes better the results? that is what's the relation of LB and size of the image
", "rawMarkdown": "Has anyone experimented with sizes? bigger the sizes better the results? that is what's the relation of LB and size of the image", "replies": [ { "id": 1095155, "postDate": "2020-11-29T09:54:36.360Z", "content": "From my experiments after a particular resolution model will stop getting better. 512x512 is an overkill the same or better results can be made with a lower resolution.
", "rawMarkdown": "From my experiments after a particular resolution model will stop getting better. 512x512 is an overkill the same or better results can be made with a lower resolution.", "votes": 1 }, { "id": 1102941, "postDate": "2020-12-05T13:51:44.390Z", "content": "The efficientnet paper has this to say:
Looking from all the results CV/LB relation is very close to 1; (still, can't trust what happens in private data), do think same imbalance is there in private
", "rawMarkdown": "Looking from all the results CV/LB relation is very close to 1; (still, can't trust what happens in private data), do think same imbalance is there in private" }, { "id": 1088650, "postDate": "2020-11-23T20:35:24.870Z", "content": "Sorry, I am a novice here :) What do you mean by single fold?
Typically you hold out 20% of the train data for a local validation dataset.
\nYou vary the 20%, so you have 5 different combinations of train data and validation data. Each one is called a \"fold\". \"Stratefied\" folds attempt to distribute the train data evenly, so you don't have one train dataset with all the class-3 data and another with all the \"class-1\" data.
\nFor your local cross validation, you combine the results of your five \"folds\" to get a local cross-validation (CV) score.
\nThen you run the test data against your five models and combine the scores. That gives you a Leaderboard score.
\nThe hope is that the five \"folds\" will average out any errors and be more accurate than one fold or just using the leaderboard score as a guide. This is usually true, and you have to learn to \"trust\" your cross-validation, rather than be guided by the leaderboard score (which is usually on a small subset of the real test data).
\nThere is nothing magic about 20% and 5 folds. Just tradition. You could do fewer folds or more, depending on your data set size and your processing time.
\nSo, the final answer - \"single fold\" is when somebody has not yet done multiple training runs, so they only have one validation dataset. Since it takes typically 5 times as long to run five folds, you typically try and get parameters close with a single fold and then start running five-folds. Also depends on how long a fold takes.
\n-Rich
", "rawMarkdown": "Typically you hold out 20% of the train data for a local validation dataset.\n\nYou vary the 20%, so you have 5 different combinations of train data and validation data. Each one is called a \"fold\". \"Stratefied\" folds attempt to distribute the train data evenly, so you don't have one train dataset with all the class-3 data and another with all the \"class-1\" data.\n\nFor your local cross validation, you combine the results of your five \"folds\" to get a local cross-validation (CV) score.\n\nThen you run the test data against your five models and combine the scores. That gives you a Leaderboard score.\n\nThe hope is that the five \"folds\" will average out any errors and be more accurate than one fold or just using the leaderboard score as a guide. This is usually true, and you have to learn to \"trust\" your cross-validation, rather than be guided by the leaderboard score (which is usually on a small subset of the real test data).\n\nThere is nothing magic about 20% and 5 folds. Just tradition. You could do fewer folds or more, depending on your data set size and your processing time.\n\nSo, the final answer - \"single fold\" is when somebody has not yet done multiple training runs, so they only have one validation dataset. Since it takes typically 5 times as long to run five folds, you typically try and get parameters close with a single fold and then start running five-folds. Also depends on how long a fold takes.\n\n-Rich", "votes": 12 }, { "id": 1088711, "postDate": "2020-11-23T21:54:04.760Z", "content": "Thank you very match for description! As I understand, single fold it the same thing as hold-out CV. Hope I can beat something serious and will write it here later)
", "rawMarkdown": "Thank you very match for description! As I understand, single fold it the same thing as hold-out CV. Hope I can beat something serious and will write it here later)", "votes": 1 } ] }, { "id": 1087598, "postDate": "2020-11-22T22:26:19.037Z", "content": "Model: fast-resnest50
\nsingle fold
\nAug: rotation, flip, brightness, contrast, cutout
\nTTA: 0
\nCV: 0.8715
\nLB: 0.8800
I believe I can push my CV/LB even further with Fast-Resnest.
", "rawMarkdown": "Model: fast-resnest50\nsingle fold\nAug: rotation, flip, brightness, contrast, cutout\nTTA: 0\nCV: 0.8715\nLB: 0.8800\n\nI believe I can push my CV/LB even further with Fast-Resnest. " }, { "id": 1087530, "postDate": "2020-11-22T20:10:58.467Z", "content": "Resnest50
\nsingle fold
\nCV: 0.9009
\nLB: 0.890
my lb is always lower than cv, should probably test other folds
", "rawMarkdown": "Resnest50\nsingle fold\nCV: 0.9009\nLB: 0.890\n\nmy lb is always lower than cv, should probably test other folds" }, { "id": 1086702, "postDate": "2020-11-22T00:40:31.070Z", "content": "resnest50d_4s2x40d
\n5 epochs
\nno TTA
\nCV .879
\nLB .892
resnext50_32x4d
\n256x256
\n10 epochs
\nno TTA
\nCV: 0.87372 (https://www.kaggle.com/yasufuminakama/cassava-resnext50-32x4d-starter-training)
\nLB: 0.890 (https://www.kaggle.com/yasufuminakama/cassava-resnext50-32x4d-starter-inference)
seresnext101_32x4d
\n320x320
\n10 epochs
\nCV: 0.887
\nLB(0TTA): 0.894
\nLB(3TTA): 0.898
resnet18, single fold
\nCV: 0.895
\nLB: 0.898
Your augmentations must be powerful. I tried simple augs with tta using ResNet18 and can't cross 0.87 on LB.
", "votes": 0, "replies": [] }, { "id": 1093454, "author_name": "Atharva Phatak", "author_url": "", "post_date": "2020-11-27T18:17:40.647000", "content": "@phalanx can you tell what Image size you are using ? also are you using TTA ?
", "votes": 0, "replies": [] } ] }, { "id": 1091621, "author_name": "Wang Xinliang", "author_url": "", "post_date": "2020-11-26T06:24:20.117000", "content": "resnext50
\n448*448
\n5folds
\nCV 0.893
\nLB 0.903
efficientnet-b0
\n5 epochs
\nno TTA
\nCV .892
\nLB .893
Good result for this network! What image size did you use?
", "votes": 0, "replies": [] }, { "id": 1094040, "author_name": "SiNpcw", "author_url": "", "post_date": "2020-11-28T09:05:41.123000", "content": "I use 512x512
", "votes": 1, "replies": [] }, { "id": 1094136, "author_name": "Oleg Khlevnov", "author_url": "", "post_date": "2020-11-28T10:54:21.887000", "content": "Thanks! I will try larger sizes too
", "votes": 0, "replies": [] } ] }, { "id": 1169365, "author_name": "Phaedrus", "author_url": "", "post_date": "2021-01-25T13:08:28.847000", "content": "Ensemble TTA CV 0.903 - TTA lb 0.904
", "votes": 1, "replies": [] }, { "id": 1100260, "author_name": "Tolga", "author_url": "", "post_date": "2020-12-03T00:19:09.663000", "content": "Wow, a sight for sore eyes after the MoA competition! The CV and LB scores are extremely close to each other here.
", "votes": 1, "replies": [] }, { "id": 1095029, "author_name": "DeepUnderstanding", "author_url": "", "post_date": "2020-11-29T07:44:00.177000", "content": "efficientnet-b3
\nsingle fold
\n20 epochs
no TTA
\nLB .892
\nCV 0.896
with TTA
\nLB 894
efficientnet-b6
\n12 epochs
\nno TTA
\nCV .8910
\nLB .9010
actually, I'm having a very hard time improving the best single fold score.
\nAll improvements come from ensembling models and its getting quite boring..
What ensembling technique are u using?
", "votes": 0, "replies": [] }, { "id": 1091040, "author_name": "Yann Majewski", "author_url": "", "post_date": "2020-11-25T18:31:54.770000", "content": "Same for me, I have similar CV as you and can't seem to increase it
", "votes": 0, "replies": [] }, { "id": 1091371, "author_name": "arutema47", "author_url": "", "post_date": "2020-11-26T00:59:03.880000", "content": "@mrinath I'm just adding up all the predictions. Since argmax is taken, that's fine.
", "votes": 0, "replies": [] }, { "id": 1094150, "author_name": "Serigne ", "author_url": "", "post_date": "2020-11-28T11:07:42.503000", "content": "\n\nAll improvements come from ensembling models and its getting quite boring..
\n
I think one could achieve +0.910 with single model before the end.
\nThe competion just begins. There are still almost 3 months to go.
\nEdit 07/12/2020: I made this comment before the LB was updated. The current best LB after update (+0.905) would like be +0.910 before update.
\nMy new prediction for best single model before the end is +0.908
", "votes": 2, "replies": [] }, { "id": 1094156, "author_name": "Wang Xinliang", "author_url": "", "post_date": "2020-11-28T11:11:47.417000", "content": "'single model' means only one fold or k-folds?
", "votes": 0, "replies": [] }, { "id": 1094405, "author_name": "Serigne ", "author_url": "", "post_date": "2020-11-28T15:38:22.573000", "content": "one or k-folds whatever, but most importantly improvements from data (and external data) processing, model architecture and training pipleline.
", "votes": 1, "replies": [] }, { "id": 1094442, "author_name": "gao-hongnan", "author_url": "", "post_date": "2020-11-28T16:06:29.147000", "content": "It’s kinda weird the scores are all stagnant at 0.90x
", "votes": 0, "replies": [] }, { "id": 1095070, "author_name": "DeepUnderstanding", "author_url": "", "post_date": "2020-11-29T08:16:46.083000", "content": "@reighns do you think this is because of some labelling issue where almost everyone's model is failing to predict
", "votes": 2, "replies": [] }, { "id": 1109666, "author_name": "", "author_url": "", "post_date": "2020-12-12T00:02:35.923000", "content": "", "votes": 0, "replies": [] } ] }, { "id": 1096131, "author_name": "Serigne ", "author_url": "", "post_date": "2020-11-30T08:31:06.477000", "content": "EffNet B4
\n5 folds
\nCV 0.901
\nLB 0.905 (No TTA)
Hi buddy, CV seems to be consistent with LB?
", "votes": 0, "replies": [] }, { "id": 1096548, "author_name": "Serigne ", "author_url": "", "post_date": "2020-11-30T15:35:05.913000", "content": "Yeah it seems stable and pretty much aligned with the LB after some modification of my training pipeline
\nGlad to see you here :)
", "votes": 0, "replies": [] }, { "id": 1167316, "author_name": "Jaideep", "author_url": "", "post_date": "2021-01-24T07:12:57.067000", "content": "@serigne what is img size you using 512 ?
", "votes": 0, "replies": [] } ] }, { "id": 1169348, "author_name": "FGPC", "author_url": "", "post_date": "2021-01-25T12:59:19.937000", "content": "seresnext101_32x4d
\n512x512
\n5-folds
\nBest single Fold: 0.905
\nCV: 0.901; LB: 0.902
best single model score .90 and ensemble 90.6 effnet and others
", "votes": 0, "replies": [ { "id": 1168324, "author_name": "ayu055", "author_url": "", "post_date": "2021-01-24T21:51:36.577000", "content": "Could you provide more details on the ensemble?
", "votes": 0, "replies": [] }, { "id": 1168325, "author_name": "Alexey Pronin", "author_url": "", "post_date": "2021-01-24T21:58:41.307000", "content": "@jaideepvalani Are these your CV or LB scores?
", "votes": 0, "replies": [] }, { "id": 1168973, "author_name": "Jaideep", "author_url": "", "post_date": "2021-01-25T09:17:41.217000", "content": "Lbs, cvs are either same or less by .003 to .004
", "votes": 1, "replies": [] }, { "id": 1171524, "author_name": "Debarshi Chanda", "author_url": "", "post_date": "2021-01-26T22:44:20.093000", "content": "@jaideepvalani Can you tell the type of ensembling are you using? Is it simple averaging or something complex?
", "votes": 0, "replies": [] } ] }, { "id": 1100462, "author_name": "zhangeng", "author_url": "", "post_date": "2020-12-03T05:16:05.217000", "content": "vit cv 90+ lb87+
", "votes": 0, "replies": [] }, { "id": 1095244, "author_name": "Heroseo", "author_url": "", "post_date": "2020-11-29T11:54:46.663000", "content": "resnext50
\n5-Folds
\nno TTA
CV 0.893
\nLB 0.902
resnext50
\n5-Folds
\nno TTA
CV 0.895
\nLB 0.903
Has anyone experimented with sizes? bigger the sizes better the results? that is what's the relation of LB and size of the image
", "votes": 0, "replies": [ { "id": 1095155, "author_name": "Harshit Sheoran", "author_url": "", "post_date": "2020-11-29T09:54:36.360000", "content": "From my experiments after a particular resolution model will stop getting better. 512x512 is an overkill the same or better results can be made with a lower resolution.
", "votes": 1, "replies": [] }, { "id": 1102941, "author_name": "vikram reddy", "author_url": "", "post_date": "2020-12-05T13:51:44.390000", "content": "The efficientnet paper has this to say:
Looking from all the results CV/LB relation is very close to 1; (still, can't trust what happens in private data), do think same imbalance is there in private
", "votes": 0, "replies": [] }, { "id": 1088650, "author_name": "Oleg Khlevnov", "author_url": "", "post_date": "2020-11-23T20:35:24.870000", "content": "Sorry, I am a novice here :) What do you mean by single fold?
Typically you hold out 20% of the train data for a local validation dataset.
\nYou vary the 20%, so you have 5 different combinations of train data and validation data. Each one is called a \"fold\". \"Stratefied\" folds attempt to distribute the train data evenly, so you don't have one train dataset with all the class-3 data and another with all the \"class-1\" data.
\nFor your local cross validation, you combine the results of your five \"folds\" to get a local cross-validation (CV) score.
\nThen you run the test data against your five models and combine the scores. That gives you a Leaderboard score.
\nThe hope is that the five \"folds\" will average out any errors and be more accurate than one fold or just using the leaderboard score as a guide. This is usually true, and you have to learn to \"trust\" your cross-validation, rather than be guided by the leaderboard score (which is usually on a small subset of the real test data).
\nThere is nothing magic about 20% and 5 folds. Just tradition. You could do fewer folds or more, depending on your data set size and your processing time.
\nSo, the final answer - \"single fold\" is when somebody has not yet done multiple training runs, so they only have one validation dataset. Since it takes typically 5 times as long to run five folds, you typically try and get parameters close with a single fold and then start running five-folds. Also depends on how long a fold takes.
\n-Rich
", "votes": 12, "replies": [] }, { "id": 1088711, "author_name": "Oleg Khlevnov", "author_url": "", "post_date": "2020-11-23T21:54:04.760000", "content": "Thank you very match for description! As I understand, single fold it the same thing as hold-out CV. Hope I can beat something serious and will write it here later)
", "votes": 1, "replies": [] } ] }, { "id": 1087598, "author_name": "Tahsin Mostafiz", "author_url": "", "post_date": "2020-11-22T22:26:19.037000", "content": "Model: fast-resnest50
\nsingle fold
\nAug: rotation, flip, brightness, contrast, cutout
\nTTA: 0
\nCV: 0.8715
\nLB: 0.8800
I believe I can push my CV/LB even further with Fast-Resnest.
", "votes": 0, "replies": [] }, { "id": 1087530, "author_name": "Yann Majewski", "author_url": "", "post_date": "2020-11-22T20:10:58.467000", "content": "Resnest50
\nsingle fold
\nCV: 0.9009
\nLB: 0.890
my lb is always lower than cv, should probably test other folds
", "votes": 0, "replies": [] }, { "id": 1086702, "author_name": "wakame", "author_url": "", "post_date": "2020-11-22T00:40:31.070000", "content": "resnest50d_4s2x40d
\n5 epochs
\nno TTA
\nCV .879
\nLB .892