{ "id": 253385, "title": "CV vs LB (New Dataset Only)", "url": "/competitions/seti-breakthrough-listen/discussion/253385", "author_name": "Nemuri", "post_date": "2021-07-16T10:27:00.390000", "votes": 44, "comment_count": 52, "views": 0, "content": "

ResNet18d 5Fold

\n

CV: 0.8243
\nLB 0.714

\n

It looks like a different competition than before the replacement.

", "messages": [ { "id": 1390058, "postDate": "2021-07-16T10:27:00.390Z", "content": "

ResNet18d 5Fold

\n

CV: 0.8243
\nLB 0.714

\n

It looks like a different competition than before the replacement.

", "rawMarkdown": "ResNet18d 5Fold\n\nCV: 0.8243\nLB 0.714\n\nIt looks like a different competition than before the replacement.", "votes": 44 }, { "id": 1392643, "postDate": "2021-07-18T23:43:41.160Z", "content": "

Efficientnet b0 5fold 10epochs
\ncv:0.8729617
\nlb:0.774

\n

more 5epcoh
\ncv:0.87358338425
\nlb:0.777😊

\n

b3 10epochs
\ncv: 0.8829833
\nlb:0.784

\n

b3 20epochs
\ncv:0.8876651
\nlb:0.791 😄

", "rawMarkdown": "Efficientnet b0 5fold 10epochs\ncv:0.8729617\nlb:0.774\n\nmore 5epcoh\ncv:0.87358338425\nlb:0.777😊\n\nb3 10epochs\ncv: 0.8829833\nlb:0.784\n\nb3 20epochs\ncv:0.8876651\nlb:0.791 😄", "votes": 12, "replies": [ { "id": 1394896, "postDate": "2021-07-20T17:47:53.467Z", "content": "

I see everyone is training for more than 15 epochs. But in my case, the valid loss starts to increase from epoch 8-9 and never ever decrease after that 😪. I performed many experiments trying different schedulers and lrs but still with same results. Can you share what scheduler you used. Ignore if you don't want to share. Thanks.
\nI tried experimenting with efficientnet_b0 and efficientnet_b4

", "rawMarkdown": "I see everyone is training for more than 15 epochs. But in my case, the valid loss starts to increase from epoch 8-9 and never ever decrease after that 😪. I performed many experiments trying different schedulers and lrs but still with same results. Can you share what scheduler you used. Ignore if you don't want to share. Thanks.\nI tried experimenting with `efficientnet_b0` and `efficientnet_b4`", "votes": 2 }, { "id": 1395079, "postDate": "2021-07-20T22:08:08.983Z", "content": "

I only use CosineAnnealingLR because i can't handle others…

", "rawMarkdown": "I only use CosineAnnealingLR because i can't handle others...", "votes": 2 }, { "id": 1395323, "postDate": "2021-07-21T06:19:36.043Z", "content": "

Thanks for sharing your approach @abebe9849. Do you use mixup?

", "rawMarkdown": "Thanks for sharing your approach @abebe9849. Do you use mixup? " }, { "id": 1400514, "postDate": "2021-07-26T11:06:39.690Z", "rawMarkdown": "", "isDeleted": true }, { "id": 1401285, "postDate": "2021-07-27T07:18:04.840Z", "content": "

i use it from the beginning.

", "rawMarkdown": "i use it from the beginning.", "votes": 1 }, { "id": 1403454, "postDate": "2021-07-29T06:12:20.497Z", "rawMarkdown": "", "votes": -4, "isDeleted": true }, { "id": 1403957, "postDate": "2021-07-29T13:01:50.540Z", "content": "

Exploring augmentation is also fun for the competition. I think it's senseless to ask that question.

", "rawMarkdown": "Exploring augmentation is also fun for the competition. I think it's senseless to ask that question.", "votes": 1 }, { "id": 1403975, "postDate": "2021-07-29T13:15:15.647Z", "content": "

Did you just change the “batch size” or “input size” when you use b3 instead of b0? I mean whether we need to change other places when we adopt a different model. Could you give me some pointers? Thanks in advance.

", "rawMarkdown": "Did you just change the “batch size” or “input size” when you use b3 instead of b0? I mean whether we need to change other places when we adopt a different model. Could you give me some pointers? Thanks in advance." }, { "id": 1404011, "postDate": "2021-07-29T13:47:08.160Z", "content": "

I use q rtx 8000 ,so i don't care about batch size.

", "rawMarkdown": "I use q rtx 8000 ,so i don't care about batch size.", "votes": 3 }, { "id": 1404361, "postDate": "2021-07-29T19:22:16.197Z", "content": "

@patriot How to optimize tmax for cosigneannealingLR?

", "rawMarkdown": "@patriot How to optimize tmax for cosigneannealingLR?", "votes": -1 }, { "id": 1404659, "postDate": "2021-07-30T06:23:50.547Z", "content": "

Thanks for sharing @abebe9849. What image size do you use?

", "rawMarkdown": "Thanks for sharing @abebe9849. What image size do you use?", "votes": -1 } ] }, { "id": 1456555, "postDate": "2021-08-07T01:28:30.103Z", "content": "

I think some people are asking too much.
\nThe purpose of discussion is to clarify general tendency of CV and LB discrepancy, not to asking precise information about high score models, isn’t it?

", "rawMarkdown": "I think some people are asking too much.\nThe purpose of discussion is to clarify general tendency of CV and LB discrepancy, not to asking precise information about high score models, isn’t it?", "votes": 8 }, { "id": 1473030, "postDate": "2021-08-15T09:52:48.697Z", "content": "

Solo Model (Single Fold): CV 0.902 LB 0.791

", "rawMarkdown": "Solo Model (Single Fold): CV 0.902 LB 0.791", "votes": 6, "replies": [ { "id": 1473111, "postDate": "2021-08-15T11:06:46.517Z", "content": "

That's really a great result @ks2019 can you tell us the model you have used for this result?

", "rawMarkdown": "That's really a great result @ks2019 can you tell us the model you have used for this result?" }, { "id": 1473266, "postDate": "2021-08-15T13:21:11.517Z", "content": "

EfficientNet B5. Really sorry, I can't share more details at this end of the competition.

", "rawMarkdown": "EfficientNet B5. Really sorry, I can't share more details at this end of the competition. ", "votes": 5 }, { "id": 1473272, "postDate": "2021-08-15T13:30:06.777Z", "content": "

Thanks for sharing, I'd given up hope on converging larger models but might have to try again.. =]

", "rawMarkdown": "Thanks for sharing, I'd given up hope on converging larger models but might have to try again.. =]", "votes": 1 }, { "id": 1473629, "postDate": "2021-08-15T16:34:14.453Z", "content": "

@ks2019 no worries all I wanted was the model name , thanks for that 😉

", "rawMarkdown": "@ks2019 no worries all I wanted was the model name , thanks for that 😉", "votes": 1 } ] }, { "id": 1390828, "postDate": "2021-07-17T05:35:33.247Z", "content": "

Validation: 0.8889
\nLB: 0.762

\n

Single fold efficientnet, Colab Pro, no old data used

", "rawMarkdown": "Validation: 0.8889\nLB: 0.762\n\nSingle fold efficientnet, Colab Pro, no old data used", "votes": 5, "replies": [ { "id": 1391365, "postDate": "2021-07-17T14:26:50.617Z", "content": "

@nyanpn : I tried using kaggle api in colab and could download only entire data. Is there any way to download only train and test of new data. Your inputs will be highly appreciated.

", "rawMarkdown": "@nyanpn : I tried using kaggle api in colab and could download only entire data. Is there any way to download only train and test of new data. Your inputs will be highly appreciated. " }, { "id": 1391471, "postDate": "2021-07-17T16:24:23.717Z", "content": "

I download whole data locally, then upload only the training data to google drive and mount it from colab.

", "rawMarkdown": "I download whole data locally, then upload only the training data to google drive and mount it from colab." }, { "id": 1391831, "postDate": "2021-07-18T06:06:39.020Z", "content": "

@nyanpn May I ask for how many epochs are you training?
\nThanks.

", "rawMarkdown": "@nyanpn May I ask for how many epochs are you training?\nThanks." }, { "id": 1391884, "postDate": "2021-07-18T07:26:01.227Z", "content": "

CV: 0.8882
\nLB: 0.768

\n

1fold(5fold) efficientnetb4 no old data used

", "rawMarkdown": "CV: 0.8882\nLB: 0.768\n\n1fold(5fold) efficientnetb4 no old data used", "votes": 2 }, { "id": 1391946, "postDate": "2021-07-18T08:42:41.500Z", "content": "

@atharvaingle 20epoch, not converged.

", "rawMarkdown": "@atharvaingle 20epoch, not converged.", "votes": 2 }, { "id": 1455795, "postDate": "2021-08-06T17:26:36.010Z", "content": "

Which augmentations you have used?

", "rawMarkdown": "Which augmentations you have used?\n", "votes": -1 } ] }, { "id": 1399735, "postDate": "2021-07-25T16:13:28.370Z", "content": "

ResNet34-D (5folds CV)

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
foldValidationPublic LB
00.88880.761
10.87550.768
20.87440.763
30.88250.765
40.87790.758
OOF0.87210.771(5fold-avg)
", "rawMarkdown": "ResNet34-D (5folds CV)\n\n| fold | Validation | Public LB |\n|:----:|:----------:|:---------:|\n| 0 | 0.8888 | 0.761 |\n| 1 | 0.8755 | 0.768 |\n| 2 | 0.8744 | 0.763 |\n| 3 | 0.8825 | 0.765 |\n| 4 | 0.8779 | 0.758 |\n| OOF | 0.8721 | 0.771(5fold-avg) |", "votes": 6, "replies": [ { "id": 1400684, "postDate": "2021-07-26T14:03:58.300Z", "content": "

do you use mixup, image size?

", "rawMarkdown": "do you use mixup, image size?", "votes": -4 } ] }, { "id": 1463051, "postDate": "2021-08-10T04:35:20.220Z", "content": "

EfficientNet-b0 val 0.8597 LB 0.760
\nIt is difficult for breaking 0.880 for val.

", "rawMarkdown": "EfficientNet-b0 val 0.8597 LB 0.760\nIt is difficult for breaking 0.880 for val.", "votes": 3, "replies": [ { "id": 1472239, "postDate": "2021-08-14T18:16:56.977Z", "content": "

Agreed. I'm finding models larger than b1 much harder to get to converge… and best solo model is 0.8791 CV / 0.772 LB, one fold.

", "rawMarkdown": "Agreed. I'm finding models larger than b1 much harder to get to converge... and best solo model is 0.8791 CV / 0.772 LB, one fold.", "votes": 1 } ] }, { "id": 1394875, "postDate": "2021-07-20T17:34:26.520Z", "content": "

Single model CV 0.880, LB 0.787. So far LB and CV are correlated for me.

", "rawMarkdown": "Single model CV 0.880, LB 0.787. So far LB and CV are correlated for me.", "votes": 4, "replies": [ { "id": 1395198, "postDate": "2021-07-21T02:56:13.360Z", "content": "

are you used adamw or adam? how many epochs

", "rawMarkdown": "are you used adamw or adam? how many epochs ", "votes": -9 } ] }, { "id": 1456472, "postDate": "2021-08-06T23:31:34.687Z", "content": "

solo model
\n5folds CV: 0.8770
\nPublic LB: 0.774

\n

As others have mentioned, Public LB and my local CV are correlated.

", "rawMarkdown": "solo model\n5folds CV: 0.8770\nPublic LB: 0.774\n\nAs others have mentioned, Public LB and my local CV are correlated.", "votes": 1 }, { "id": 1396960, "postDate": "2021-07-22T16:10:36.140Z", "content": "

resnet18d, single fold
\nCV:0.841
\nLB:0.725

", "rawMarkdown": "resnet18d, single fold\nCV:0.841\nLB:0.725", "votes": 1 }, { "id": 1396938, "postDate": "2021-07-22T15:43:19.180Z", "content": "

Efnetb0 - new data only - 4 folds
\nCV: 0.8732
\nLB: 0.769

", "rawMarkdown": "Efnetb0 - new data only - 4 folds\nCV: 0.8732\nLB: 0.769", "votes": 1, "replies": [ { "id": 1399207, "postDate": "2021-07-25T02:55:38.103Z", "content": "

Is it convenient to share the image size and epoch? Thanks.

", "rawMarkdown": "Is it convenient to share the image size and epoch? Thanks.\n" }, { "id": 1399389, "postDate": "2021-07-25T09:08:38.780Z", "content": "

512, 15 epochs

", "rawMarkdown": "512, 15 epochs", "votes": 2 }, { "id": 1400294, "postDate": "2021-07-26T07:15:29.573Z", "content": "

Thanks,bro.

", "rawMarkdown": "Thanks,bro." } ] }, { "id": 1395493, "postDate": "2021-07-21T09:06:43.833Z", "content": "

Solo model, new data only. CV 0.862 (5 fold). LB 767.

", "rawMarkdown": "Solo model, new data only. CV 0.862 (5 fold). LB 767.", "votes": 1, "replies": [ { "id": 1395514, "postDate": "2021-07-21T09:22:20.993Z", "content": "

what is the solo model?

", "rawMarkdown": "what is the solo model?" }, { "id": 1396176, "postDate": "2021-07-21T21:27:35.053Z", "content": "

Model is Nfnet.

", "rawMarkdown": "Model is Nfnet." }, { "id": 1398700, "postDate": "2021-07-24T12:09:19.390Z", "content": "

What image size do you use?

", "rawMarkdown": "What image size do you use?", "votes": 1 }, { "id": 1398756, "postDate": "2021-07-24T13:15:29.387Z", "content": "

The size of input images is 256x256x3.

", "rawMarkdown": "The size of input images is 256x256x3.", "votes": 5 } ] }, { "id": 1394016, "postDate": "2021-07-20T05:33:06.740Z", "content": "

Efficientnet b4 5fold 15epochs
\ncv:0.8585
\nlb:0.765

", "rawMarkdown": "Efficientnet b4 5fold 15epochs\ncv:0.8585\nlb:0.765", "votes": 1 }, { "id": 1390265, "postDate": "2021-07-16T13:59:44.417Z", "content": "

EfficientNetB5

\n

CV:0.861
\nLB 0.751

", "rawMarkdown": "EfficientNetB5\n\nCV:0.861\nLB 0.751\n", "votes": 1 }, { "id": 1392035, "postDate": "2021-07-18T10:05:45.370Z", "content": "

Efficientnet b0 3fold
\nCV 0.7924 < = > LB 0.699 ; Condition 1
\nCV 0.8090 < = > LB 0.710 ; Condition 2
\nCV 0.8150 < = > LB 0.713 ; Condition 3

\n

LB seems to have about 0.1 lower value than CV

", "rawMarkdown": "Efficientnet b0 3fold\nCV 0.7924 < = > LB 0.699 ; Condition 1\nCV 0.8090 < = > LB 0.710 ; Condition 2\nCV 0.8150 < = > LB 0.713 ; Condition 3 \n\nLB seems to have about 0.1 lower value than CV", "votes": 2 }, { "id": 1390968, "postDate": "2021-07-17T07:53:06.153Z", "content": "

Effnet b0 single fold
\nval: 0.8801
\nLb: 0.754

", "rawMarkdown": "Effnet b0 single fold\nval: 0.8801\nLb: 0.754", "votes": 2 }, { "id": 1390962, "postDate": "2021-07-17T07:44:27.280Z", "content": "

val: 0.865
\nLB: 0.753
\nsingle fold effcientnet-b0

", "rawMarkdown": "val: 0.865\nLB: 0.753\nsingle fold effcientnet-b0", "votes": 2 }, { "id": 1390431, "postDate": "2021-07-16T16:29:45.517Z", "content": "

efficientnetb0 5fold
\ncv: 0.85607
\nlb: 0.747

", "rawMarkdown": "efficientnetb0 5fold\ncv: 0.85607\nlb: 0.747", "votes": 2 }, { "id": 1390181, "postDate": "2021-07-16T12:32:20.630Z", "content": "

Efficientnet b0: cv 0.876, lb 0.756

", "rawMarkdown": "Efficientnet b0: cv 0.876, lb 0.756", "votes": 2 }, { "id": 1390090, "postDate": "2021-07-16T10:49:55.930Z", "content": "

Same here with EfficientNet B3 (5-fold):

\n", "rawMarkdown": "Same here with EfficientNet B3 (5-fold):\n- CV: 0.821\n- LB: 0.717", "votes": 2 }, { "id": 1464030, "postDate": "2021-08-10T12:27:34.643Z", "content": "

EfficientNet B1 (5 Fold) CV / LB: 0.854 / 0.757

", "rawMarkdown": "EfficientNet B1 (5 Fold) CV / LB: 0.854 / 0.757 " }, { "id": 1443113, "postDate": "2021-08-04T01:49:02.457Z", "content": "

resnet18d, 5 fold, 320x320, 40epoch
\nCV: 0.840
\nLB: 0.720

", "rawMarkdown": "resnet18d, 5 fold, 320x320, 40epoch\nCV: 0.840\nLB: 0.720" } ], "comments": [ { "id": 1392643, "author_name": "patriot", "author_url": "", "post_date": "2021-07-18T23:43:41.160000", "content": "

Efficientnet b0 5fold 10epochs
\ncv:0.8729617
\nlb:0.774

\n

more 5epcoh
\ncv:0.87358338425
\nlb:0.777😊

\n

b3 10epochs
\ncv: 0.8829833
\nlb:0.784

\n

b3 20epochs
\ncv:0.8876651
\nlb:0.791 😄

", "votes": 12, "replies": [ { "id": 1394896, "author_name": "Atharva Ingle", "author_url": "", "post_date": "2021-07-20T17:47:53.467000", "content": "

I see everyone is training for more than 15 epochs. But in my case, the valid loss starts to increase from epoch 8-9 and never ever decrease after that 😪. I performed many experiments trying different schedulers and lrs but still with same results. Can you share what scheduler you used. Ignore if you don't want to share. Thanks.
\nI tried experimenting with efficientnet_b0 and efficientnet_b4

", "votes": 2, "replies": [] }, { "id": 1395079, "author_name": "patriot", "author_url": "", "post_date": "2021-07-20T22:08:08.983000", "content": "

I only use CosineAnnealingLR because i can't handle others…

", "votes": 2, "replies": [] }, { "id": 1395323, "author_name": "_lev_lipinski", "author_url": "", "post_date": "2021-07-21T06:19:36.043000", "content": "

Thanks for sharing your approach @abebe9849. Do you use mixup?

", "votes": 0, "replies": [] }, { "id": 1400514, "author_name": "", "author_url": "", "post_date": "2021-07-26T11:06:39.690000", "content": "", "votes": 0, "replies": [] }, { "id": 1401285, "author_name": "patriot", "author_url": "", "post_date": "2021-07-27T07:18:04.840000", "content": "

i use it from the beginning.

", "votes": 1, "replies": [] }, { "id": 1403454, "author_name": "", "author_url": "", "post_date": "2021-07-29T06:12:20.497000", "content": "", "votes": -4, "replies": [] }, { "id": 1403957, "author_name": "patriot", "author_url": "", "post_date": "2021-07-29T13:01:50.540000", "content": "

Exploring augmentation is also fun for the competition. I think it's senseless to ask that question.

", "votes": 1, "replies": [] }, { "id": 1403975, "author_name": "Gainover", "author_url": "", "post_date": "2021-07-29T13:15:15.647000", "content": "

Did you just change the “batch size” or “input size” when you use b3 instead of b0? I mean whether we need to change other places when we adopt a different model. Could you give me some pointers? Thanks in advance.

", "votes": 0, "replies": [] }, { "id": 1404011, "author_name": "patriot", "author_url": "", "post_date": "2021-07-29T13:47:08.160000", "content": "

I use q rtx 8000 ,so i don't care about batch size.

", "votes": 3, "replies": [] }, { "id": 1404361, "author_name": "Eduardo Farina", "author_url": "", "post_date": "2021-07-29T19:22:16.197000", "content": "

@patriot How to optimize tmax for cosigneannealingLR?

", "votes": -1, "replies": [] }, { "id": 1404659, "author_name": "ld", "author_url": "", "post_date": "2021-07-30T06:23:50.547000", "content": "

Thanks for sharing @abebe9849. What image size do you use?

", "votes": -1, "replies": [] } ] }, { "id": 1456555, "author_name": "Bilzard", "author_url": "", "post_date": "2021-08-07T01:28:30.103000", "content": "

I think some people are asking too much.
\nThe purpose of discussion is to clarify general tendency of CV and LB discrepancy, not to asking precise information about high score models, isn’t it?

", "votes": 8, "replies": [] }, { "id": 1473030, "author_name": "Kumar Shubham", "author_url": "", "post_date": "2021-08-15T09:52:48.697000", "content": "

Solo Model (Single Fold): CV 0.902 LB 0.791

", "votes": 6, "replies": [ { "id": 1473111, "author_name": "Mr_KnowNothing", "author_url": "", "post_date": "2021-08-15T11:06:46.517000", "content": "

That's really a great result @ks2019 can you tell us the model you have used for this result?

", "votes": 0, "replies": [] }, { "id": 1473266, "author_name": "Kumar Shubham", "author_url": "", "post_date": "2021-08-15T13:21:11.517000", "content": "

EfficientNet B5. Really sorry, I can't share more details at this end of the competition.

", "votes": 5, "replies": [] }, { "id": 1473272, "author_name": "عثمان", "author_url": "", "post_date": "2021-08-15T13:30:06.777000", "content": "

Thanks for sharing, I'd given up hope on converging larger models but might have to try again.. =]

", "votes": 1, "replies": [] }, { "id": 1473629, "author_name": "Mr_KnowNothing", "author_url": "", "post_date": "2021-08-15T16:34:14.453000", "content": "

@ks2019 no worries all I wanted was the model name , thanks for that 😉

", "votes": 1, "replies": [] } ] }, { "id": 1390828, "author_name": "nyanp", "author_url": "", "post_date": "2021-07-17T05:35:33.247000", "content": "

Validation: 0.8889
\nLB: 0.762

\n

Single fold efficientnet, Colab Pro, no old data used

", "votes": 5, "replies": [ { "id": 1391365, "author_name": "Manoj Prabhakar", "author_url": "", "post_date": "2021-07-17T14:26:50.617000", "content": "

@nyanpn : I tried using kaggle api in colab and could download only entire data. Is there any way to download only train and test of new data. Your inputs will be highly appreciated.

", "votes": 0, "replies": [] }, { "id": 1391471, "author_name": "nyanp", "author_url": "", "post_date": "2021-07-17T16:24:23.717000", "content": "

I download whole data locally, then upload only the training data to google drive and mount it from colab.

", "votes": 0, "replies": [] }, { "id": 1391831, "author_name": "Atharva Ingle", "author_url": "", "post_date": "2021-07-18T06:06:39.020000", "content": "

@nyanpn May I ask for how many epochs are you training?
\nThanks.

", "votes": 0, "replies": [] }, { "id": 1391884, "author_name": "zhangeng", "author_url": "", "post_date": "2021-07-18T07:26:01.227000", "content": "

CV: 0.8882
\nLB: 0.768

\n

1fold(5fold) efficientnetb4 no old data used

", "votes": 2, "replies": [] }, { "id": 1391946, "author_name": "nyanp", "author_url": "", "post_date": "2021-07-18T08:42:41.500000", "content": "

@atharvaingle 20epoch, not converged.

", "votes": 2, "replies": [] }, { "id": 1455795, "author_name": "Vatsal Mavani", "author_url": "", "post_date": "2021-08-06T17:26:36.010000", "content": "

Which augmentations you have used?

", "votes": -1, "replies": [] } ] }, { "id": 1399735, "author_name": "Tawara", "author_url": "", "post_date": "2021-07-25T16:13:28.370000", "content": "

ResNet34-D (5folds CV)

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
foldValidationPublic LB
00.88880.761
10.87550.768
20.87440.763
30.88250.765
40.87790.758
OOF0.87210.771(5fold-avg)
", "votes": 6, "replies": [ { "id": 1400684, "author_name": "Eevee", "author_url": "", "post_date": "2021-07-26T14:03:58.300000", "content": "

do you use mixup, image size?

", "votes": -4, "replies": [] } ] }, { "id": 1463051, "author_name": "imori", "author_url": "", "post_date": "2021-08-10T04:35:20.220000", "content": "

EfficientNet-b0 val 0.8597 LB 0.760
\nIt is difficult for breaking 0.880 for val.

", "votes": 3, "replies": [ { "id": 1472239, "author_name": "عثمان", "author_url": "", "post_date": "2021-08-14T18:16:56.977000", "content": "

Agreed. I'm finding models larger than b1 much harder to get to converge… and best solo model is 0.8791 CV / 0.772 LB, one fold.

", "votes": 1, "replies": [] } ] }, { "id": 1394875, "author_name": "sin", "author_url": "", "post_date": "2021-07-20T17:34:26.520000", "content": "

Single model CV 0.880, LB 0.787. So far LB and CV are correlated for me.

", "votes": 4, "replies": [ { "id": 1395198, "author_name": "Eevee", "author_url": "", "post_date": "2021-07-21T02:56:13.360000", "content": "

are you used adamw or adam? how many epochs

", "votes": -9, "replies": [] } ] }, { "id": 1456472, "author_name": "hayashida", "author_url": "", "post_date": "2021-08-06T23:31:34.687000", "content": "

solo model
\n5folds CV: 0.8770
\nPublic LB: 0.774

\n

As others have mentioned, Public LB and my local CV are correlated.

", "votes": 1, "replies": [] }, { "id": 1396960, "author_name": "mizoo", "author_url": "", "post_date": "2021-07-22T16:10:36.140000", "content": "

resnet18d, single fold
\nCV:0.841
\nLB:0.725

", "votes": 1, "replies": [] }, { "id": 1396938, "author_name": "Ed Yanakov", "author_url": "", "post_date": "2021-07-22T15:43:19.180000", "content": "

Efnetb0 - new data only - 4 folds
\nCV: 0.8732
\nLB: 0.769

", "votes": 1, "replies": [ { "id": 1399207, "author_name": "Gainover", "author_url": "", "post_date": "2021-07-25T02:55:38.103000", "content": "

Is it convenient to share the image size and epoch? Thanks.

", "votes": 0, "replies": [] }, { "id": 1399389, "author_name": "Ed Yanakov", "author_url": "", "post_date": "2021-07-25T09:08:38.780000", "content": "

512, 15 epochs

", "votes": 2, "replies": [] }, { "id": 1400294, "author_name": "Gainover", "author_url": "", "post_date": "2021-07-26T07:15:29.573000", "content": "

Thanks,bro.

", "votes": 0, "replies": [] } ] }, { "id": 1395493, "author_name": "Sergey Bryansky", "author_url": "", "post_date": "2021-07-21T09:06:43.833000", "content": "

Solo model, new data only. CV 0.862 (5 fold). LB 767.

", "votes": 1, "replies": [ { "id": 1395514, "author_name": "Eevee", "author_url": "", "post_date": "2021-07-21T09:22:20.993000", "content": "

what is the solo model?

", "votes": 0, "replies": [] }, { "id": 1396176, "author_name": "Sergey Bryansky", "author_url": "", "post_date": "2021-07-21T21:27:35.053000", "content": "

Model is Nfnet.

", "votes": 0, "replies": [] }, { "id": 1398700, "author_name": "Debarshi Chanda", "author_url": "", "post_date": "2021-07-24T12:09:19.390000", "content": "

What image size do you use?

", "votes": 1, "replies": [] }, { "id": 1398756, "author_name": "Sergey Bryansky", "author_url": "", "post_date": "2021-07-24T13:15:29.387000", "content": "

The size of input images is 256x256x3.

", "votes": 5, "replies": [] } ] }, { "id": 1394016, "author_name": "Ctrl_CV", "author_url": "", "post_date": "2021-07-20T05:33:06.740000", "content": "

Efficientnet b4 5fold 15epochs
\ncv:0.8585
\nlb:0.765

", "votes": 1, "replies": [] }, { "id": 1390265, "author_name": "Manoj Prabhakar", "author_url": "", "post_date": "2021-07-16T13:59:44.417000", "content": "

EfficientNetB5

\n

CV:0.861
\nLB 0.751

", "votes": 1, "replies": [] }, { "id": 1392035, "author_name": "Beluga", "author_url": "", "post_date": "2021-07-18T10:05:45.370000", "content": "

Efficientnet b0 3fold
\nCV 0.7924 < = > LB 0.699 ; Condition 1
\nCV 0.8090 < = > LB 0.710 ; Condition 2
\nCV 0.8150 < = > LB 0.713 ; Condition 3

\n

LB seems to have about 0.1 lower value than CV

", "votes": 2, "replies": [] }, { "id": 1390968, "author_name": "Debarshi Chanda", "author_url": "", "post_date": "2021-07-17T07:53:06.153000", "content": "

Effnet b0 single fold
\nval: 0.8801
\nLb: 0.754

", "votes": 2, "replies": [] }, { "id": 1390962, "author_name": "rincha", "author_url": "", "post_date": "2021-07-17T07:44:27.280000", "content": "

val: 0.865
\nLB: 0.753
\nsingle fold effcientnet-b0

", "votes": 2, "replies": [] }, { "id": 1390431, "author_name": "yabea", "author_url": "", "post_date": "2021-07-16T16:29:45.517000", "content": "

efficientnetb0 5fold
\ncv: 0.85607
\nlb: 0.747

", "votes": 2, "replies": [] }, { "id": 1390181, "author_name": "Eevee", "author_url": "", "post_date": "2021-07-16T12:32:20.630000", "content": "

Efficientnet b0: cv 0.876, lb 0.756

", "votes": 2, "replies": [] }, { "id": 1390090, "author_name": "Nikita Kozodoi", "author_url": "", "post_date": "2021-07-16T10:49:55.930000", "content": "

Same here with EfficientNet B3 (5-fold):

\n", "votes": 2, "replies": [] }, { "id": 1464030, "author_name": "Miltos", "author_url": "", "post_date": "2021-08-10T12:27:34.643000", "content": "

EfficientNet B1 (5 Fold) CV / LB: 0.854 / 0.757

", "votes": 0, "replies": [] }, { "id": 1443113, "author_name": "Bilzard", "author_url": "", "post_date": "2021-08-04T01:49:02.457000", "content": "

resnet18d, 5 fold, 320x320, 40epoch
\nCV: 0.840
\nLB: 0.720

", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1390058": "ResNet18d 5Fold\n\nCV: 0.8243\nLB 0.714\n\nIt looks like a different competition than before the replacement.", "1392643": "Efficientnet b0 5fold 10epochs\ncv:0.8729617\nlb:0.774\n\nmore 5epcoh\ncv:0.87358338425\nlb:0.777😊\n\nb3 10epochs\ncv: 0.8829833\nlb:0.784\n\nb3 20epochs\ncv:0.8876651\nlb:0.791 😄", "1456555": "I think some people are asking too much.\nThe purpose of discussion is to clarify general tendency of CV and LB discrepancy, not to asking precise information about high score models, isn’t it?", "1473030": "Solo Model (Single Fold): CV 0.902 LB 0.791", "1390828": "Validation: 0.8889\nLB: 0.762\n\nSingle fold efficientnet, Colab Pro, no old data used", "1399735": "ResNet34-D (5folds CV)\n\n| fold | Validation | Public LB |\n|:----:|:----------:|:---------:|\n| 0 | 0.8888 | 0.761 |\n| 1 | 0.8755 | 0.768 |\n| 2 | 0.8744 | 0.763 |\n| 3 | 0.8825 | 0.765 |\n| 4 | 0.8779 | 0.758 |\n| OOF | 0.8721 | 0.771(5fold-avg) |", "1463051": "EfficientNet-b0 val 0.8597 LB 0.760\nIt is difficult for breaking 0.880 for val.", "1394875": "Single model CV 0.880, LB 0.787. So far LB and CV are correlated for me.", "1456472": "solo model\n5folds CV: 0.8770\nPublic LB: 0.774\n\nAs others have mentioned, Public LB and my local CV are correlated.", "1396960": "resnet18d, single fold\nCV:0.841\nLB:0.725", "1396938": "Efnetb0 - new data only - 4 folds\nCV: 0.8732\nLB: 0.769", "1395493": "Solo model, new data only. CV 0.862 (5 fold). LB 767.", "1394016": "Efficientnet b4 5fold 15epochs\ncv:0.8585\nlb:0.765", "1390265": "EfficientNetB5\n\nCV:0.861\nLB 0.751\n", "1392035": "Efficientnet b0 3fold\nCV 0.7924 < = > LB 0.699 ; Condition 1\nCV 0.8090 < = > LB 0.710 ; Condition 2\nCV 0.8150 < = > LB 0.713 ; Condition 3 \n\nLB seems to have about 0.1 lower value than CV", "1390968": "Effnet b0 single fold\nval: 0.8801\nLb: 0.754", "1390962": "val: 0.865\nLB: 0.753\nsingle fold effcientnet-b0", "1390431": "efficientnetb0 5fold\ncv: 0.85607\nlb: 0.747", "1390181": "Efficientnet b0: cv 0.876, lb 0.756", "1390090": "Same here with EfficientNet B3 (5-fold):\n- CV: 0.821\n- LB: 0.717", "1464030": "EfficientNet B1 (5 Fold) CV / LB: 0.854 / 0.757 ", "1443113": "resnet18d, 5 fold, 320x320, 40epoch\nCV: 0.840\nLB: 0.720" } }