{ "id": 314135, "title": "【talktalk】What is your single model score?", "url": "/competitions/happy-whale-and-dolphin/discussion/314135", "author_name": "Yangranran", "post_date": "2022-03-21T06:49:08.535000", "votes": 19, "comment_count": 32, "views": 0, "content": "

Let's have a talk!

", "messages": [ { "id": 1730344, "postDate": "2022-03-21T06:49:08.537Z", "content": "

Let's have a talk!

", "rawMarkdown": "Let's have a talk!", "votes": 19 }, { "id": 1730588, "postDate": "2022-03-21T12:35:07.733Z", "content": "

effnet b6 768 , lb 809

", "rawMarkdown": "effnet b6 768 , lb 809", "votes": 10, "replies": [ { "id": 1730613, "postDate": "2022-03-21T13:00:03.150Z", "content": "

Outstanding!
\nWaiting for solution desctiption and good luck! πŸ‘ We are trying to close the gap between us … and still experimenting - looking for game changer.

", "rawMarkdown": "Outstanding! \nWaiting for solution desctiption and good luck! πŸ‘ We are trying to close the gap between us ... and still experimenting - looking for game changer. \n ", "votes": 2 }, { "id": 1731078, "postDate": "2022-03-22T01:28:34.843Z", "content": "

Great job! one fold or 5 folds?

", "rawMarkdown": "Great job! one fold or 5 folds?" }, { "id": 1731103, "postDate": "2022-03-22T01:49:24.713Z", "content": "

one fold. Random split.

", "rawMarkdown": "one fold. Random split.", "votes": 3 }, { "id": 1731286, "postDate": "2022-03-22T07:45:38.513Z", "content": "

What is your CV?

", "rawMarkdown": "What is your CV?" } ] }, { "id": 1730581, "postDate": "2022-03-21T12:20:18.837Z", "content": "

Our single is:

\n", "rawMarkdown": "Our single is:\n- CV .820 -> LB .775 (EffNetB6)", "votes": 5, "replies": [ { "id": 1730595, "postDate": "2022-03-21T12:44:30.690Z", "content": "

Nice! Could you give some hints on how you managed to reduce the drop from CV to LB?

", "rawMarkdown": "Nice! Could you give some hints on how you managed to reduce the drop from CV to LB? " }, { "id": 1730621, "postDate": "2022-03-21T13:06:44.180Z", "content": "

Most of improvements comes from dataset in our case. We went from 0.671 to 0.775 (for one fold) only changing DS but I am still not happy with solution. We have not found any biggest game changer so far.

", "rawMarkdown": "Most of improvements comes from dataset in our case. We went from 0.671 to 0.775 (for one fold) only changing DS but I am still not happy with solution. We have not found any biggest game changer so far. ", "votes": 2 }, { "id": 1731270, "postDate": "2022-03-22T07:16:16Z", "content": "

If our dataset has a lot of classes with just 1 example, then how do you guy split that in training and validation. I

", "rawMarkdown": "If our dataset has a lot of classes with just 1 example, then how do you guy split that in training and validation. I" } ] }, { "id": 1733610, "postDate": "2022-03-24T12:48:50.113Z", "content": "

MODEL / SIZE : EfficientNet7
\nSingle Fold, Tensorflow
\nCV : 0.830
\nLB : 0.800

", "rawMarkdown": "MODEL / SIZE : EfficientNet7\nSingle Fold, Tensorflow\nCV : 0.830\nLB : 0.800", "votes": 6 }, { "id": 1731547, "postDate": "2022-03-22T13:27:54.413Z", "content": "

UPDATE 03/27:
\nsingle fold
\neffnet-b6
\ncv=0.8257
\nlb=0.801

\n

OLD:
\nsingle fold
\neffnet-b5
\ncv=0.817
\nlb=0.779

", "rawMarkdown": "UPDATE 03/27:\nsingle fold\neffnet-b6\ncv=0.8257\nlb=0.801\n\nOLD:\nsingle fold\neffnet-b5\ncv=0.817\nlb=0.779\n", "votes": 3, "replies": [ { "id": 1732038, "postDate": "2022-03-23T00:27:28.447Z", "content": "

How do you split in training and validation? Random split? Some classes has just 1 example so you are not validating on those right?

", "rawMarkdown": "How do you split in training and validation? Random split? Some classes has just 1 example so you are not validating on those right?" }, { "id": 1736140, "postDate": "2022-03-27T03:20:20.313Z", "content": "

Random split

", "rawMarkdown": "Random split", "votes": 1 }, { "id": 1736145, "postDate": "2022-03-27T03:26:23.120Z", "content": "

You change model b5 to b6 and get so huge improvement, great!

", "rawMarkdown": "You change model b5 to b6 and get so huge improvement, great!", "votes": 1 } ] }, { "id": 1731122, "postDate": "2022-03-22T02:15:37.693Z", "content": "

effnet -b7, cv=0.815, lb=0.819 with 5 folds.

", "rawMarkdown": "effnet -b7, cv=0.815, lb=0.819 with 5 folds.", "votes": 3 }, { "id": 1733313, "postDate": "2022-03-24T07:14:50.943Z", "content": "

effnet-v2m, single fold, val. 8413, lb 809
\nconvnext-large, single fold, val. 8407, lb 805

", "rawMarkdown": "effnet-v2m, single fold, val. 8413, lb 809\nconvnext-large, single fold, val. 8407, lb 805 ", "votes": 4, "replies": [ { "id": 1733342, "postDate": "2022-03-24T07:35:41.267Z", "content": "

Very good score!! I do not ask how … but looking for solution description after competition. Looks your dataset provide great source of features to NN.

", "rawMarkdown": "Very good score!! I do not ask how … but looking for solution description after competition. Looks your dataset provide great source of features to NN." }, { "id": 1733356, "postDate": "2022-03-24T07:50:11.163Z", "content": "

wow…good job πŸ‘ An Effnet-v2m that has 809 lb is an incredible thing for me 😲. Could you give us some hints on image size and any other trick that you feel comfortable sharing? Also, I have trained Convnext too but only tried a small one because I want my image size to be at least 512, which turns out to be a long and painful journey. So, I am curious whether you reduce the image size while training a Convnext-large?

", "rawMarkdown": "wow...good job πŸ‘ An Effnet-v2m that has 809 lb is an incredible thing for me 😲. Could you give us some hints on image size and any other trick that you feel comfortable sharing? Also, I have trained Convnext too but only tried a small one because I want my image size to be at least 512, which turns out to be a long and painful journey. So, I am curious whether you reduce the image size while training a Convnext-large?", "votes": 2 }, { "id": 1733614, "postDate": "2022-03-24T12:49:40.997Z", "content": "

I tried convnext, efficientnetv2, but its score was lower than my efficientnet around 5 % when i used convnext + arcface

", "rawMarkdown": "I tried convnext, efficientnetv2, but its score was lower than my efficientnet around 5 % when i used convnext + arcface" }, { "id": 1734201, "postDate": "2022-03-25T04:53:08.587Z", "content": "

@remekkinas @qiuyanxinjupiter @deepkim
\nI can share the tweaks to achieve those scores after the competition ends πŸ˜„
\nYou don't need to train very long (>30 epochs) or at a very large image size (>768) though. Effnetv1 models (b6, b7) are better than effnetv2 and convnext models
\nWith post-processing, I think a single fold model can cross 0.84 on lb.

", "rawMarkdown": "@remekkinas @qiuyanxinjupiter @deepkim \nI can share the tweaks to achieve those scores after the competition ends πŸ˜„\nYou don't need to train very long (>30 epochs) or at a very large image size (>768) though. Effnetv1 models (b6, b7) are better than effnetv2 and convnext models\nWith post-processing, I think a single fold model can cross 0.84 on lb.", "votes": 8 }, { "id": 1734350, "postDate": "2022-03-25T08:32:31.547Z", "content": "
\n

I think a single fold model can cross 0.84 on lb.

\n
\n

@andy2709 I see you have made it, good job!

", "rawMarkdown": "> I think a single fold model can cross 0.84 on lb.\n\n@andy2709 I see you have made it, good job!", "votes": 1 }, { "id": 1734490, "postDate": "2022-03-25T11:51:25.177Z", "rawMarkdown": "", "isDeleted": true }, { "id": 1734497, "postDate": "2022-03-25T11:56:51.947Z", "content": "

@qiuyanxinjupiter οΌŒδ½ ιœ€θ¦ι˜Ÿε‹ε—οΌŸζˆ‘δ»¬δΈ€θ΅·η»„δΈͺ队？πŸ’ͺπŸ’ͺπŸ’ͺπŸ˜ƒ

", "rawMarkdown": "@qiuyanxinjupiter οΌŒδ½ ιœ€θ¦ι˜Ÿε‹ε—οΌŸζˆ‘δ»¬δΈ€θ΅·η»„δΈͺ队？πŸ’ͺπŸ’ͺπŸ’ͺπŸ˜ƒ" }, { "id": 1734634, "postDate": "2022-03-25T14:15:02.360Z", "content": "

@andy2709 than you for sharing. We came to the some conclutions. We have different solution in inference and data part (I suppose) which make difference. You have better one. We are not able to cross 0.8 from single fold so far. Looking for solution :)

", "rawMarkdown": "@andy2709 than you for sharing. We came to the some conclutions. We have different solution in inference and data part (I suppose) which make difference. You have better one. We are not able to cross 0.8 from single fold so far. Looking for solution :)", "votes": 1 } ] }, { "id": 1730387, "postDate": "2022-03-21T07:49:43.093Z", "content": "

LB 0.786 EffnetB7
\n1 fold
\n40 epochs

", "rawMarkdown": "LB 0.786 EffnetB7\n1 fold \n40 epochs ", "votes": 1, "replies": [ { "id": 1730394, "postDate": "2022-03-21T07:56:24.253Z", "content": "

Nice work!

", "rawMarkdown": "Nice work!", "votes": 1 }, { "id": 1730496, "postDate": "2022-03-21T10:40:46.747Z", "content": "

I am also using EffnetB7 with the detic crop dataset but my highest score is like 0.7. Could you share some thoughts about how you reach the 0.786 score? like dataset or matching technique for embedding? Thanks so much

", "rawMarkdown": "I am also using EffnetB7 with the detic crop dataset but my highest score is like 0.7. Could you share some thoughts about how you reach the 0.786 score? like dataset or matching technique for embedding? Thanks so much" }, { "id": 1730503, "postDate": "2022-03-21T11:04:27.410Z", "content": "

Lot of experiments were done on

\n
    \n
  1. Model architecture {Completely different from the ones that are made public}
    2. \n
  2. Changes in the augmentations
    3. \n
\n

My initial models were getting around ~0.7
\nAfter updating model structure and changing augmentation schemes, performance increased a lot

\n

Another thing is
\nFull Body annotation > Detic > Yolo

\n

I think the backfin dataset yields better results than all the above-mentioned datasets.
\nI believe that most top-scoring kernels are using that.

\n

I think my model performance would also increase with that dataset. I am planning to set up a dataset of my own as that's the best booster for this challenge

\n

Check further details at https://www.kaggle.com/c/happy-whale-and-dolphin/discussion/313987

", "rawMarkdown": "Lot of experiments were done on \n\n1. Model architecture {Completely different from the ones that are made public}\n2. Changes in the augmentations\n\nMy initial models were getting around ~0.7\nAfter updating model structure and changing augmentation schemes, performance increased a lot\n\nAnother thing is \nFull Body annotation > Detic > Yolo\n\nI think the backfin dataset yields better results than all the above-mentioned datasets. \nI believe that most top-scoring kernels are using that.\n\nI think my model performance would also increase with that dataset. I am planning to set up a dataset of my own as that's the best booster for this challenge \n\nCheck further details at https://www.kaggle.com/c/happy-whale-and-dolphin/discussion/313987", "votes": 9 }, { "id": 1731809, "postDate": "2022-03-22T17:53:15.993Z", "content": "

May I ask what image size are you using?

", "rawMarkdown": "May I ask what image size are you using?" }, { "id": 1732204, "postDate": "2022-03-23T05:54:42.983Z", "content": "

Image size 600

", "rawMarkdown": "Image size 600 " } ] }, { "id": 1737021, "postDate": "2022-03-28T03:43:02.363Z", "content": "

b5-384, .780 LB single fold

", "rawMarkdown": "b5-384, .780 LB single fold" }, { "id": 1736666, "postDate": "2022-03-27T15:27:51.053Z", "content": "

thanks everyone for insights

", "rawMarkdown": "thanks everyone for insights" } ], "comments": [ { "id": 1730588, "author_name": "lafe", "author_url": "", "post_date": "2022-03-21T12:35:07.733000", "content": "

effnet b6 768 , lb 809

", "votes": 10, "replies": [ { "id": 1730613, "author_name": "Remek Kinas", "author_url": "", "post_date": "2022-03-21T13:00:03.150000", "content": "

Outstanding!
\nWaiting for solution desctiption and good luck! πŸ‘ We are trying to close the gap between us … and still experimenting - looking for game changer.

", "votes": 2, "replies": [] }, { "id": 1731078, "author_name": "Roc", "author_url": "", "post_date": "2022-03-22T01:28:34.843000", "content": "

Great job! one fold or 5 folds?

", "votes": 0, "replies": [] }, { "id": 1731103, "author_name": "lafe", "author_url": "", "post_date": "2022-03-22T01:49:24.713000", "content": "

one fold. Random split.

", "votes": 3, "replies": [] }, { "id": 1731286, "author_name": "Salman Ahmed", "author_url": "", "post_date": "2022-03-22T07:45:38.513000", "content": "

What is your CV?

", "votes": 0, "replies": [] } ] }, { "id": 1730581, "author_name": "Remek Kinas", "author_url": "", "post_date": "2022-03-21T12:20:18.837000", "content": "

Our single is:

\n", "votes": 5, "replies": [ { "id": 1730595, "author_name": "clem-chris", "author_url": "", "post_date": "2022-03-21T12:44:30.690000", "content": "

Nice! Could you give some hints on how you managed to reduce the drop from CV to LB?

", "votes": 0, "replies": [] }, { "id": 1730621, "author_name": "Remek Kinas", "author_url": "", "post_date": "2022-03-21T13:06:44.180000", "content": "

Most of improvements comes from dataset in our case. We went from 0.671 to 0.775 (for one fold) only changing DS but I am still not happy with solution. We have not found any biggest game changer so far.

", "votes": 2, "replies": [] }, { "id": 1731270, "author_name": "Salman Ahmed", "author_url": "", "post_date": "2022-03-22T07:16:16", "content": "

If our dataset has a lot of classes with just 1 example, then how do you guy split that in training and validation. I

", "votes": 0, "replies": [] } ] }, { "id": 1733610, "author_name": "kaggler", "author_url": "", "post_date": "2022-03-24T12:48:50.113000", "content": "

MODEL / SIZE : EfficientNet7
\nSingle Fold, Tensorflow
\nCV : 0.830
\nLB : 0.800

", "votes": 6, "replies": [] }, { "id": 1731547, "author_name": "MOONMOON", "author_url": "", "post_date": "2022-03-22T13:27:54.413000", "content": "

UPDATE 03/27:
\nsingle fold
\neffnet-b6
\ncv=0.8257
\nlb=0.801

\n

OLD:
\nsingle fold
\neffnet-b5
\ncv=0.817
\nlb=0.779

", "votes": 3, "replies": [ { "id": 1732038, "author_name": "Salman Ahmed", "author_url": "", "post_date": "2022-03-23T00:27:28.447000", "content": "

How do you split in training and validation? Random split? Some classes has just 1 example so you are not validating on those right?

", "votes": 0, "replies": [] }, { "id": 1736140, "author_name": "MOONMOON", "author_url": "", "post_date": "2022-03-27T03:20:20.313000", "content": "

Random split

", "votes": 1, "replies": [] }, { "id": 1736145, "author_name": "老θ‚₯", "author_url": "", "post_date": "2022-03-27T03:26:23.120000", "content": "

You change model b5 to b6 and get so huge improvement, great!

", "votes": 1, "replies": [] } ] }, { "id": 1731122, "author_name": "Roc", "author_url": "", "post_date": "2022-03-22T02:15:37.693000", "content": "

effnet -b7, cv=0.815, lb=0.819 with 5 folds.

", "votes": 3, "replies": [] }, { "id": 1733313, "author_name": "NguyenThanhNhan", "author_url": "", "post_date": "2022-03-24T07:14:50.943000", "content": "

effnet-v2m, single fold, val. 8413, lb 809
\nconvnext-large, single fold, val. 8407, lb 805

", "votes": 4, "replies": [ { "id": 1733342, "author_name": "Remek Kinas", "author_url": "", "post_date": "2022-03-24T07:35:41.267000", "content": "

Very good score!! I do not ask how … but looking for solution description after competition. Looks your dataset provide great source of features to NN.

", "votes": 0, "replies": [] }, { "id": 1733356, "author_name": "Jupiter", "author_url": "", "post_date": "2022-03-24T07:50:11.163000", "content": "

wow…good job πŸ‘ An Effnet-v2m that has 809 lb is an incredible thing for me 😲. Could you give us some hints on image size and any other trick that you feel comfortable sharing? Also, I have trained Convnext too but only tried a small one because I want my image size to be at least 512, which turns out to be a long and painful journey. So, I am curious whether you reduce the image size while training a Convnext-large?

", "votes": 2, "replies": [] }, { "id": 1733614, "author_name": "kaggler", "author_url": "", "post_date": "2022-03-24T12:49:40.997000", "content": "

I tried convnext, efficientnetv2, but its score was lower than my efficientnet around 5 % when i used convnext + arcface

", "votes": 0, "replies": [] }, { "id": 1734201, "author_name": "NguyenThanhNhan", "author_url": "", "post_date": "2022-03-25T04:53:08.587000", "content": "

@remekkinas @qiuyanxinjupiter @deepkim
\nI can share the tweaks to achieve those scores after the competition ends πŸ˜„
\nYou don't need to train very long (>30 epochs) or at a very large image size (>768) though. Effnetv1 models (b6, b7) are better than effnetv2 and convnext models
\nWith post-processing, I think a single fold model can cross 0.84 on lb.

", "votes": 8, "replies": [] }, { "id": 1734350, "author_name": "老θ‚₯", "author_url": "", "post_date": "2022-03-25T08:32:31.547000", "content": "
\n

I think a single fold model can cross 0.84 on lb.

\n
\n

@andy2709 I see you have made it, good job!

", "votes": 1, "replies": [] }, { "id": 1734490, "author_name": "", "author_url": "", "post_date": "2022-03-25T11:51:25.177000", "content": "", "votes": 0, "replies": [] }, { "id": 1734497, "author_name": "huyahuya", "author_url": "", "post_date": "2022-03-25T11:56:51.947000", "content": "

@qiuyanxinjupiter οΌŒδ½ ιœ€θ¦ι˜Ÿε‹ε—οΌŸζˆ‘δ»¬δΈ€θ΅·η»„δΈͺ队？πŸ’ͺπŸ’ͺπŸ’ͺπŸ˜ƒ

", "votes": 0, "replies": [] }, { "id": 1734634, "author_name": "Remek Kinas", "author_url": "", "post_date": "2022-03-25T14:15:02.360000", "content": "

@andy2709 than you for sharing. We came to the some conclutions. We have different solution in inference and data part (I suppose) which make difference. You have better one. We are not able to cross 0.8 from single fold so far. Looking for solution :)

", "votes": 1, "replies": [] } ] }, { "id": 1730387, "author_name": "Balaji Selvaraj", "author_url": "", "post_date": "2022-03-21T07:49:43.093000", "content": "

LB 0.786 EffnetB7
\n1 fold
\n40 epochs

", "votes": 1, "replies": [ { "id": 1730394, "author_name": "Yangranran", "author_url": "", "post_date": "2022-03-21T07:56:24.253000", "content": "

Nice work!

", "votes": 1, "replies": [] }, { "id": 1730496, "author_name": "Rickyinferno", "author_url": "", "post_date": "2022-03-21T10:40:46.747000", "content": "

I am also using EffnetB7 with the detic crop dataset but my highest score is like 0.7. Could you share some thoughts about how you reach the 0.786 score? like dataset or matching technique for embedding? Thanks so much

", "votes": 0, "replies": [] }, { "id": 1730503, "author_name": "Balaji Selvaraj", "author_url": "", "post_date": "2022-03-21T11:04:27.410000", "content": "

Lot of experiments were done on

\n
    \n
  1. Model architecture {Completely different from the ones that are made public}
    2. \n
  2. Changes in the augmentations
    3. \n
\n

My initial models were getting around ~0.7
\nAfter updating model structure and changing augmentation schemes, performance increased a lot

\n

Another thing is
\nFull Body annotation > Detic > Yolo

\n

I think the backfin dataset yields better results than all the above-mentioned datasets.
\nI believe that most top-scoring kernels are using that.

\n

I think my model performance would also increase with that dataset. I am planning to set up a dataset of my own as that's the best booster for this challenge

\n

Check further details at https://www.kaggle.com/c/happy-whale-and-dolphin/discussion/313987

", "votes": 9, "replies": [] }, { "id": 1731809, "author_name": "Zhongkai Shangguan", "author_url": "", "post_date": "2022-03-22T17:53:15.993000", "content": "

May I ask what image size are you using?

", "votes": 0, "replies": [] }, { "id": 1732204, "author_name": "Balaji Selvaraj", "author_url": "", "post_date": "2022-03-23T05:54:42.983000", "content": "

Image size 600

", "votes": 0, "replies": [] } ] }, { "id": 1737021, "author_name": "Harshit Sheoran", "author_url": "", "post_date": "2022-03-28T03:43:02.363000", "content": "

b5-384, .780 LB single fold

", "votes": 0, "replies": [] }, { "id": 1736666, "author_name": "Arunkr", "author_url": "", "post_date": "2022-03-27T15:27:51.053000", "content": "

thanks everyone for insights

", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1730344": "Let's have a talk!", "1730588": "effnet b6 768 , lb 809", "1730581": "Our single is:\n- CV .820 -> LB .775 (EffNetB6)", "1733610": "MODEL / SIZE : EfficientNet7\nSingle Fold, Tensorflow\nCV : 0.830\nLB : 0.800", "1731547": "UPDATE 03/27:\nsingle fold\neffnet-b6\ncv=0.8257\nlb=0.801\n\nOLD:\nsingle fold\neffnet-b5\ncv=0.817\nlb=0.779\n", "1731122": "effnet -b7, cv=0.815, lb=0.819 with 5 folds.", "1733313": "effnet-v2m, single fold, val. 8413, lb 809\nconvnext-large, single fold, val. 8407, lb 805 ", "1730387": "LB 0.786 EffnetB7\n1 fold \n40 epochs ", "1737021": "b5-384, .780 LB single fold", "1736666": "thanks everyone for insights" } }