{ "id": 132118, "title": "Things that work (Or did not work?) (For me)", "url": "/competitions/bengaliai-cv19/discussion/132118", "author_name": "", "post_date": "2020-02-24T08:55:05.460943300Z", "votes": 43, "comment_count": 35, "views": 0, "content": "

Preprocessing:\n- Just plain resize. Other methods like crop black pixels or centering did not work as expected (I seriously have no idea why??)\n- Larger input size

\n\n

Augmentation:\n- Cutout works really (emphasized really) well. It's the key factor contributing to my CV/LB.\n- ShiftScaleRotate provides minor improvement.

\n\n

Models:\n- Densenet121 works best for me. Maybe deeper version of densenet will be better?

\n\n

Training:\n- More epochs (100-150). I haven't tried training longer.\n- ReduceLROnPlateau over OneLRCycle.

", "messages": [ { "id": "754943", "postDate": "02/24/2020 08:55:05", "content": "

Preprocessing:\n- Just plain resize. Other methods like crop black pixels or centering did not work as expected (I seriously have no idea why??)\n- Larger input size

\n\n

Augmentation:\n- Cutout works really (emphasized really) well. It's the key factor contributing to my CV/LB.\n- ShiftScaleRotate provides minor improvement.

\n\n

Models:\n- Densenet121 works best for me. Maybe deeper version of densenet will be better?

\n\n

Training:\n- More epochs (100-150). I haven't tried training longer.\n- ReduceLROnPlateau over OneLRCycle.

", "rawMarkdown": "Preprocessing:\n- Just plain resize. Other methods like crop black pixels or centering did not work as expected (I seriously have no idea why??)\n- Larger input size\n\nAugmentation:\n- Cutout works really (emphasized really) well. It's the key factor contributing to my CV/LB.\n- ShiftScaleRotate provides minor improvement.\n\nModels:\n- Densenet121 works best for me. Maybe deeper version of densenet will be better?\n\nTraining:\n- More epochs (100-150). I haven't tried training longer.\n- ReduceLROnPlateau over OneLRCycle.", "votes": null }, { "id": "755045", "postDate": "02/24/2020 11:28:40", "content": "

Center character didn't work for me. Surprise just resize work well than center character. I think center character help model recognize character stability. Finally, i decided to train with 128x128 just resize.

", "rawMarkdown": "Center character didn't work for me. Surprise just resize work well than center character. I think center character help model recognize character stability. Finally, i decided to train with 128x128 just resize.", "votes": null }, { "id": "755082", "postDate": "02/24/2020 12:08:44", "content": "

@quandapro I am thinking about adding some baseline transformations to the mix (grid-distortion, elastic-distortion, blurring etc. ) Have you tried those? If you don't mind disclosing

", "rawMarkdown": "quandapro I am thinking about adding some baseline transformations to the mix (grid-distortion, elastic-distortion, blurring etc. ) Have you tried those? If you don't mind disclosing", "votes": null }, { "id": "755096", "postDate": "02/24/2020 12:32:58", "content": "

Yes, I also tried distortion but it didn't help...

", "rawMarkdown": "Yes, I also tried distortion but it didn't help...", "votes": null }, { "id": "755098", "postDate": "02/24/2020 12:33:59", "content": "

I don't know why centering loses to plain resize though...

", "rawMarkdown": "I don't know why centering loses to plain resize though...", "votes": null }, { "id": "755349", "postDate": "02/24/2020 17:33:55", "content": "

I feel like i'm doing something wrong.. I've tried all of those things but can't get to 98 :( i will try just plain resize as image preprocessing, maybe it will give me better results!

", "rawMarkdown": "I feel like i'm doing something wrong.. I've tried all of those things but can't get to 98 :( i will try just plain resize as image preprocessing, maybe it will give me better results!", "votes": null }, { "id": "755367", "postDate": "02/24/2020 17:54:04", "content": "

+1 for \"plain resize\". Various crops or centering didn't give me any boost. What is more, I also do not know how to explain it :)

", "rawMarkdown": "1 for \"plain resize\". Various crops or centering didn't give me any boost. What is more, I also do not know how to explain it :)", "votes": null }, { "id": "755370", "postDate": "02/24/2020 17:58:04", "content": "

This is really interesting

\n\n

Quite naively, I've only been trying with re-centered images without ever trying with the plain resized ones, I'll give that a try!

", "rawMarkdown": "This is really interesting\n\nQuite naively, I've only been trying with re-centered images without ever trying with the plain resized ones, I'll give that a try!", "votes": null }, { "id": "755393", "postDate": "02/24/2020 18:33:19", "content": "

That's what I've experienced, and my frustration as well. Why cropping the image doesn't work?

", "rawMarkdown": "That's what I've experienced, and my frustration as well. Why cropping the image doesn't work?", "votes": null }, { "id": "755458", "postDate": "02/24/2020 19:41:10", "content": "

Interesting so the public 128x128 dataset might not work since it centers the characters instead of plain resizing. Something to add to my list!

", "rawMarkdown": "Interesting so the public 128x128 dataset might not work since it centers the characters instead of plain resizing. Something to add to my list!", "votes": null }, { "id": "755477", "postDate": "02/24/2020 20:02:02", "content": "

For me @maxlenormand's scaled cropped version still works best ;)

", "rawMarkdown": "For me @maxlenormand's scaled cropped [version](https://www.kaggle.com/maxlenormand/cropping-to-character-resizing-images) still works best ;)", "votes": null }, { "id": "755739", "postDate": "02/25/2020 04:35:33", "content": "

Are you using cutout in albumentations and how do you tune the parameters? Cutout doesn't work much for me.

", "rawMarkdown": "Are you using cutout in `albumentations` and how do you tune the parameters? Cutout doesn't work much for me.", "votes": null }, { "id": "755828", "postDate": "02/25/2020 06:43:44", "content": "

Hey, thanks for this helpful post. What optimizer are you using?

", "rawMarkdown": "Hey, thanks for this helpful post. What optimizer are you using?", "votes": null }, { "id": "755831", "postDate": "02/25/2020 06:48:17", "content": "

Glad to read that @joatom ! I’ll give a try to the rescaled original images nontheless! Maybe an ensemble gets with the original resized + cropped and resized might actually be even better?

", "rawMarkdown": "Glad to read that @joatom ! I’ll give a try to the rescaled original images nontheless! Maybe an ensemble gets with the original resized + cropped and resized might actually be even better?", "votes": null }, { "id": "755850", "postDate": "02/25/2020 07:22:20", "content": "

It's worth a try. \nI only applied basic augmentation (rotation, cutout, symmetric_warp) in addition with mixup.\nI didn't generate additional train data, yet.

", "rawMarkdown": "It's worth a try. \nI only applied basic augmentation (rotation, cutout, symmetric_warp) in addition with mixup.\nI didn't generate additional train data, yet.", "votes": null }, { "id": "756007", "postDate": "02/25/2020 10:45:33", "content": "

I don't know how to explain this though 😅

", "rawMarkdown": "I don't know how to explain this though 😅", "votes": null }, { "id": "756008", "postDate": "02/25/2020 10:48:25", "content": "

Yes. Cutout 1 hole with max_w, max_h equal to w // 2, h // 2 (which cover maximum 1/4 of the image). I will try more holes that cover the same area, but more spread out and random.

", "rawMarkdown": "Yes. Cutout 1 hole with max_w, max_h equal to w // 2, h // 2 (which cover maximum 1/4 of the image). I will try more holes that cover the same area, but more spread out and random.", "votes": null }, { "id": "756011", "postDate": "02/25/2020 10:49:39", "content": "

Thank you. Just good-ol-Adam 😁

", "rawMarkdown": "Thank you. Just good-ol-Adam 😁", "votes": null }, { "id": "756231", "postDate": "02/25/2020 14:33:29", "content": "

Hi @quandapro Do you mind share your factor of ReduceLROnPlateau?

", "rawMarkdown": "Hi @quandapro Do you mind share your factor of ReduceLROnPlateau?", "votes": null }, { "id": "756233", "postDate": "02/25/2020 14:35:27", "content": "

Did you just use Cutout and ShiftScaleRotate ?\nIf you don't mind to share

", "rawMarkdown": "Did you just use Cutout and ShiftScaleRotate ?\nIf you don't mind to share", "votes": null }, { "id": "756324", "postDate": "02/25/2020 16:02:56", "content": "

Your maxw, maxh are quite aggressive compared with the setting in the original paper. Thanks! I would give it a try.

", "rawMarkdown": "Your maxw, maxh are quite aggressive compared with the setting in the original paper. Thanks! I would give it a try.", "votes": null }, { "id": "756490", "postDate": "02/25/2020 19:03:44", "content": "

My explanation is that a lot of the grapheme roots appear not only by themselves, but as parts of other characters. So when the model is forced to recognized non-centered grapheme roots, it is also able to recognize them in more complex characters.

\n\n

As opposed to centering the model maybe is just able to memorize everything.

\n\n

This is just my hunch though.

", "rawMarkdown": "My explanation is that a lot of the grapheme roots appear not only by themselves, but as parts of other characters. So when the model is forced to recognized non-centered grapheme roots, it is also able to recognize them in more complex characters.\n\nAs opposed to centering the model maybe is just able to memorize everything.\n\nThis is just my hunch though.", "votes": null }, { "id": "756644", "postDate": "02/25/2020 23:10:56", "content": "

So the data you use is just purely resized 137x236 -> 128x128 even though that distorts the image?

", "rawMarkdown": "So the data you use is just purely resized 137x236 -> 128x128 even though that distorts the image?", "votes": null }, { "id": "756729", "postDate": "02/26/2020 02:41:07", "content": "

I use ReduceLROnPlateau with factor of 0.5

", "rawMarkdown": "I use ReduceLROnPlateau with factor of 0.5", "votes": null }, { "id": "756731", "postDate": "02/26/2020 02:42:46", "content": "

Yes I'm currently only use Cutout and ShiftScaleRotate. I will make a post on my augmentation experiments soon.

", "rawMarkdown": "Yes I'm currently only use Cutout and ShiftScaleRotate. I will make a post on my augmentation experiments soon.", "votes": null }, { "id": "756796", "postDate": "02/26/2020 04:55:32", "content": "

For me resize + center crop gives some boost.

\n\n

Did you try some blur augmentations/BrightnessContrast?

", "rawMarkdown": "For me resize + center crop gives some boost.\n\nDid you try some blur augmentations/BrightnessContrast?", "votes": null }, { "id": "756869", "postDate": "02/26/2020 06:44:21", "content": "

Hey @quandapro Are you using ReduceLROnPlateau with patience 5 and factor 0.5??

", "rawMarkdown": "Hey @quandapro Are you using ReduceLROnPlateau with patience 5 and factor 0.5??", "votes": null }, { "id": "757012", "postDate": "02/26/2020 10:42:37", "content": "

Maybe because when we crop by bbox and resize we lose information about how big the symbol was. Look at data - some chars take a lot of original image and some - just a little bit. It can be a predictive feature and one may try to add it to features before classification heads to see if it really the case.

", "rawMarkdown": "Maybe because when we crop by bbox and resize we lose information about how big the symbol was. Look at data - some chars take a lot of original image and some - just a little bit. It can be a predictive feature and one may try to add it to features before classification heads to see if it really the case.", "votes": null }, { "id": "757076", "postDate": "02/26/2020 12:06:41", "content": "

@kostyaatarik Thank you, this sounds logical!

", "rawMarkdown": "kostyaatarik Thank you, this sounds logical!", "votes": null }, { "id": "757223", "postDate": "02/26/2020 14:37:57", "content": "

Thank you for your sharing

", "rawMarkdown": "Thank you for your sharing", "votes": null }, { "id": "757338", "postDate": "02/26/2020 16:48:44", "content": "

Might be irrelevant to the header.When I try to run the prediction cell in which I have ensembled the outputs of two models, the kernel refreshes. It is happening for that cell alone. Btw, I am using a combination of FastAI and PyTorch in my inference notebook. What might be the issue?

", "rawMarkdown": "Might be irrelevant to the header.When I try to run the prediction cell in which I have ensembled the outputs of two models, the kernel refreshes. It is happening for that cell alone. Btw, I am using a combination of FastAI and PyTorch in my inference notebook. What might be the issue?", "votes": null }, { "id": "758166", "postDate": "02/27/2020 14:04:46", "content": "

May the symbol has more distinct detail in vertical direction, resize 128x128 also increase the vertical 🤔 detail

", "rawMarkdown": "May the symbol has more distinct detail in vertical direction, resize 128x128 also increase the vertical 🤔 detail", "votes": null }, { "id": "759066", "postDate": "02/28/2020 14:26:19", "content": "

@quandapro I agree that only plain resized images are working great rather than cropped images. Nearly gave a 0.022 boost on leaderboard! Also training for longer epochs(80-150) is the key (at least for me).

", "rawMarkdown": "quandapro I agree that only plain resized images are working great rather than cropped images. Nearly gave a 0.022 boost on leaderboard! Also training for longer epochs(80-150) is the key (at least for me).", "votes": null }, { "id": "762631", "postDate": "03/03/2020 16:59:27", "content": "

did it help?

", "rawMarkdown": "did it help?", "votes": null }, { "id": "762688", "postDate": "03/03/2020 18:08:50", "content": "

Yes it did :) you should try it

", "rawMarkdown": "Yes it did :) you should try it", "votes": null }, { "id": "763257", "postDate": "03/04/2020 09:53:14", "content": "

may I ask what learning rate are you starting with?

", "rawMarkdown": "may I ask what learning rate are you starting with?", "votes": null } ], "comments": [ { "id": 755045, "author_name": "quan0095", "author_url": "", "post_date": "02/24/2020 11:28:40", "content": "

Center character didn't work for me. Surprise just resize work well than center character. I think center character help model recognize character stability. Finally, i decided to train with 128x128 just resize.

", "votes": null, "replies": [ { "id": 755098, "author_name": "quandapro", "author_url": "", "post_date": "02/24/2020 12:33:59", "content": "

I don't know why centering loses to plain resize though...

", "votes": null, "replies": [] }, { "id": 755370, "author_name": "maxlenormand", "author_url": "", "post_date": "02/24/2020 17:58:04", "content": "

This is really interesting

\n\n

Quite naively, I've only been trying with re-centered images without ever trying with the plain resized ones, I'll give that a try!

", "votes": null, "replies": [] }, { "id": 755477, "author_name": "joatom", "author_url": "", "post_date": "02/24/2020 20:02:02", "content": "

For me @maxlenormand's scaled cropped version still works best ;)

", "votes": null, "replies": [] }, { "id": 755831, "author_name": "maxlenormand", "author_url": "", "post_date": "02/25/2020 06:48:17", "content": "

Glad to read that @joatom ! I’ll give a try to the rescaled original images nontheless! Maybe an ensemble gets with the original resized + cropped and resized might actually be even better?

", "votes": null, "replies": [] }, { "id": 755850, "author_name": "joatom", "author_url": "", "post_date": "02/25/2020 07:22:20", "content": "

It's worth a try. \nI only applied basic augmentation (rotation, cutout, symmetric_warp) in addition with mixup.\nI didn't generate additional train data, yet.

", "votes": null, "replies": [] } ] }, { "id": 755082, "author_name": "roguekk007", "author_url": "", "post_date": "02/24/2020 12:08:44", "content": "

@quandapro I am thinking about adding some baseline transformations to the mix (grid-distortion, elastic-distortion, blurring etc. ) Have you tried those? If you don't mind disclosing

", "votes": null, "replies": [ { "id": 755096, "author_name": "quandapro", "author_url": "", "post_date": "02/24/2020 12:32:58", "content": "

Yes, I also tried distortion but it didn't help...

", "votes": null, "replies": [] } ] }, { "id": 755349, "author_name": "yannmajewski", "author_url": "", "post_date": "02/24/2020 17:33:55", "content": "

I feel like i'm doing something wrong.. I've tried all of those things but can't get to 98 :( i will try just plain resize as image preprocessing, maybe it will give me better results!

", "votes": null, "replies": [ { "id": 762631, "author_name": "ankitsajwan", "author_url": "", "post_date": "03/03/2020 16:59:27", "content": "

did it help?

", "votes": null, "replies": [] }, { "id": 762688, "author_name": "yannmajewski", "author_url": "", "post_date": "03/03/2020 18:08:50", "content": "

Yes it did :) you should try it

", "votes": null, "replies": [] } ] }, { "id": 755367, "author_name": "dukhovnik", "author_url": "", "post_date": "02/24/2020 17:54:04", "content": "

+1 for \"plain resize\". Various crops or centering didn't give me any boost. What is more, I also do not know how to explain it :)

", "votes": null, "replies": [ { "id": 755393, "author_name": "axiostpc", "author_url": "", "post_date": "02/24/2020 18:33:19", "content": "

That's what I've experienced, and my frustration as well. Why cropping the image doesn't work?

", "votes": null, "replies": [] }, { "id": 757012, "author_name": "kostyaatarik", "author_url": "", "post_date": "02/26/2020 10:42:37", "content": "

Maybe because when we crop by bbox and resize we lose information about how big the symbol was. Look at data - some chars take a lot of original image and some - just a little bit. It can be a predictive feature and one may try to add it to features before classification heads to see if it really the case.

", "votes": null, "replies": [] }, { "id": 757076, "author_name": "dukhovnik", "author_url": "", "post_date": "02/26/2020 12:06:41", "content": "

@kostyaatarik Thank you, this sounds logical!

", "votes": null, "replies": [] }, { "id": 758166, "author_name": "cswwp347724", "author_url": "", "post_date": "02/27/2020 14:04:46", "content": "

May the symbol has more distinct detail in vertical direction, resize 128x128 also increase the vertical 🤔 detail

", "votes": null, "replies": [] } ] }, { "id": 755458, "author_name": "greatgamedota", "author_url": "", "post_date": "02/24/2020 19:41:10", "content": "

Interesting so the public 128x128 dataset might not work since it centers the characters instead of plain resizing. Something to add to my list!

", "votes": null, "replies": [ { "id": 756007, "author_name": "quandapro", "author_url": "", "post_date": "02/25/2020 10:45:33", "content": "

I don't know how to explain this though 😅

", "votes": null, "replies": [] }, { "id": 756490, "author_name": "linainversez", "author_url": "", "post_date": "02/25/2020 19:03:44", "content": "

My explanation is that a lot of the grapheme roots appear not only by themselves, but as parts of other characters. So when the model is forced to recognized non-centered grapheme roots, it is also able to recognize them in more complex characters.

\n\n

As opposed to centering the model maybe is just able to memorize everything.

\n\n

This is just my hunch though.

", "votes": null, "replies": [] }, { "id": 756644, "author_name": "greatgamedota", "author_url": "", "post_date": "02/25/2020 23:10:56", "content": "

So the data you use is just purely resized 137x236 -> 128x128 even though that distorts the image?

", "votes": null, "replies": [] } ] }, { "id": 755739, "author_name": "syoya1997", "author_url": "", "post_date": "02/25/2020 04:35:33", "content": "

Are you using cutout in albumentations and how do you tune the parameters? Cutout doesn't work much for me.

", "votes": null, "replies": [ { "id": 756008, "author_name": "quandapro", "author_url": "", "post_date": "02/25/2020 10:48:25", "content": "

Yes. Cutout 1 hole with max_w, max_h equal to w // 2, h // 2 (which cover maximum 1/4 of the image). I will try more holes that cover the same area, but more spread out and random.

", "votes": null, "replies": [] }, { "id": 756324, "author_name": "syoya1997", "author_url": "", "post_date": "02/25/2020 16:02:56", "content": "

Your maxw, maxh are quite aggressive compared with the setting in the original paper. Thanks! I would give it a try.

", "votes": null, "replies": [] } ] }, { "id": 755828, "author_name": "p4rallax", "author_url": "", "post_date": "02/25/2020 06:43:44", "content": "

Hey, thanks for this helpful post. What optimizer are you using?

", "votes": null, "replies": [ { "id": 756011, "author_name": "quandapro", "author_url": "", "post_date": "02/25/2020 10:49:39", "content": "

Thank you. Just good-ol-Adam 😁

", "votes": null, "replies": [] } ] }, { "id": 756231, "author_name": "cswwp347724", "author_url": "", "post_date": "02/25/2020 14:33:29", "content": "

Hi @quandapro Do you mind share your factor of ReduceLROnPlateau?

", "votes": null, "replies": [ { "id": 756729, "author_name": "quandapro", "author_url": "", "post_date": "02/26/2020 02:41:07", "content": "

I use ReduceLROnPlateau with factor of 0.5

", "votes": null, "replies": [] }, { "id": 756869, "author_name": "ratan123", "author_url": "", "post_date": "02/26/2020 06:44:21", "content": "

Hey @quandapro Are you using ReduceLROnPlateau with patience 5 and factor 0.5??

", "votes": null, "replies": [] } ] }, { "id": 756233, "author_name": "kasim0226", "author_url": "", "post_date": "02/25/2020 14:35:27", "content": "

Did you just use Cutout and ShiftScaleRotate ?\nIf you don't mind to share

", "votes": null, "replies": [ { "id": 756731, "author_name": "quandapro", "author_url": "", "post_date": "02/26/2020 02:42:46", "content": "

Yes I'm currently only use Cutout and ShiftScaleRotate. I will make a post on my augmentation experiments soon.

", "votes": null, "replies": [] }, { "id": 757223, "author_name": "kasim0226", "author_url": "", "post_date": "02/26/2020 14:37:57", "content": "

Thank you for your sharing

", "votes": null, "replies": [] } ] }, { "id": 756796, "author_name": "kupchanski", "author_url": "", "post_date": "02/26/2020 04:55:32", "content": "

For me resize + center crop gives some boost.

\n\n

Did you try some blur augmentations/BrightnessContrast?

", "votes": null, "replies": [] }, { "id": 757338, "author_name": "venky2506", "author_url": "", "post_date": "02/26/2020 16:48:44", "content": "

Might be irrelevant to the header.When I try to run the prediction cell in which I have ensembled the outputs of two models, the kernel refreshes. It is happening for that cell alone. Btw, I am using a combination of FastAI and PyTorch in my inference notebook. What might be the issue?

", "votes": null, "replies": [] }, { "id": 759066, "author_name": "ratan123", "author_url": "", "post_date": "02/28/2020 14:26:19", "content": "

@quandapro I agree that only plain resized images are working great rather than cropped images. Nearly gave a 0.022 boost on leaderboard! Also training for longer epochs(80-150) is the key (at least for me).

", "votes": null, "replies": [] }, { "id": 763257, "author_name": "yl1202", "author_url": "", "post_date": "03/04/2020 09:53:14", "content": "

may I ask what learning rate are you starting with?

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "754943": "Preprocessing:\n- Just plain resize. Other methods like crop black pixels or centering did not work as expected (I seriously have no idea why??)\n- Larger input size\n\nAugmentation:\n- Cutout works really (emphasized really) well. It's the key factor contributing to my CV/LB.\n- ShiftScaleRotate provides minor improvement.\n\nModels:\n- Densenet121 works best for me. Maybe deeper version of densenet will be better?\n\nTraining:\n- More epochs (100-150). I haven't tried training longer.\n- ReduceLROnPlateau over OneLRCycle.", "755045": "Center character didn't work for me. Surprise just resize work well than center character. I think center character help model recognize character stability. Finally, i decided to train with 128x128 just resize.", "755082": "quandapro I am thinking about adding some baseline transformations to the mix (grid-distortion, elastic-distortion, blurring etc. ) Have you tried those? If you don't mind disclosing", "755096": "Yes, I also tried distortion but it didn't help...", "755098": "I don't know why centering loses to plain resize though...", "755349": "I feel like i'm doing something wrong.. I've tried all of those things but can't get to 98 :( i will try just plain resize as image preprocessing, maybe it will give me better results!", "755367": "1 for \"plain resize\". Various crops or centering didn't give me any boost. What is more, I also do not know how to explain it :)", "755370": "This is really interesting\n\nQuite naively, I've only been trying with re-centered images without ever trying with the plain resized ones, I'll give that a try!", "755393": "That's what I've experienced, and my frustration as well. Why cropping the image doesn't work?", "755458": "Interesting so the public 128x128 dataset might not work since it centers the characters instead of plain resizing. Something to add to my list!", "755477": "For me @maxlenormand's scaled cropped [version](https://www.kaggle.com/maxlenormand/cropping-to-character-resizing-images) still works best ;)", "755739": "Are you using cutout in `albumentations` and how do you tune the parameters? Cutout doesn't work much for me.", "755828": "Hey, thanks for this helpful post. What optimizer are you using?", "755831": "Glad to read that @joatom ! I’ll give a try to the rescaled original images nontheless! Maybe an ensemble gets with the original resized + cropped and resized might actually be even better?", "755850": "It's worth a try. \nI only applied basic augmentation (rotation, cutout, symmetric_warp) in addition with mixup.\nI didn't generate additional train data, yet.", "756007": "I don't know how to explain this though 😅", "756008": "Yes. Cutout 1 hole with max_w, max_h equal to w // 2, h // 2 (which cover maximum 1/4 of the image). I will try more holes that cover the same area, but more spread out and random.", "756011": "Thank you. Just good-ol-Adam 😁", "756231": "Hi @quandapro Do you mind share your factor of ReduceLROnPlateau?", "756233": "Did you just use Cutout and ShiftScaleRotate ?\nIf you don't mind to share", "756324": "Your maxw, maxh are quite aggressive compared with the setting in the original paper. Thanks! I would give it a try.", "756490": "My explanation is that a lot of the grapheme roots appear not only by themselves, but as parts of other characters. So when the model is forced to recognized non-centered grapheme roots, it is also able to recognize them in more complex characters.\n\nAs opposed to centering the model maybe is just able to memorize everything.\n\nThis is just my hunch though.", "756644": "So the data you use is just purely resized 137x236 -> 128x128 even though that distorts the image?", "756729": "I use ReduceLROnPlateau with factor of 0.5", "756731": "Yes I'm currently only use Cutout and ShiftScaleRotate. I will make a post on my augmentation experiments soon.", "756796": "For me resize + center crop gives some boost.\n\nDid you try some blur augmentations/BrightnessContrast?", "756869": "Hey @quandapro Are you using ReduceLROnPlateau with patience 5 and factor 0.5??", "757012": "Maybe because when we crop by bbox and resize we lose information about how big the symbol was. Look at data - some chars take a lot of original image and some - just a little bit. It can be a predictive feature and one may try to add it to features before classification heads to see if it really the case.", "757076": "kostyaatarik Thank you, this sounds logical!", "757223": "Thank you for your sharing", "757338": "Might be irrelevant to the header.When I try to run the prediction cell in which I have ensembled the outputs of two models, the kernel refreshes. It is happening for that cell alone. Btw, I am using a combination of FastAI and PyTorch in my inference notebook. What might be the issue?", "758166": "May the symbol has more distinct detail in vertical direction, resize 128x128 also increase the vertical 🤔 detail", "759066": "quandapro I agree that only plain resized images are working great rather than cropped images. Nearly gave a 0.022 boost on leaderboard! Also training for longer epochs(80-150) is the key (at least for me).", "762631": "did it help?", "762688": "Yes it did :) you should try it", "763257": "may I ask what learning rate are you starting with?" }, "source": "meta" }