{ "id": 104818, "title": "How far can 256x256 image size go?", "url": "/competitions/aptos2019-blindness-detection/discussion/104818", "author_name": "Quan", "post_date": "2019-08-19T11:12:40.935000", "votes": 24, "comment_count": 37, "views": 0, "content": "

For me is 0.811 on LB. What about you?

", "messages": [ { "id": 602674, "postDate": "2019-08-19T11:12:40.937Z", "content": "

For me is 0.811 on LB. What about you?

", "rawMarkdown": "For me is 0.811 on LB. What about you?", "votes": 23 }, { "id": 602693, "postDate": "2019-08-19T11:52:56.013Z", "content": "

Nice! Did not explore how far I could push 256x256. Started at 224x224 and went from there. Curious to hear what architecture you used!

", "rawMarkdown": "Nice! Did not explore how far I could push 256x256. Started at 224x224 and went from there. Curious to hear what architecture you used!", "votes": 1 }, { "id": 602689, "postDate": "2019-08-19T11:49:54.390Z", "content": "

What kind of model do you use?

", "rawMarkdown": "What kind of model do you use?", "votes": 1, "replies": [ { "id": 602760, "postDate": "2019-08-19T13:12:51.373Z", "content": "

I use EfficientNetB2 (Though it's default image size is 260)

", "rawMarkdown": "I use EfficientNetB2 (Though it's default image size is 260)", "votes": 2 }, { "id": 602762, "postDate": "2019-08-19T13:14:35.177Z", "content": "

I also tried 224x224 with EfficientNetB0 and got a suprisingly good result, pretty close to EfficientNetB2 with 256x256 images

", "rawMarkdown": "I also tried 224x224 with EfficientNetB0 and got a suprisingly good result, pretty close to EfficientNetB2 with 256x256 images" }, { "id": 602790, "postDate": "2019-08-19T13:47:32.393Z", "content": "

What kind of preprocessing you used?

", "rawMarkdown": "What kind of preprocessing you used?" }, { "id": 602812, "postDate": "2019-08-19T14:16:50.390Z", "content": "

I don't do any preprocessing in this experiment. I do a lot of image augmenting like rotation, zoom, flip, constrast, etc... Surprisingly 224 with B0 are almost as good as 256 with B2, and faster to train too!

", "rawMarkdown": "I don't do any preprocessing in this experiment. I do a lot of image augmenting like rotation, zoom, flip, constrast, etc... Surprisingly 224 with B0 are almost as good as 256 with B2, and faster to train too!", "votes": 1 }, { "id": 602817, "postDate": "2019-08-19T14:22:44.077Z", "content": "

That's awesome! Are you also using a lot of additional data (like APTOS 2015)?

", "rawMarkdown": "That's awesome! Are you also using a lot of additional data (like APTOS 2015)?", "votes": 1 }, { "id": 602823, "postDate": "2019-08-19T14:26:36.720Z", "content": "

TTA or training augmentations only?

", "rawMarkdown": "TTA or training augmentations only?" }, { "id": 602831, "postDate": "2019-08-19T14:33:22.930Z", "content": "

Thanks for sharing :)

", "rawMarkdown": "Thanks for sharing :)" }, { "id": 603164, "postDate": "2019-08-20T01:00:41.097Z", "content": "

@carlolepelaars Yes I use both current and 2015 training data.\n@abhishek No I don't do any TTA, I do a lot of training augmentations.\n@tahsin You're welcome!

", "rawMarkdown": "@carlolepelaars Yes I use both current and 2015 training data.\n@abhishek No I don't do any TTA, I do a lot of training augmentations.\n@tahsin You're welcome!", "votes": 1 }, { "id": 603190, "postDate": "2019-08-20T02:09:29.377Z", "content": "

Using only this competition data? CV consistent with LB?

", "rawMarkdown": "Using only this competition data? CV consistent with LB?" }, { "id": 603221, "postDate": "2019-08-20T03:06:42.403Z", "content": "

hi ! can you tell me more about your augment ? and did you use the both old data and new data(not pretrain in old data then train in new data) to train your model ? thx :)

", "rawMarkdown": "hi ! can you tell me more about your augment ? and did you use the both old data and new data(not pretrain in old data then train in new data) to train your model ? thx :)" }, { "id": 603349, "postDate": "2019-08-20T07:12:10.927Z", "content": "

@ggbrother I just mix them up, split into train and valid set then train them in one go. I only do augmentation on training set.

", "rawMarkdown": "@ggbrother I just mix them up, split into train and valid set then train them in one go. I only do augmentation on training set.", "votes": 1 }, { "id": 603461, "postDate": "2019-08-20T09:49:17.543Z", "content": "

@quandapro Are you treating the problem as multilabel classification, multiclass classification or regression ?

", "rawMarkdown": "@quandapro Are you treating the problem as multilabel classification, multiclass classification or regression ? " }, { "id": 603664, "postDate": "2019-08-20T14:34:42.803Z", "content": "

@quandapro mind to share you used all 2015 data or just fraction of them ?

", "rawMarkdown": "@quandapro mind to share you used all 2015 data or just fraction of them ?", "votes": 1 }, { "id": 603797, "postDate": "2019-08-20T17:11:23.783Z", "content": "

@quandapro did you use 2015 test data as well? For me it elevated LB score when i used training data of 2015 but gave a lower LB after adding 2015 test data

", "rawMarkdown": "@quandapro did you use 2015 test data as well? For me it elevated LB score when i used training data of 2015 but gave a lower LB after adding 2015 test data" }, { "id": 604172, "postDate": "2019-08-21T05:39:50.690Z", "content": "

@sabbiracoustic1006 I don't know about 2015 test data. Maybe it has different distribution compared to current test data?

", "rawMarkdown": "@sabbiracoustic1006 I don't know about 2015 test data. Maybe it has different distribution compared to current test data?", "votes": 1 } ] }, { "id": 603251, "postDate": "2019-08-20T04:21:36.833Z", "content": "

For me is 0.815, B3 as classifier, with a lot of image augment, TTA. \nI am not sure Public LB is reliable.

", "rawMarkdown": "For me is 0.815, B3 as classifier, with a lot of image augment, TTA. \nI am not sure Public LB is reliable. \n", "votes": 2, "replies": [ { "id": 604000, "postDate": "2019-08-20T23:46:43.773Z", "content": "

Great result! Are you using classification or regression approach?

", "rawMarkdown": "Great result! Are you using classification or regression approach?" }, { "id": 604015, "postDate": "2019-08-21T00:13:16.447Z", "content": "

classification

", "rawMarkdown": "classification", "votes": 1 }, { "id": 604021, "postDate": "2019-08-21T00:29:59.573Z", "content": "

what type of TTA you use? how much it improves on lb?

", "rawMarkdown": "what type of TTA you use? how much it improves on lb?" }, { "id": 604053, "postDate": "2019-08-21T01:26:53.390Z", "content": "

Just rotation, zoom, flip.

", "rawMarkdown": "Just rotation, zoom, flip." }, { "id": 604149, "postDate": "2019-08-21T05:03:55.213Z", "content": "

@chanhu Did you use 2015 data as well? And you treated the problem as multilabel classification or multiclass classification? And what preprocessing did you use for training?

", "rawMarkdown": "@chanhu Did you use 2015 data as well? And you treated the problem as multilabel classification or multiclass classification? And what preprocessing did you use for training? " }, { "id": 604253, "postDate": "2019-08-21T07:31:35.950Z", "content": "
    \n
  1. using 2015 data（only train data）
    2. \n
  2. considering problem as simple classification.
    3. \n
", "rawMarkdown": "1. using 2015 data（only train data）\n2. considering problem as simple classification. ", "votes": 1 } ] }, { "id": 604199, "postDate": "2019-08-21T06:19:54.760Z", "content": "

Pay attention to image preprocessing! Heavy preprocessing may reduce the performance of EfficientNet (0.811 -> 0.732). Maybe will try to play around with parameters like width, depth, resolution coefficient of EfficientNet.

", "rawMarkdown": "Pay attention to image preprocessing! Heavy preprocessing may reduce the performance of EfficientNet (0.811 -> 0.732). Maybe will try to play around with parameters like width, depth, resolution coefficient of EfficientNet.", "replies": [ { "id": 604204, "postDate": "2019-08-21T06:23:54.277Z", "content": "

In EfficientNet paper https://arxiv.org/abs/1905.11946, they did grid search to find the balanced width, depth and resolution of the model to achieve high accuracy and low computational cost. But they didn't do any heavy preprocessing on training images (Just crop). So default settings might not be optimized if we do heavy preprocessing!

", "rawMarkdown": "In EfficientNet paper https://arxiv.org/abs/1905.11946, they did grid search to find the balanced width, depth and resolution of the model to achieve high accuracy and low computational cost. But they didn't do any heavy preprocessing on training images (Just crop). So default settings might not be optimized if we do heavy preprocessing!", "votes": 2 }, { "id": 605038, "postDate": "2019-08-22T02:06:06.207Z", "content": "

yes,i have try alot of data augment like autoaugment and some heavy,it seems didn't work and get a worse score about 0.71 ,can you share me more detail about your data augment ?

", "rawMarkdown": "yes,i have try alot of data augment like autoaugment and some heavy,it seems didn't work and get a worse score about 0.71 ,can you share me more detail about your data augment ? " } ] }, { "id": 603852, "postDate": "2019-08-20T18:11:25.793Z", "content": "

Hi @quandapro, I saw in a comment that you added the previous competition dataset, I tried to add it aswell. Using the same model (EfficientNetB4) that I used for the 2019 dataset, however I can't make the model converge. May I ask you if you have stratified the dataset or removed some images ?

", "rawMarkdown": "Hi @quandapro, I saw in a comment that you added the previous competition dataset, I tried to add it aswell. Using the same model (EfficientNetB4) that I used for the 2019 dataset, however I can't make the model converge. May I ask you if you have stratified the dataset or removed some images ?", "replies": [ { "id": 604154, "postDate": "2019-08-21T05:12:52.727Z", "content": "

I just include the whole dataset as it is. Also start with imagenet weights for faster training.

", "rawMarkdown": "I just include the whole dataset as it is. Also start with imagenet weights for faster training.", "votes": 1 }, { "id": 604187, "postDate": "2019-08-21T06:04:03.830Z", "content": "

Hi, Quan. Did you just train the 2015+2019dataset, not pretrained on 2015 and then train on 2019 dataset?

", "rawMarkdown": "Hi, Quan. Did you just train the 2015+2019dataset, not pretrained on 2015 and then train on 2019 dataset? " }, { "id": 604192, "postDate": "2019-08-21T06:14:09.737Z", "content": "

Thank you, I will try this

", "rawMarkdown": "Thank you, I will try this" }, { "id": 604195, "postDate": "2019-08-21T06:16:47.163Z", "content": "

@garybios Exactly!

", "rawMarkdown": "@garybios Exactly!" } ] }, { "id": 603143, "postDate": "2019-08-19T23:30:46.477Z", "content": "

From my experience, using default image size for EfficientNet always helps no matter what. To stay on topic though, I managed to hit .80 with 256.

", "rawMarkdown": "From my experience, using default image size for EfficientNet always helps no matter what. To stay on topic though, I managed to hit .80 with 256.", "replies": [ { "id": 603171, "postDate": "2019-08-20T01:13:50.593Z", "content": "

Sound great! I find that getting 1-2% above 80+% on LB is extremely challenging!

", "rawMarkdown": "Sound great! I find that getting 1-2% above 80+% on LB is extremely challenging!" } ] }, { "id": 602792, "postDate": "2019-08-19T13:50:15.933Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 602815, "postDate": "2019-08-19T14:17:09.513Z", "content": "

No I don't.

", "rawMarkdown": "No I don't." }, { "id": 603823, "postDate": "2019-08-20T17:40:06.893Z", "content": "

What was the deleted question?

", "rawMarkdown": "What was the deleted question? ", "votes": 1 } ] } ], "comments": [ { "id": 602693, "author_name": "Carlo", "author_url": "", "post_date": "2019-08-19T11:52:56.013000", "content": "

Nice! Did not explore how far I could push 256x256. Started at 224x224 and went from there. Curious to hear what architecture you used!

", "votes": 1, "replies": [] }, { "id": 602689, "author_name": "Abhishek Thakur", "author_url": "", "post_date": "2019-08-19T11:49:54.390000", "content": "

What kind of model do you use?

", "votes": 1, "replies": [ { "id": 602760, "author_name": "Quan", "author_url": "", "post_date": "2019-08-19T13:12:51.373000", "content": "

I use EfficientNetB2 (Though it's default image size is 260)

", "votes": 2, "replies": [] }, { "id": 602762, "author_name": "Quan", "author_url": "", "post_date": "2019-08-19T13:14:35.177000", "content": "

I also tried 224x224 with EfficientNetB0 and got a suprisingly good result, pretty close to EfficientNetB2 with 256x256 images

", "votes": 0, "replies": [] }, { "id": 602790, "author_name": "Tahsin Mostafiz", "author_url": "", "post_date": "2019-08-19T13:47:32.393000", "content": "

What kind of preprocessing you used?

", "votes": 0, "replies": [] }, { "id": 602812, "author_name": "Quan", "author_url": "", "post_date": "2019-08-19T14:16:50.390000", "content": "

I don't do any preprocessing in this experiment. I do a lot of image augmenting like rotation, zoom, flip, constrast, etc... Surprisingly 224 with B0 are almost as good as 256 with B2, and faster to train too!

", "votes": 1, "replies": [] }, { "id": 602817, "author_name": "Carlo", "author_url": "", "post_date": "2019-08-19T14:22:44.077000", "content": "

That's awesome! Are you also using a lot of additional data (like APTOS 2015)?

", "votes": 1, "replies": [] }, { "id": 602823, "author_name": "Abhishek Thakur", "author_url": "", "post_date": "2019-08-19T14:26:36.720000", "content": "

TTA or training augmentations only?

", "votes": 0, "replies": [] }, { "id": 602831, "author_name": "Tahsin Mostafiz", "author_url": "", "post_date": "2019-08-19T14:33:22.930000", "content": "

Thanks for sharing :)

", "votes": 0, "replies": [] }, { "id": 603164, "author_name": "Quan", "author_url": "", "post_date": "2019-08-20T01:00:41.097000", "content": "

@carlolepelaars Yes I use both current and 2015 training data.\n@abhishek No I don't do any TTA, I do a lot of training augmentations.\n@tahsin You're welcome!

", "votes": 1, "replies": [] }, { "id": 603190, "author_name": "Bibek", "author_url": "", "post_date": "2019-08-20T02:09:29.377000", "content": "

Using only this competition data? CV consistent with LB?

", "votes": 0, "replies": [] }, { "id": 603221, "author_name": "YILING X", "author_url": "", "post_date": "2019-08-20T03:06:42.403000", "content": "

hi ! can you tell me more about your augment ? and did you use the both old data and new data(not pretrain in old data then train in new data) to train your model ? thx :)

", "votes": 0, "replies": [] }, { "id": 603349, "author_name": "Quan", "author_url": "", "post_date": "2019-08-20T07:12:10.927000", "content": "

@ggbrother I just mix them up, split into train and valid set then train them in one go. I only do augmentation on training set.

", "votes": 1, "replies": [] }, { "id": 603461, "author_name": "Viraj Bagal", "author_url": "", "post_date": "2019-08-20T09:49:17.543000", "content": "

@quandapro Are you treating the problem as multilabel classification, multiclass classification or regression ?

", "votes": 0, "replies": [] }, { "id": 603664, "author_name": "CC Joshua ", "author_url": "", "post_date": "2019-08-20T14:34:42.803000", "content": "

@quandapro mind to share you used all 2015 data or just fraction of them ?

", "votes": 1, "replies": [] }, { "id": 603797, "author_name": "Sabbir Ahmed", "author_url": "", "post_date": "2019-08-20T17:11:23.783000", "content": "

@quandapro did you use 2015 test data as well? For me it elevated LB score when i used training data of 2015 but gave a lower LB after adding 2015 test data

", "votes": 0, "replies": [] }, { "id": 604172, "author_name": "Quan", "author_url": "", "post_date": "2019-08-21T05:39:50.690000", "content": "

@sabbiracoustic1006 I don't know about 2015 test data. Maybe it has different distribution compared to current test data?

", "votes": 1, "replies": [] } ] }, { "id": 603251, "author_name": "Chanhu", "author_url": "", "post_date": "2019-08-20T04:21:36.833000", "content": "

For me is 0.815, B3 as classifier, with a lot of image augment, TTA. \nI am not sure Public LB is reliable.

", "votes": 2, "replies": [ { "id": 604000, "author_name": "Salil Mishra", "author_url": "", "post_date": "2019-08-20T23:46:43.773000", "content": "

Great result! Are you using classification or regression approach?

", "votes": 0, "replies": [] }, { "id": 604015, "author_name": "Chanhu", "author_url": "", "post_date": "2019-08-21T00:13:16.447000", "content": "

classification

", "votes": 1, "replies": [] }, { "id": 604021, "author_name": "Nguyen Quan Anh Minh", "author_url": "", "post_date": "2019-08-21T00:29:59.573000", "content": "

what type of TTA you use? how much it improves on lb?

", "votes": 0, "replies": [] }, { "id": 604053, "author_name": "Chanhu", "author_url": "", "post_date": "2019-08-21T01:26:53.390000", "content": "

Just rotation, zoom, flip.

", "votes": 0, "replies": [] }, { "id": 604149, "author_name": "Viraj Bagal", "author_url": "", "post_date": "2019-08-21T05:03:55.213000", "content": "

@chanhu Did you use 2015 data as well? And you treated the problem as multilabel classification or multiclass classification? And what preprocessing did you use for training?

", "votes": 0, "replies": [] }, { "id": 604253, "author_name": "Chanhu", "author_url": "", "post_date": "2019-08-21T07:31:35.950000", "content": "
    \n
  1. using 2015 data（only train data）
    2. \n
  2. considering problem as simple classification.
    3. \n
", "votes": 1, "replies": [] } ] }, { "id": 604199, "author_name": "Quan", "author_url": "", "post_date": "2019-08-21T06:19:54.760000", "content": "

Pay attention to image preprocessing! Heavy preprocessing may reduce the performance of EfficientNet (0.811 -> 0.732). Maybe will try to play around with parameters like width, depth, resolution coefficient of EfficientNet.

", "votes": 0, "replies": [ { "id": 604204, "author_name": "Quan", "author_url": "", "post_date": "2019-08-21T06:23:54.277000", "content": "

In EfficientNet paper https://arxiv.org/abs/1905.11946, they did grid search to find the balanced width, depth and resolution of the model to achieve high accuracy and low computational cost. But they didn't do any heavy preprocessing on training images (Just crop). So default settings might not be optimized if we do heavy preprocessing!

", "votes": 2, "replies": [] }, { "id": 605038, "author_name": "YILING X", "author_url": "", "post_date": "2019-08-22T02:06:06.207000", "content": "

yes,i have try alot of data augment like autoaugment and some heavy,it seems didn't work and get a worse score about 0.71 ,can you share me more detail about your data augment ?

", "votes": 0, "replies": [] } ] }, { "id": 603852, "author_name": "Jonas", "author_url": "", "post_date": "2019-08-20T18:11:25.793000", "content": "

Hi @quandapro, I saw in a comment that you added the previous competition dataset, I tried to add it aswell. Using the same model (EfficientNetB4) that I used for the 2019 dataset, however I can't make the model converge. May I ask you if you have stratified the dataset or removed some images ?

", "votes": 0, "replies": [ { "id": 604154, "author_name": "Quan", "author_url": "", "post_date": "2019-08-21T05:12:52.727000", "content": "

I just include the whole dataset as it is. Also start with imagenet weights for faster training.

", "votes": 1, "replies": [] }, { "id": 604187, "author_name": "Gary", "author_url": "", "post_date": "2019-08-21T06:04:03.830000", "content": "

Hi, Quan. Did you just train the 2015+2019dataset, not pretrained on 2015 and then train on 2019 dataset?

", "votes": 0, "replies": [] }, { "id": 604192, "author_name": "Jonas", "author_url": "", "post_date": "2019-08-21T06:14:09.737000", "content": "

Thank you, I will try this

", "votes": 0, "replies": [] }, { "id": 604195, "author_name": "Quan", "author_url": "", "post_date": "2019-08-21T06:16:47.163000", "content": "

@garybios Exactly!

", "votes": 0, "replies": [] } ] }, { "id": 603143, "author_name": "Drei", "author_url": "", "post_date": "2019-08-19T23:30:46.477000", "content": "

From my experience, using default image size for EfficientNet always helps no matter what. To stay on topic though, I managed to hit .80 with 256.

", "votes": 0, "replies": [ { "id": 603171, "author_name": "Quan", "author_url": "", "post_date": "2019-08-20T01:13:50.593000", "content": "

Sound great! I find that getting 1-2% above 80+% on LB is extremely challenging!

", "votes": 0, "replies": [] } ] }, { "id": 602792, "author_name": "", "author_url": "", "post_date": "2019-08-19T13:50:15.933000", "content": "", "votes": 0, "replies": [ { "id": 602815, "author_name": "Quan", "author_url": "", "post_date": "2019-08-19T14:17:09.513000", "content": "

No I don't.

", "votes": 0, "replies": [] }, { "id": 603823, "author_name": "S D", "author_url": "", "post_date": "2019-08-20T17:40:06.893000", "content": "

What was the deleted question?

", "votes": 1, "replies": [] } ] } ], "raw_markdown_by_id": { "602674": "For me is 0.811 on LB. What about you?", "602693": "Nice! Did not explore how far I could push 256x256. Started at 224x224 and went from there. Curious to hear what architecture you used!", "602689": "What kind of model do you use?", "603251": "For me is 0.815, B3 as classifier, with a lot of image augment, TTA. \nI am not sure Public LB is reliable. \n", "604199": "Pay attention to image preprocessing! Heavy preprocessing may reduce the performance of EfficientNet (0.811 -> 0.732). Maybe will try to play around with parameters like width, depth, resolution coefficient of EfficientNet.", "603852": "Hi @quandapro, I saw in a comment that you added the previous competition dataset, I tried to add it aswell. Using the same model (EfficientNetB4) that I used for the 2019 dataset, however I can't make the model converge. May I ask you if you have stratified the dataset or removed some images ?", "603143": "From my experience, using default image size for EfficientNet always helps no matter what. To stay on topic though, I managed to hit .80 with 256.", "602792": "" } }