Hi all, I have tested my models on different size of images. However, models feed by 256 * 256 images are much better than 512 * 512. In my opinion, a larger image means more information. Why this would happen?
", "messages": [ { "id": "947299", "postDate": "07/27/2020 07:13:20", "content": "Hi all, I have tested my models on different size of images. However, models feed by 256 * 256 images are much better than 512 * 512. In my opinion, a larger image means more information. Why this would happen?
", "rawMarkdown": "Hi all, I have tested my models on different size of images. However, models feed by 256 * 256 images are much better than 512 * 512. In my opinion, a larger image means more information. Why this would happen?", "votes": null }, { "id": "947356", "postDate": "07/27/2020 07:53:56", "content": "This may help! :)
", "rawMarkdown": "[This](https://www.kaggle.com/c/siim-isic-melanoma-classification/discussion/160147) may help! :)", "votes": null }, { "id": "947373", "postDate": "07/27/2020 08:15:32", "content": "Large image size have more pixels thus contains large data about the images. So if you train your images on large size images then definitely your model perform well. But there is a cost that is It will increase computation of your model.
", "rawMarkdown": "Large image size have more pixels thus contains large data about the images. So if you train your images on large size images then definitely your model perform well. But there is a cost that is It will increase computation of your model.", "votes": null }, { "id": "947396", "postDate": "07/27/2020 08:40:24", "content": "Hi, larger sizes mean larger neural networks, larger neural network results in a more complex problem to solve, that's probably why it works better with smaller images (as long as they are still understandable).
", "rawMarkdown": "Hi, larger sizes mean larger neural networks, larger neural network results in a more complex problem to solve, that's probably why it works better with smaller images (as long as they are still understandable).", "votes": null }, { "id": "947432", "postDate": "07/27/2020 09:09:51", "content": "Hi, with 512 I succeed to have 0.95X+ results (single model) contrary to 256 so I suspect 512 seems better for me .
", "rawMarkdown": "Hi, with 512 I succeed to have 0.95X+ results (single model) contrary to 256 so I suspect 512 seems better for me .", "votes": null }, { "id": "947506", "postDate": "07/27/2020 10:04:10", "content": "I also noted this...it is performing well only in low size images. Don't find a reason...
", "rawMarkdown": "I also noted this...it is performing well only in low size images. Don't find a reason...", "votes": null }, { "id": "947567", "postDate": "07/27/2020 10:52:53", "content": "I started from 32x32, in the last days I was working on 128x128 now I switched to 256x256
", "rawMarkdown": "I started from 32x32, in the last days I was working on 128x128 now I switched to 256x256", "votes": null }, { "id": "947660", "postDate": "07/27/2020 12:12:15", "content": "Thanks a lot!
", "rawMarkdown": "Thanks a lot!", "votes": null }, { "id": "959796", "postDate": "08/05/2020 22:38:49", "content": "If the only thing you changed was image size, then maybe your parameters are calibrated so much to 256x256 that you need to change some of them for it to be more optimal for 512x512. I assume you changed batch size if you moved up to 512x512. 512x512 is not double 256x256, it is 4x. So typically you would have to use 1/4 (or so) batch size. Batch Size can effect learning rate, it can also effect batch normalization.
\n\nAre you still using 256? If so great job, you are 357 on LB so you are doing great.
", "rawMarkdown": "If the only thing you changed was image size, then maybe your parameters are calibrated so much to 256x256 that you need to change some of them for it to be more optimal for 512x512. I assume you changed batch size if you moved up to 512x512. 512x512 is not double 256x256, it is 4x. So typically you would have to use 1/4 (or so) batch size. Batch Size can effect learning rate, it can also effect batch normalization. \n\nAre you still using 256? If so great job, you are 357 on LB so you are doing great.", "votes": null }, { "id": "959797", "postDate": "08/05/2020 22:40:25", "content": "@ludovick what were you getting with 256? Doesn't it make sense to work with 256 and optmize and then switch to 512 just before competition ends?
", "rawMarkdown": "ludovick what were you getting with 256? Doesn't it make sense to work with 256 and optmize and then switch to 512 just before competition ends?", "votes": null }, { "id": "959805", "postDate": "08/05/2020 22:53:41", "content": "with 256 I had difficulty to get good results. It was around 0.92X - 0.93X (single model). With ensembling of 256 model and TTA I can get 0.94X occasionnaly. But I may have bad hyper-parameters. I am not doing K - fold, just doing a 90/10 % (train/val) splits.
", "rawMarkdown": "with 256 I had difficulty to get good results. It was around 0.92X - 0.93X (single model). With ensembling of 256 model and TTA I can get 0.94X occasionnaly. But I may have bad hyper-parameters. I am not doing K - fold, just doing a 90/10 % (train/val) splits.", "votes": null }, { "id": "959806", "postDate": "08/05/2020 22:58:33", "content": "Regarding your second question, as the competition ends soon, I will advice you to start with 384/512 in order to have few strong models and doing an ensembling models as a baseline.
", "rawMarkdown": "Regarding your second question, as the competition ends soon, I will advice you to start with 384/512 in order to have few strong models and doing an ensembling models as a baseline.", "votes": null } ], "comments": [ { "id": 947356, "author_name": "sarques", "author_url": "", "post_date": "07/27/2020 07:53:56", "content": "This may help! :)
", "votes": null, "replies": [ { "id": 947660, "author_name": "vincentrenzw", "author_url": "", "post_date": "07/27/2020 12:12:15", "content": "Thanks a lot!
", "votes": null, "replies": [] } ] }, { "id": 947373, "author_name": "aman2000jaiswal", "author_url": "", "post_date": "07/27/2020 08:15:32", "content": "Large image size have more pixels thus contains large data about the images. So if you train your images on large size images then definitely your model perform well. But there is a cost that is It will increase computation of your model.
", "votes": null, "replies": [] }, { "id": 947396, "author_name": "grumbach", "author_url": "", "post_date": "07/27/2020 08:40:24", "content": "Hi, larger sizes mean larger neural networks, larger neural network results in a more complex problem to solve, that's probably why it works better with smaller images (as long as they are still understandable).
", "votes": null, "replies": [] }, { "id": 947432, "author_name": "ludovick", "author_url": "", "post_date": "07/27/2020 09:09:51", "content": "Hi, with 512 I succeed to have 0.95X+ results (single model) contrary to 256 so I suspect 512 seems better for me .
", "votes": null, "replies": [ { "id": 959797, "author_name": "brianfeeny", "author_url": "", "post_date": "08/05/2020 22:40:25", "content": "@ludovick what were you getting with 256? Doesn't it make sense to work with 256 and optmize and then switch to 512 just before competition ends?
", "votes": null, "replies": [] }, { "id": 959805, "author_name": "ludovick", "author_url": "", "post_date": "08/05/2020 22:53:41", "content": "with 256 I had difficulty to get good results. It was around 0.92X - 0.93X (single model). With ensembling of 256 model and TTA I can get 0.94X occasionnaly. But I may have bad hyper-parameters. I am not doing K - fold, just doing a 90/10 % (train/val) splits.
", "votes": null, "replies": [] }, { "id": 959806, "author_name": "ludovick", "author_url": "", "post_date": "08/05/2020 22:58:33", "content": "Regarding your second question, as the competition ends soon, I will advice you to start with 384/512 in order to have few strong models and doing an ensembling models as a baseline.
", "votes": null, "replies": [] } ] }, { "id": 947506, "author_name": "timothyalexjohn", "author_url": "", "post_date": "07/27/2020 10:04:10", "content": "I also noted this...it is performing well only in low size images. Don't find a reason...
", "votes": null, "replies": [] }, { "id": 947567, "author_name": "jacekpoplawski", "author_url": "", "post_date": "07/27/2020 10:52:53", "content": "I started from 32x32, in the last days I was working on 128x128 now I switched to 256x256
", "votes": null, "replies": [] }, { "id": 959796, "author_name": "brianfeeny", "author_url": "", "post_date": "08/05/2020 22:38:49", "content": "If the only thing you changed was image size, then maybe your parameters are calibrated so much to 256x256 that you need to change some of them for it to be more optimal for 512x512. I assume you changed batch size if you moved up to 512x512. 512x512 is not double 256x256, it is 4x. So typically you would have to use 1/4 (or so) batch size. Batch Size can effect learning rate, it can also effect batch normalization.
\n\nAre you still using 256? If so great job, you are 357 on LB so you are doing great.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "947299": "Hi all, I have tested my models on different size of images. However, models feed by 256 * 256 images are much better than 512 * 512. In my opinion, a larger image means more information. Why this would happen?", "947356": "[This](https://www.kaggle.com/c/siim-isic-melanoma-classification/discussion/160147) may help! :)", "947373": "Large image size have more pixels thus contains large data about the images. So if you train your images on large size images then definitely your model perform well. But there is a cost that is It will increase computation of your model.", "947396": "Hi, larger sizes mean larger neural networks, larger neural network results in a more complex problem to solve, that's probably why it works better with smaller images (as long as they are still understandable).", "947432": "Hi, with 512 I succeed to have 0.95X+ results (single model) contrary to 256 so I suspect 512 seems better for me .", "947506": "I also noted this...it is performing well only in low size images. Don't find a reason...", "947567": "I started from 32x32, in the last days I was working on 128x128 now I switched to 256x256", "947660": "Thanks a lot!", "959796": "If the only thing you changed was image size, then maybe your parameters are calibrated so much to 256x256 that you need to change some of them for it to be more optimal for 512x512. I assume you changed batch size if you moved up to 512x512. 512x512 is not double 256x256, it is 4x. So typically you would have to use 1/4 (or so) batch size. Batch Size can effect learning rate, it can also effect batch normalization. \n\nAre you still using 256? If so great job, you are 357 on LB so you are doing great.", "959797": "ludovick what were you getting with 256? Doesn't it make sense to work with 256 and optmize and then switch to 512 just before competition ends?", "959805": "with 256 I had difficulty to get good results. It was around 0.92X - 0.93X (single model). With ensembling of 256 model and TTA I can get 0.94X occasionnaly. But I may have bad hyper-parameters. I am not doing K - fold, just doing a 90/10 % (train/val) splits.", "959806": "Regarding your second question, as the competition ends soon, I will advice you to start with 384/512 in order to have few strong models and doing an ensembling models as a baseline." }, "source": "meta" }