![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1943421%2F9c55f4b8208645a5d717575419f8a563%2FScreenshot%20from%202020-02-29%2011-25-21.png?generation=1582946762554639&alt=media\")
Well, it is no surprise now that we should NOT use cropped images.\nMerely doing this and using OneCycle schedule gave me a ~.015 boost and pulled me out of bronze
\n\nI noticed this comment, the 1st place's comments are always worth surveying\n
Wonder if anyone did experiments on this? It just seems counter-intuitive that resizing the image beyond original size (fundamentally not adding more information) might provide a boost to the model.
", "messages": [ { "id": "759451", "postDate": "02/29/2020 03:26:35", "content": "Well, it is no surprise now that we should NOT use cropped images.\nMerely doing this and using OneCycle schedule gave me a ~.015 boost and pulled me out of bronze
\n\nI noticed this comment, the 1st place's comments are always worth surveying\n
Wonder if anyone did experiments on this? It just seems counter-intuitive that resizing the image beyond original size (fundamentally not adding more information) might provide a boost to the model.
", "rawMarkdown": "Well, it is no surprise now that we should NOT use cropped images.\nMerely doing this and using OneCycle schedule gave me a ~.015 boost and pulled me out of bronze\n\nI noticed this comment, the 1st place's comments are always worth surveying\n\n\nWonder if anyone did experiments on this? It just seems counter-intuitive that resizing the image beyond original size (fundamentally not adding more information) might provide a boost to the model.", "votes": null }, { "id": "759457", "postDate": "02/29/2020 04:01:44", "content": "\"fundamentally not adding more information\"
\n\nit does, bucause of pooling and stride
", "rawMarkdown": "\"fundamentally not adding more information\"\n\nit does, bucause of pooling and stride", "votes": null }, { "id": "759795", "postDate": "02/29/2020 13:24:59", "content": "@hengck23 Not really getting what you mean there😁 . We are only creating pixels out of existing ones, so it is kind of like feature engineering. The only explanation I can come up for possible improvement is that CNN prefers square images and square feature maps (which sounds tenure).
", "rawMarkdown": "hengck23 Not really getting what you mean there😁 . We are only creating pixels out of existing ones, so it is kind of like feature engineering. The only explanation I can come up for possible improvement is that CNN prefers square images and square feature maps (which sounds tenure).", "votes": null }, { "id": "759796", "postDate": "02/29/2020 13:25:36", "content": "And new experiments are taking the hella long time to train :(
", "rawMarkdown": "And new experiments are taking the hella long time to train :(", "votes": null }, { "id": "759859", "postDate": "02/29/2020 14:52:39", "content": "assuming the original data is [0,1,2,3,4...]\nbecause of striding, you only access data [0,2,4,...]
\n\nnow i upsize by 2 to get [0,0,1,1,2,2,3,3,4,4, ...]\nwith the same network, i can get data [0,1,2,3,4,...]
", "rawMarkdown": "assuming the original data is [0,1,2,3,4...]\nbecause of striding, you only access data [0,2,4,...]\n\nnow i upsize by 2 to get [0,0,1,1,2,2,3,3,4,4, ...]\nwith the same network, i can get data [0,1,2,3,4,...]", "votes": null }, { "id": "759862", "postDate": "02/29/2020 14:54:48", "content": "How much memory is required for this. On 64x64 for all training data keeping float32 I am getting memory error on 16 GB machine
", "rawMarkdown": "How much memory is required for this. On 64x64 for all training data keeping float32 I am getting memory error on 16 GB machine", "votes": null }, { "id": "759863", "postDate": "02/29/2020 14:55:13", "content": "there is a paper that did experiments on different test/train image sizes. you can search for it
", "rawMarkdown": "there is a paper that did experiments on different test/train image sizes. you can search for it", "votes": null }, { "id": "759924", "postDate": "02/29/2020 16:12:00", "content": "It might affect CutMix.\nThe image that is added over the base image is reduced in size and, as a consequence, in resolution. As such, the model might have difficulty dealing with the added image. If everything is scaled up beforehand, the added image would be closer to its original resolution and as such the model might be able to better deal with it. This is especially true as some graphemes take up only a small portion of the total image - if the resolution is reduced further, it might be very hard to deal with it.
", "rawMarkdown": "It might affect CutMix.\nThe image that is added over the base image is reduced in size and, as a consequence, in resolution. As such, the model might have difficulty dealing with the added image. If everything is scaled up beforehand, the added image would be closer to its original resolution and as such the model might be able to better deal with it. This is especially true as some graphemes take up only a small portion of the total image - if the resolution is reduced further, it might be very hard to deal with it.", "votes": null }, { "id": "760064", "postDate": "02/29/2020 19:58:21", "content": "Reduce batch size. Try batch size of 128. If you get error, then reduce it to 64
", "rawMarkdown": "Reduce batch size. Try batch size of 128. If you get error, then reduce it to 64", "votes": null }, { "id": "760088", "postDate": "02/29/2020 21:10:17", "content": "\n\n\nWell, it is no surprise now that we should NOT use cropped images.
\n
Thanks for confirming. That was ;y conclusion based on local validation, but not being a computer vision expert by any mean I was in doubt.
", "rawMarkdown": "> Well, it is no surprise now that we should NOT use cropped images.\n\nThanks for confirming. That was ;y conclusion based on local validation, but not being a computer vision expert by any mean I was in doubt.", "votes": null }, { "id": "760185", "postDate": "03/01/2020 00:12:54", "content": "@hengck23 There, I got that, thanks! Did not think of the innate nature of CNNs before. As helpful as ever
", "rawMarkdown": "hengck23 There, I got that, thanks! Did not think of the innate nature of CNNs before. As helpful as ever", "votes": null }, { "id": "760361", "postDate": "03/01/2020 07:15:34", "content": "Thanks @santosh16k . Its kinda helped
", "rawMarkdown": "Thanks @santosh16k . Its kinda helped", "votes": null }, { "id": "762984", "postDate": "03/04/2020 01:54:34", "content": "My results? 224x224 did not work for me. Experiments take 1.3x longer (at least) and does not yield significant improvements. Well I did not train until the end and simply early stopped the experiments where my original-resolution experiment achieved its best score.
", "rawMarkdown": "My results? 224x224 did not work for me. Experiments take 1.3x longer (at least) and does not yield significant improvements. Well I did not train until the end and simply early stopped the experiments where my original-resolution experiment achieved its best score.", "votes": null } ], "comments": [ { "id": 759457, "author_name": "hengck23", "author_url": "", "post_date": "02/29/2020 04:01:44", "content": "\"fundamentally not adding more information\"
\n\nit does, bucause of pooling and stride
", "votes": null, "replies": [] }, { "id": 759795, "author_name": "roguekk007", "author_url": "", "post_date": "02/29/2020 13:24:59", "content": "@hengck23 Not really getting what you mean there😁 . We are only creating pixels out of existing ones, so it is kind of like feature engineering. The only explanation I can come up for possible improvement is that CNN prefers square images and square feature maps (which sounds tenure).
", "votes": null, "replies": [ { "id": 759859, "author_name": "hengck23", "author_url": "", "post_date": "02/29/2020 14:52:39", "content": "assuming the original data is [0,1,2,3,4...]\nbecause of striding, you only access data [0,2,4,...]
\n\nnow i upsize by 2 to get [0,0,1,1,2,2,3,3,4,4, ...]\nwith the same network, i can get data [0,1,2,3,4,...]
", "votes": null, "replies": [] }, { "id": 759863, "author_name": "hengck23", "author_url": "", "post_date": "02/29/2020 14:55:13", "content": "there is a paper that did experiments on different test/train image sizes. you can search for it
", "votes": null, "replies": [] }, { "id": 760185, "author_name": "roguekk007", "author_url": "", "post_date": "03/01/2020 00:12:54", "content": "@hengck23 There, I got that, thanks! Did not think of the innate nature of CNNs before. As helpful as ever
", "votes": null, "replies": [] } ] }, { "id": 759796, "author_name": "roguekk007", "author_url": "", "post_date": "02/29/2020 13:25:36", "content": "And new experiments are taking the hella long time to train :(
", "votes": null, "replies": [] }, { "id": 759862, "author_name": "mks2192", "author_url": "", "post_date": "02/29/2020 14:54:48", "content": "How much memory is required for this. On 64x64 for all training data keeping float32 I am getting memory error on 16 GB machine
", "votes": null, "replies": [ { "id": 760064, "author_name": "santosh16k", "author_url": "", "post_date": "02/29/2020 19:58:21", "content": "Reduce batch size. Try batch size of 128. If you get error, then reduce it to 64
", "votes": null, "replies": [] }, { "id": 760361, "author_name": "mks2192", "author_url": "", "post_date": "03/01/2020 07:15:34", "content": "Thanks @santosh16k . Its kinda helped
", "votes": null, "replies": [] } ] }, { "id": 759924, "author_name": "larswigger", "author_url": "", "post_date": "02/29/2020 16:12:00", "content": "It might affect CutMix.\nThe image that is added over the base image is reduced in size and, as a consequence, in resolution. As such, the model might have difficulty dealing with the added image. If everything is scaled up beforehand, the added image would be closer to its original resolution and as such the model might be able to better deal with it. This is especially true as some graphemes take up only a small portion of the total image - if the resolution is reduced further, it might be very hard to deal with it.
", "votes": null, "replies": [] }, { "id": 760088, "author_name": "cpmpml", "author_url": "", "post_date": "02/29/2020 21:10:17", "content": "\n\n\nWell, it is no surprise now that we should NOT use cropped images.
\n
Thanks for confirming. That was ;y conclusion based on local validation, but not being a computer vision expert by any mean I was in doubt.
", "votes": null, "replies": [] }, { "id": 762984, "author_name": "roguekk007", "author_url": "", "post_date": "03/04/2020 01:54:34", "content": "My results? 224x224 did not work for me. Experiments take 1.3x longer (at least) and does not yield significant improvements. Well I did not train until the end and simply early stopped the experiments where my original-resolution experiment achieved its best score.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "759451": "Well, it is no surprise now that we should NOT use cropped images.\nMerely doing this and using OneCycle schedule gave me a ~.015 boost and pulled me out of bronze\n\nI noticed this comment, the 1st place's comments are always worth surveying\n\n\nWonder if anyone did experiments on this? It just seems counter-intuitive that resizing the image beyond original size (fundamentally not adding more information) might provide a boost to the model.", "759457": "\"fundamentally not adding more information\"\n\nit does, bucause of pooling and stride", "759795": "hengck23 Not really getting what you mean there😁 . We are only creating pixels out of existing ones, so it is kind of like feature engineering. The only explanation I can come up for possible improvement is that CNN prefers square images and square feature maps (which sounds tenure).", "759796": "And new experiments are taking the hella long time to train :(", "759859": "assuming the original data is [0,1,2,3,4...]\nbecause of striding, you only access data [0,2,4,...]\n\nnow i upsize by 2 to get [0,0,1,1,2,2,3,3,4,4, ...]\nwith the same network, i can get data [0,1,2,3,4,...]", "759862": "How much memory is required for this. On 64x64 for all training data keeping float32 I am getting memory error on 16 GB machine", "759863": "there is a paper that did experiments on different test/train image sizes. you can search for it", "759924": "It might affect CutMix.\nThe image that is added over the base image is reduced in size and, as a consequence, in resolution. As such, the model might have difficulty dealing with the added image. If everything is scaled up beforehand, the added image would be closer to its original resolution and as such the model might be able to better deal with it. This is especially true as some graphemes take up only a small portion of the total image - if the resolution is reduced further, it might be very hard to deal with it.", "760064": "Reduce batch size. Try batch size of 128. If you get error, then reduce it to 64", "760088": "> Well, it is no surprise now that we should NOT use cropped images.\n\nThanks for confirming. That was ;y conclusion based on local validation, but not being a computer vision expert by any mean I was in doubt.", "760185": "hengck23 There, I got that, thanks! Did not think of the innate nature of CNNs before. As helpful as ever", "760361": "Thanks @santosh16k . Its kinda helped", "762984": "My results? 224x224 did not work for me. Experiments take 1.3x longer (at least) and does not yield significant improvements. Well I did not train until the end and simply early stopped the experiments where my original-resolution experiment achieved its best score." }, "source": "meta" }