![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2234817%2Fec22b1a660156022aa137cd5986cf6a2%2Fsmall.gif?generation=1573010361978521&alt=media\")
As we know that the size of all images in this competition's dataset is 1400x2100. Sometimes, it is impossible to directly use these images as the input of our neural networks because some architectures need images that satisfy some conditions, such as divisible by 32, etc... So, I've seen many competitors resized the dataset's images before or during the training process. Obviously, the smaller resized images are, the higher information loss. But, we can train our models much faster with smaller ones. I think choosing the size of images is an important step which we should not ignore.
\n\nWhat size have you guys chosen? Mine is 384x576 ๐ .
", "messages": [ { "id": 666268, "postDate": "2019-11-06T00:24:11.423Z", "content": "As we know that the size of all images in this competition's dataset is 1400x2100. Sometimes, it is impossible to directly use these images as the input of our neural networks because some architectures need images that satisfy some conditions, such as divisible by 32, etc... So, I've seen many competitors resized the dataset's images before or during the training process. Obviously, the smaller resized images are, the higher information loss. But, we can train our models much faster with smaller ones. I think choosing the size of images is an important step which we should not ignore.
\n\nWhat size have you guys chosen? Mine is 384x576 ๐ .
", "rawMarkdown": "As we know that the size of all images in this competition's dataset is *1400x2100*. Sometimes, it is impossible to directly use these images as the input of our neural networks because some architectures need images that satisfy some conditions, such as divisible by *32*, etc... So, I've seen many competitors resized the dataset's images before or during the training process. Obviously, the smaller resized images are, the higher information loss. But, we can train our models much faster with smaller ones. I think choosing the size of images is an important step which we should not ignore.\n\nWhat size have you guys chosen? Mine is *384x576* ๐ .", "votes": 8 }, { "id": 666733, "postDate": "2019-11-06T12:49:46.480Z", "content": "I use (384 , 576) at first , it gives me 0.660 , then I think use (350, 525) directly will reduce the distortion , but it drop to 0.655 . At second , I think bigger img_size will gives me more information about mask's position , so I transfer to (416, 608) , it is still 0.655 . \nGod , what should I do ?
", "rawMarkdown": "I use (384 , 576) at first , it gives me 0.660 , then I think use (350, 525) directly will reduce the distortion , but it drop to 0.655 . At second , I think bigger img_size will gives me more information about mask's position , so I transfer to (416, 608) , it is still 0.655 . \nGod , what should I do ?", "votes": 1, "replies": [ { "id": 666807, "postDate": "2019-11-06T14:16:42.960Z", "content": "analyse the reason behind the score. number alone don't tell the whole story.
\n\n0.660 may be statistically close to 0.655, i.e. there is no significant difference between these two number
", "rawMarkdown": "analyse the reason behind the score. number alone don't tell the whole story.\n\n0.660 may be statistically close to 0.655, i.e. there is no significant difference between these two number", "votes": 1 }, { "id": 666829, "postDate": "2019-11-06T14:44:05.600Z", "content": "Yes , I admit I am a little sensitive about LB score , I am about to try a much bigger image size to prove my thoughts , if it is useful , I will post here
", "rawMarkdown": "Yes , I admit I am a little sensitive about LB score , I am about to try a much bigger image size to prove my thoughts , if it is useful , I will post here" }, { "id": 666840, "postDate": "2019-11-06T14:53:25.373Z", "content": "@xujingzhao Hope that bigger-size images can help you increase your LB score ๐
", "rawMarkdown": "@xujingzhao Hope that bigger-size images can help you increase your LB score ๐ " } ] }, { "id": 666344, "postDate": "2019-11-06T02:39:31.250Z", "content": "Mine is 32x64 :)
", "rawMarkdown": "Mine is 32x64 :)", "replies": [ { "id": 666362, "postDate": "2019-11-06T03:19:55.537Z", "content": "Segmentation model is like :P \n
I think he meant 320 x 640
", "rawMarkdown": "I think he meant 320 x 640", "votes": 1 }, { "id": 666377, "postDate": "2019-11-06T03:41:02.757Z", "content": "@khahuras Did you use 32x64 images to train segmentation models ๐ฌ ? I think it is possible to train classifiers with these images but segmentation models ๐ค . Or, do you mean 32x64 cropped images from the original?
", "rawMarkdown": "@khahuras Did you use 32x64 images to train segmentation models ๐ฌ ? I think it is possible to train classifiers with these images but segmentation models ๐ค . Or, do you mean 32x64 cropped images from the original?" }, { "id": 666408, "postDate": "2019-11-06T04:40:59.410Z", "content": "I starter a joke, which started a whole world wondering... ;)\nBtw, I use 320x480 for testing ideas, and will increase a little bit when final training.
", "rawMarkdown": "I starter a joke, which started a whole world wondering... ;)\nBtw, I use 320x480 for testing ideas, and will increase a little bit when final training.", "votes": 1 }, { "id": 666419, "postDate": "2019-11-06T04:50:41.910Z", "content": "@khahuras You really did shock me ๐
", "rawMarkdown": "@khahuras You really did shock me ๐ " }, { "id": 666431, "postDate": "2019-11-06T05:07:30.420Z", "content": "@phunghieu starting training on 320x512 and then for next 10 epochs scaling to 320x640 increase you model score. Start training on small image and then scale to larger images.
", "rawMarkdown": "@phunghieu starting training on 320x512 and then for next 10 epochs scaling to 320x640 increase you model score. Start training on small image and then scale to larger images.", "votes": 2 }, { "id": 666436, "postDate": "2019-11-06T05:11:54.713Z", "content": "@axel81 cool, I will try it. Thanks for sharing ๐ !
", "rawMarkdown": "@axel81 cool, I will try it. Thanks for sharing ๐ !" }, { "id": 666630, "postDate": "2019-11-06T09:45:03.947Z", "content": "@axel81 How do you change the input shape at runtime? Or do you just save the weights after first n epochs and then upload to the model with different input shape?
", "rawMarkdown": "@axel81 How do you change the input shape at runtime? Or do you just save the weights after first n epochs and then upload to the model with different input shape?", "votes": 1 }, { "id": 666728, "postDate": "2019-11-06T12:44:13.863Z", "content": "@lightnezzofbeing different epoch give the dataset different transform.Resize() , and model end with nn.AdptiveAvg2D()
\n\n@khahuras @phoenix9032 evertime I feel boring then I will come to kaggle , peoples are so funny here , I can't help laughing when I see your GIF .
", "rawMarkdown": "@lightnezzofbeing different epoch give the dataset different transform.Resize() , and model end with nn.AdptiveAvg2D()\n\n@khahuras @phoenix9032 evertime I feel boring then I will come to kaggle , peoples are so funny here , I can't help laughing when I see your GIF .\n\n", "votes": 3 }, { "id": 666907, "postDate": "2019-11-06T15:59:27.830Z", "content": "@lightnezzofbeing I checkpoint my model after every 10 epoch, where I decide based on loss/dice to change image size or not.
", "rawMarkdown": "@lightnezzofbeing I checkpoint my model after every 10 epoch, where I decide based on loss/dice to change image size or not.", "votes": 1 } ] }, { "id": 768962, "postDate": "2020-03-11T11:43:19.530Z", "content": "Thank you for sharing
", "rawMarkdown": "Thank you for sharing", "votes": 1, "isDeleted": true } ], "comments": [ { "id": 666733, "author_name": "ๅๅฐ็็งปๅจๅๅ ก", "author_url": "", "post_date": "2019-11-06T12:49:46.480000", "content": "I use (384 , 576) at first , it gives me 0.660 , then I think use (350, 525) directly will reduce the distortion , but it drop to 0.655 . At second , I think bigger img_size will gives me more information about mask's position , so I transfer to (416, 608) , it is still 0.655 . \nGod , what should I do ?
", "votes": 1, "replies": [ { "id": 666807, "author_name": "hengck23", "author_url": "", "post_date": "2019-11-06T14:16:42.960000", "content": "analyse the reason behind the score. number alone don't tell the whole story.
\n\n0.660 may be statistically close to 0.655, i.e. there is no significant difference between these two number
", "votes": 1, "replies": [] }, { "id": 666829, "author_name": "ๅๅฐ็็งปๅจๅๅ ก", "author_url": "", "post_date": "2019-11-06T14:44:05.600000", "content": "Yes , I admit I am a little sensitive about LB score , I am about to try a much bigger image size to prove my thoughts , if it is useful , I will post here
", "votes": 0, "replies": [] }, { "id": 666840, "author_name": "Hieu Phung", "author_url": "", "post_date": "2019-11-06T14:53:25.373000", "content": "@xujingzhao Hope that bigger-size images can help you increase your LB score ๐
", "votes": 0, "replies": [] } ] }, { "id": 666344, "author_name": "Kha Vo", "author_url": "", "post_date": "2019-11-06T02:39:31.250000", "content": "Mine is 32x64 :)
", "votes": 0, "replies": [ { "id": 666362, "author_name": "Nirjhar Roy", "author_url": "", "post_date": "2019-11-06T03:19:55.537000", "content": "Segmentation model is like :P \n
I think he meant 320 x 640
", "votes": 1, "replies": [] }, { "id": 666377, "author_name": "Hieu Phung", "author_url": "", "post_date": "2019-11-06T03:41:02.757000", "content": "@khahuras Did you use 32x64 images to train segmentation models ๐ฌ ? I think it is possible to train classifiers with these images but segmentation models ๐ค . Or, do you mean 32x64 cropped images from the original?
", "votes": 0, "replies": [] }, { "id": 666408, "author_name": "Kha Vo", "author_url": "", "post_date": "2019-11-06T04:40:59.410000", "content": "I starter a joke, which started a whole world wondering... ;)\nBtw, I use 320x480 for testing ideas, and will increase a little bit when final training.
", "votes": 1, "replies": [] }, { "id": 666419, "author_name": "Hieu Phung", "author_url": "", "post_date": "2019-11-06T04:50:41.910000", "content": "@khahuras You really did shock me ๐
", "votes": 0, "replies": [] }, { "id": 666431, "author_name": "Ram Ramrakhya", "author_url": "", "post_date": "2019-11-06T05:07:30.420000", "content": "@phunghieu starting training on 320x512 and then for next 10 epochs scaling to 320x640 increase you model score. Start training on small image and then scale to larger images.
", "votes": 2, "replies": [] }, { "id": 666436, "author_name": "Hieu Phung", "author_url": "", "post_date": "2019-11-06T05:11:54.713000", "content": "@axel81 cool, I will try it. Thanks for sharing ๐ !
", "votes": 0, "replies": [] }, { "id": 666630, "author_name": "Cyr1ll", "author_url": "", "post_date": "2019-11-06T09:45:03.947000", "content": "@axel81 How do you change the input shape at runtime? Or do you just save the weights after first n epochs and then upload to the model with different input shape?
", "votes": 1, "replies": [] }, { "id": 666728, "author_name": "ๅๅฐ็็งปๅจๅๅ ก", "author_url": "", "post_date": "2019-11-06T12:44:13.863000", "content": "@lightnezzofbeing different epoch give the dataset different transform.Resize() , and model end with nn.AdptiveAvg2D()
\n\n@khahuras @phoenix9032 evertime I feel boring then I will come to kaggle , peoples are so funny here , I can't help laughing when I see your GIF .
", "votes": 3, "replies": [] }, { "id": 666907, "author_name": "Ram Ramrakhya", "author_url": "", "post_date": "2019-11-06T15:59:27.830000", "content": "@lightnezzofbeing I checkpoint my model after every 10 epoch, where I decide based on loss/dice to change image size or not.
", "votes": 1, "replies": [] } ] }, { "id": 768962, "author_name": "", "author_url": "", "post_date": "2020-03-11T11:43:19.530000", "content": "Thank you for sharing
", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "666268": "As we know that the size of all images in this competition's dataset is *1400x2100*. Sometimes, it is impossible to directly use these images as the input of our neural networks because some architectures need images that satisfy some conditions, such as divisible by *32*, etc... So, I've seen many competitors resized the dataset's images before or during the training process. Obviously, the smaller resized images are, the higher information loss. But, we can train our models much faster with smaller ones. I think choosing the size of images is an important step which we should not ignore.\n\nWhat size have you guys chosen? Mine is *384x576* ๐ .", "666733": "I use (384 , 576) at first , it gives me 0.660 , then I think use (350, 525) directly will reduce the distortion , but it drop to 0.655 . At second , I think bigger img_size will gives me more information about mask's position , so I transfer to (416, 608) , it is still 0.655 . \nGod , what should I do ?", "666344": "Mine is 32x64 :)", "768962": "Thank you for sharing" } }