Due to the limitation of the local environment, I can only learn 256-size images, is there a big difference from learning 512-size images?
", "messages": [ { "id": "1195929", "postDate": "02/11/2021 06:24:28", "content": "Due to the limitation of the local environment, I can only learn 256-size images, is there a big difference from learning 512-size images?
", "rawMarkdown": "Due to the limitation of the local environment, I can only learn 256-size images, is there a big difference from learning 512-size images?", "votes": null }, { "id": "1196051", "postDate": "02/11/2021 08:13:42", "content": "I haven't tried but if you want to learn with 256 size. You can try to use ViT with 224 and see if it achieves good result.
", "rawMarkdown": "I haven't tried but if you want to learn with 256 size. You can try to use ViT with 224 and see if it achieves good result.", "votes": null }, { "id": "1196346", "postDate": "02/11/2021 11:33:46", "content": "Pearsonly, I can achieve 0.9 CV with a resnet200d model at image size 384. while scaling up to 800, CV jumped to 0.902. But I haven't submit these run to the leaderboard.
\nI guess the boost of image size is not that significant, but if you want to stand at the head of the both LB, larger resolution may required.
Hello!
\nUnfortunately, I cannot provide any direct links (as those topics are now buried somwhere in the forum depths) but numerous users here reported better results with higher rezolution (at least 384x384, 512x512 is probably best).
\nI suggest working on Kaggle, not on your local machine and access powerful resources (GPUs and TPUs), as there are more computation greedy competitions which you may not be able to enter at all. E.g. you can look through this community notebook for a quick start with TPU.
", "rawMarkdown": "Hello!\n\nUnfortunately, I cannot provide any direct links (as those topics are now buried somwhere in the forum depths) but numerous users here reported better results with higher rezolution (at least 384x384, 512x512 is probably best). \n\nI suggest working on Kaggle, not on your local machine and access powerful resources (GPUs and TPUs), as there are more computation greedy competitions which you may not be able to enter at all. E.g. you can look through this **[community notebook](https://www.kaggle.com/jessemostipak/getting-started-tpus-cassava-leaf-disease)** for a quick start with TPU.", "votes": null }, { "id": "1196828", "postDate": "02/11/2021 17:28:36", "content": "Hi,
\nAll of our models consistently improved with 512-size images compared to 256. Additional to that, in the kaggle notebooks we were able to run effnet-b4 with 512 sized images.
", "rawMarkdown": "Hi,\n\nAll of our models consistently improved with 512-size images compared to 256. Additional to that, in the kaggle notebooks we were able to run effnet-b4 with 512 sized images.", "votes": null }, { "id": "1196868", "postDate": "02/11/2021 18:14:30", "content": "256사이즈로는 안돌려봤지만 384 사이즈에서 512 사이즈로 사이즈를 키웠을 시 점수가 상승했습니다.
\n그리고 구글 코랩이나, 캐글 환경을 사용하면 512 사이즈로 돌릴 수 있어요.
I am using only kaggle notebook. It is possible to make models with 512 size images.
\nI made train notebooks for each fold. e.g) note1 for fold0, note2 for fold1-2, note3 for fold3-4.
Hi,
\nIf you want to see the difference yourself, you can decrease the BATCH size, you have to find limits of the local environment, I have RTX2070 Super with 8Gb ram, when I train my model with 512x512, the max batch size is 16.
\nAnother solution is to set memory growth to disable (it is for TensorFlow), TensorFlow will allocate ram as available.
\n tf.config.experimental.set_memory_growth(gpu, False)\n
You can check the TensorFlow Limiting GPU memory growth guide for more info.
\n\n", "rawMarkdown": "Hi,\n\nIf you want to see the difference yourself, you can decrease the BATCH size, you have to find limits of the local environment, I have RTX2070 Super with 8Gb ram, when I train my model with 512x512, the max batch size is 16. \nAnother solution is to set memory growth to disable (it is for TensorFlow), TensorFlow will allocate ram as available.\n\n`\n tf.config.experimental.set_memory_growth(gpu, False)\n`\n\nYou can check the [TensorFlow Limiting GPU memory growth ](https://www.tensorflow.org/guide/gpu#limiting_gpu_memory_growth)guide for more info.\n\n> The first option is to turn on memory growth by calling tf.config.experimental.set_memory_growth, which attempts to allocate only as much GPU memory as needed for the runtime allocations", "votes": null } ], "comments": [ { "id": 1196051, "author_name": "tom88jerry", "author_url": "", "post_date": "02/11/2021 08:13:42", "content": "The first option is to turn on memory growth by calling tf.config.experimental.set_memory_growth, which attempts to allocate only as much GPU memory as needed for the runtime allocations
\n
I haven't tried but if you want to learn with 256 size. You can try to use ViT with 224 and see if it achieves good result.
", "votes": null, "replies": [] }, { "id": 1196346, "author_name": "steamedsheep", "author_url": "", "post_date": "02/11/2021 11:33:46", "content": "Pearsonly, I can achieve 0.9 CV with a resnet200d model at image size 384. while scaling up to 800, CV jumped to 0.902. But I haven't submit these run to the leaderboard.
\nI guess the boost of image size is not that significant, but if you want to stand at the head of the both LB, larger resolution may required.
Hello!
\nUnfortunately, I cannot provide any direct links (as those topics are now buried somwhere in the forum depths) but numerous users here reported better results with higher rezolution (at least 384x384, 512x512 is probably best).
\nI suggest working on Kaggle, not on your local machine and access powerful resources (GPUs and TPUs), as there are more computation greedy competitions which you may not be able to enter at all. E.g. you can look through this community notebook for a quick start with TPU.
", "votes": null, "replies": [] }, { "id": 1196828, "author_name": "snnclsr", "author_url": "", "post_date": "02/11/2021 17:28:36", "content": "Hi,
\nAll of our models consistently improved with 512-size images compared to 256. Additional to that, in the kaggle notebooks we were able to run effnet-b4 with 512 sized images.
", "votes": null, "replies": [] }, { "id": 1196868, "author_name": "deepkim", "author_url": "", "post_date": "02/11/2021 18:14:30", "content": "256사이즈로는 안돌려봤지만 384 사이즈에서 512 사이즈로 사이즈를 키웠을 시 점수가 상승했습니다.
\n그리고 구글 코랩이나, 캐글 환경을 사용하면 512 사이즈로 돌릴 수 있어요.
I am using only kaggle notebook. It is possible to make models with 512 size images.
\nI made train notebooks for each fold. e.g) note1 for fold0, note2 for fold1-2, note3 for fold3-4.
Hi,
\nIf you want to see the difference yourself, you can decrease the BATCH size, you have to find limits of the local environment, I have RTX2070 Super with 8Gb ram, when I train my model with 512x512, the max batch size is 16.
\nAnother solution is to set memory growth to disable (it is for TensorFlow), TensorFlow will allocate ram as available.
\n tf.config.experimental.set_memory_growth(gpu, False)\n
You can check the TensorFlow Limiting GPU memory growth guide for more info.
\n\n", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1195929": "Due to the limitation of the local environment, I can only learn 256-size images, is there a big difference from learning 512-size images?", "1196051": "I haven't tried but if you want to learn with 256 size. You can try to use ViT with 224 and see if it achieves good result.", "1196346": "Pearsonly, I can achieve 0.9 CV with a resnet200d model at image size 384. while scaling up to 800, CV jumped to 0.902. But I haven't submit these run to the leaderboard. \nI guess the boost of image size is not that significant, but if you want to stand at the head of the both LB, larger resolution may required.", "1196369": "Hello!\n\nUnfortunately, I cannot provide any direct links (as those topics are now buried somwhere in the forum depths) but numerous users here reported better results with higher rezolution (at least 384x384, 512x512 is probably best). \n\nI suggest working on Kaggle, not on your local machine and access powerful resources (GPUs and TPUs), as there are more computation greedy competitions which you may not be able to enter at all. E.g. you can look through this **[community notebook](https://www.kaggle.com/jessemostipak/getting-started-tpus-cassava-leaf-disease)** for a quick start with TPU.", "1196828": "Hi,\n\nAll of our models consistently improved with 512-size images compared to 256. Additional to that, in the kaggle notebooks we were able to run effnet-b4 with 512 sized images.", "1196868": "256사이즈로는 안돌려봤지만 384 사이즈에서 512 사이즈로 사이즈를 키웠을 시 점수가 상승했습니다.\n그리고 구글 코랩이나, 캐글 환경을 사용하면 512 사이즈로 돌릴 수 있어요.", "1197367": "I am using only kaggle notebook. It is possible to make models with 512 size images.\nI made train notebooks for each fold. e.g) note1 for fold0, note2 for fold1-2, note3 for fold3-4.", "1197745": "Hi,\n\nIf you want to see the difference yourself, you can decrease the BATCH size, you have to find limits of the local environment, I have RTX2070 Super with 8Gb ram, when I train my model with 512x512, the max batch size is 16. \nAnother solution is to set memory growth to disable (it is for TensorFlow), TensorFlow will allocate ram as available.\n\n`\n tf.config.experimental.set_memory_growth(gpu, False)\n`\n\nYou can check the [TensorFlow Limiting GPU memory growth ](https://www.tensorflow.org/guide/gpu#limiting_gpu_memory_growth)guide for more info.\n\n> The first option is to turn on memory growth by calling tf.config.experimental.set_memory_growth, which attempts to allocate only as much GPU memory as needed for the runtime allocations" }, "source": "meta" }The first option is to turn on memory growth by calling tf.config.experimental.set_memory_growth, which attempts to allocate only as much GPU memory as needed for the runtime allocations
\n