I'm using image size of 1024x512 to train my model, and I intent to increase the image size for further training.
\nDoes anyone can share your experience of training model with larger input size? Will it make the model hard to converge? Does the LB/CV score increase or decrease? Please let me know, that will be really helpful:) Thx!
I'm using image size of 1024x512 to train my model, and I intent to increase the image size for further training.
\nDoes anyone can share your experience of training model with larger input size? Will it make the model hard to converge? Does the LB/CV score increase or decrease? Please let me know, that will be really helpful:) Thx!
The EfficientNet paper is a good reference to learn which parameters to increase to directly increase the performance of your model. Image size is one of these parameters. See: https://arxiv.org/pdf/1905.11946.pdf
", "rawMarkdown": "The EfficientNet paper is a good reference to learn which parameters to increase to directly increase the performance of your model. Image size is one of these parameters. See: https://arxiv.org/pdf/1905.11946.pdf", "votes": 5 }, { "id": 2129196, "postDate": "2023-02-04T12:02:54.030Z", "rawMarkdown": "", "votes": 1, "replies": [ { "id": 2129620, "postDate": "2023-02-04T18:53:50.203Z", "content": "Is this ChatGPT? Please mention it or answer the question more directly with the context of this competition. I looked through some of your latest posts and they also seem to use some LLM. It can come off as spam (in my opinion) since in situations like this it is not really helpful to the original poster.
", "rawMarkdown": "Is this ChatGPT? Please mention it or answer the question more directly with the context of this competition. I looked through some of your latest posts and they also seem to use some LLM. It can come off as spam (in my opinion) since in situations like this it is not really helpful to the original poster.", "votes": 2 } ] }, { "id": 2134821, "postDate": "2023-02-08T09:22:10.640Z", "content": "Hi @jay2333,
\nIn my experience, I've always gotten better results by increasing the size of my images. The raw scans are very large (a few thousand pixels). I started with 512x256 for my baseline then increased fairly quickly to 1024x512. The final dimensions I use are 1456x728. I kept the same aspect ratio of 1:2 while getting as close as possible to the 20GB limit of the datasets. I am convinced that with larger images it is possible to achieve higher scores. However, we then come up against a problem of storage and probably of time during the inference.
\nLast but not least, I think it's best to use a dataset with smaller size images and quality preprocessing rather than large sloppy images. I advise you to focus above all on the way you do your preprocessing.
", "rawMarkdown": "Hi @jay2333,\n\nIn my experience, I've always gotten better results by increasing the size of my images. The raw scans are very large (a few thousand pixels). I started with 512x256 for my baseline then increased fairly quickly to 1024x512. The final dimensions I use are 1456x728. I kept the same aspect ratio of 1:2 while getting as close as possible to the 20GB limit of the datasets. I am convinced that with larger images it is possible to achieve higher scores. However, we then come up against a problem of storage and probably of time during the inference.\n\nLast but not least, I think it's best to use a dataset with smaller size images and quality preprocessing rather than large sloppy images. I advise you to focus above all on the way you do your preprocessing.", "votes": 2, "replies": [ { "id": 2135171, "postDate": "2023-02-08T13:51:38.827Z", "content": "you able to store larger images and use mode then 20GB, please see https://www.kaggle.com/competitions/rsna-breast-cancer-detection/discussion/383747#2130155
", "rawMarkdown": "you able to store larger images and use mode then 20GB, please see https://www.kaggle.com/competitions/rsna-breast-cancer-detection/discussion/383747#2130155", "votes": 2 } ] }, { "id": 2129007, "postDate": "2023-02-04T08:48:38.683Z", "content": "TLDR with increasing image size model performance also increases, and seems 2048 is a good choice.
\nHere is good thread to read and you can find answers about im size https://www.kaggle.com/competitions/rsna-breast-cancer-detection/discussion/370333
Thank you guys all for replying me! I plan to process the dataset first, and I will try increasing image size after that!
", "rawMarkdown": "Thank you guys all for replying me! I plan to process the dataset first, and I will try increasing image size after that!" }, { "id": 2130468, "postDate": "2023-02-05T13:24:29.913Z", "rawMarkdown": "", "votes": -5, "isDeleted": true } ], "comments": [ { "id": 2129025, "author_name": "Ali Abdin", "author_url": "", "post_date": "2023-02-04T09:06:24.293000", "content": "The EfficientNet paper is a good reference to learn which parameters to increase to directly increase the performance of your model. Image size is one of these parameters. See: https://arxiv.org/pdf/1905.11946.pdf
", "votes": 5, "replies": [] }, { "id": 2129196, "author_name": "", "author_url": "", "post_date": "2023-02-04T12:02:54.030000", "content": "", "votes": 1, "replies": [ { "id": 2129620, "author_name": "outwrest", "author_url": "", "post_date": "2023-02-04T18:53:50.203000", "content": "Is this ChatGPT? Please mention it or answer the question more directly with the context of this competition. I looked through some of your latest posts and they also seem to use some LLM. It can come off as spam (in my opinion) since in situations like this it is not really helpful to the original poster.
", "votes": 2, "replies": [] } ] }, { "id": 2134821, "author_name": "Paul Bacher", "author_url": "", "post_date": "2023-02-08T09:22:10.640000", "content": "Hi @jay2333,
\nIn my experience, I've always gotten better results by increasing the size of my images. The raw scans are very large (a few thousand pixels). I started with 512x256 for my baseline then increased fairly quickly to 1024x512. The final dimensions I use are 1456x728. I kept the same aspect ratio of 1:2 while getting as close as possible to the 20GB limit of the datasets. I am convinced that with larger images it is possible to achieve higher scores. However, we then come up against a problem of storage and probably of time during the inference.
\nLast but not least, I think it's best to use a dataset with smaller size images and quality preprocessing rather than large sloppy images. I advise you to focus above all on the way you do your preprocessing.
", "votes": 2, "replies": [ { "id": 2135171, "author_name": "A.P.", "author_url": "", "post_date": "2023-02-08T13:51:38.827000", "content": "you able to store larger images and use mode then 20GB, please see https://www.kaggle.com/competitions/rsna-breast-cancer-detection/discussion/383747#2130155
", "votes": 2, "replies": [] } ] }, { "id": 2129007, "author_name": "A.P.", "author_url": "", "post_date": "2023-02-04T08:48:38.683000", "content": "TLDR with increasing image size model performance also increases, and seems 2048 is a good choice.
\nHere is good thread to read and you can find answers about im size https://www.kaggle.com/competitions/rsna-breast-cancer-detection/discussion/370333
Thank you guys all for replying me! I plan to process the dataset first, and I will try increasing image size after that!
", "votes": 0, "replies": [] }, { "id": 2130468, "author_name": "", "author_url": "", "post_date": "2023-02-05T13:24:29.913000", "content": "", "votes": -5, "replies": [] } ], "raw_markdown_by_id": { "2128901": "I'm using image size of 1024x512 to train my model, and I intent to increase the image size for further training.\nDoes anyone can share your experience of training model with larger input size? Will it make the model hard to converge? Does the LB/CV score increase or decrease? Please let me know, that will be really helpful:) Thx!", "2129025": "The EfficientNet paper is a good reference to learn which parameters to increase to directly increase the performance of your model. Image size is one of these parameters. See: https://arxiv.org/pdf/1905.11946.pdf", "2129196": "", "2134821": "Hi @jay2333,\n\nIn my experience, I've always gotten better results by increasing the size of my images. The raw scans are very large (a few thousand pixels). I started with 512x256 for my baseline then increased fairly quickly to 1024x512. The final dimensions I use are 1456x728. I kept the same aspect ratio of 1:2 while getting as close as possible to the 20GB limit of the datasets. I am convinced that with larger images it is possible to achieve higher scores. However, we then come up against a problem of storage and probably of time during the inference.\n\nLast but not least, I think it's best to use a dataset with smaller size images and quality preprocessing rather than large sloppy images. I advise you to focus above all on the way you do your preprocessing.", "2129007": "TLDR with increasing image size model performance also increases, and seems 2048 is a good choice.\nHere is good thread to read and you can find answers about im size https://www.kaggle.com/competitions/rsna-breast-cancer-detection/discussion/370333", "2140708": "Thank you guys all for replying me! I plan to process the dataset first, and I will try increasing image size after that!", "2130468": "" } }