I get ~0.955CV with 512 imgsize, but get lower CV with larger imgsize such as 640. Have you ever been in a similar situation? Are there adjustments needed to make when using larger images?
", "messages": [ { "id": 1184482, "postDate": "2021-02-03T14:47:26.613Z", "content": "I get ~0.955CV with 512 imgsize, but get lower CV with larger imgsize such as 640. Have you ever been in a similar situation? Are there adjustments needed to make when using larger images?
", "rawMarkdown": "I get ~0.955CV with 512 imgsize, but get lower CV with larger imgsize such as 640. Have you ever been in a similar situation? Are there adjustments needed to make when using larger images? ", "votes": 9 }, { "id": 1188204, "postDate": "2021-02-06T04:16:07.590Z", "content": "Try getting larger batchsize (using better machine) or accumulating gradients.
", "rawMarkdown": "Try getting larger batchsize (using better machine) or accumulating gradients.", "votes": 3 }, { "id": 1188391, "postDate": "2021-02-06T07:46:23.593Z", "content": "Others have mentioned good tips, but its also not guaranteed that bigger images yield better results. Yes you have more detail but with relatively few images it might actually be the case that you begin just fitting to small noise patterns in the image rather than the broader conditions occurring in the image.
", "rawMarkdown": "Others have mentioned good tips, but its also not guaranteed that bigger images yield better results. Yes you have more detail but with relatively few images it might actually be the case that you begin just fitting to small noise patterns in the image rather than the broader conditions occurring in the image. ", "votes": 1 }, { "id": 1188319, "postDate": "2021-02-06T06:14:03.180Z", "content": "If your batch size is small than 8, consider use group normalization instead of batch normalization.
\nIf you have the access to multiple gpus, try using sync batch normalization and distributed training.
\nAnd you can use gradient accumulation to get a more stable step.
@dldmw579 I suggest you try with your usual image resizing method along with image smoothing like Gaussian filter.
", "rawMarkdown": "@dldmw579 I suggest you try with your usual image resizing method along with image smoothing like Gaussian filter." }, { "id": 1186988, "postDate": "2021-02-05T06:40:29.730Z", "rawMarkdown": "", "votes": -4, "isDeleted": true } ], "comments": [ { "id": 1188204, "author_name": "sin", "author_url": "", "post_date": "2021-02-06T04:16:07.590000", "content": "Try getting larger batchsize (using better machine) or accumulating gradients.
", "votes": 3, "replies": [] }, { "id": 1188391, "author_name": "ryches", "author_url": "", "post_date": "2021-02-06T07:46:23.593000", "content": "Others have mentioned good tips, but its also not guaranteed that bigger images yield better results. Yes you have more detail but with relatively few images it might actually be the case that you begin just fitting to small noise patterns in the image rather than the broader conditions occurring in the image.
", "votes": 1, "replies": [] }, { "id": 1188319, "author_name": "sheep", "author_url": "", "post_date": "2021-02-06T06:14:03.180000", "content": "If your batch size is small than 8, consider use group normalization instead of batch normalization.
\nIf you have the access to multiple gpus, try using sync batch normalization and distributed training.
\nAnd you can use gradient accumulation to get a more stable step.
@dldmw579 I suggest you try with your usual image resizing method along with image smoothing like Gaussian filter.
", "votes": 0, "replies": [] }, { "id": 1186988, "author_name": "", "author_url": "", "post_date": "2021-02-05T06:40:29.730000", "content": "", "votes": -4, "replies": [] } ], "raw_markdown_by_id": { "1184482": "I get ~0.955CV with 512 imgsize, but get lower CV with larger imgsize such as 640. Have you ever been in a similar situation? Are there adjustments needed to make when using larger images? ", "1188204": "Try getting larger batchsize (using better machine) or accumulating gradients.", "1188391": "Others have mentioned good tips, but its also not guaranteed that bigger images yield better results. Yes you have more detail but with relatively few images it might actually be the case that you begin just fitting to small noise patterns in the image rather than the broader conditions occurring in the image. ", "1188319": "If your batch size is small than 8, consider use group normalization instead of batch normalization.\nIf you have the access to multiple gpus, try using sync batch normalization and distributed training.\nAnd you can use gradient accumulation to get a more stable step.", "1194900": "@dldmw579 I suggest you try with your usual image resizing method along with image smoothing like Gaussian filter.", "1186988": "" } }