Does everyone use 128*128 for their image size? I use 64*64 since it is too slow to train on 128*128. I think using a low resolution is important for people like me who have limited GPUs
", "messages": [ { "id": "730931", "postDate": "01/28/2020 06:10:23", "content": "Does everyone use 128*128 for their image size? I use 64*64 since it is too slow to train on 128*128. I think using a low resolution is important for people like me who have limited GPUs
", "rawMarkdown": "Does everyone use 128*128 for their image size? I use 64*64 since it is too slow to train on 128*128. I think using a low resolution is important for people like me who have limited GPUs", "votes": null }, { "id": "730944", "postDate": "01/28/2020 06:42:19", "content": "I use 64x64 too, since I have limited resources.
", "rawMarkdown": "I use 64x64 too, since I have limited resources.", "votes": null }, { "id": "732110", "postDate": "01/29/2020 13:57:49", "content": "Is your current score achieved by using 64x64 image size?\nIn my case, 224x224 image size is better than 128x128. I haven't tried 64x64 image size yet..
", "rawMarkdown": "Is your current score achieved by using 64x64 image size?\nIn my case, 224x224 image size is better than 128x128. I haven't tried 64x64 image size yet..", "votes": null }, { "id": "732239", "postDate": "01/29/2020 15:50:43", "content": "Yeah my current score is using 64x64
", "rawMarkdown": "Yeah my current score is using 64x64", "votes": null }, { "id": "732240", "postDate": "01/29/2020 15:51:35", "content": "Also I'm pretty sure increasing resolution gives you better score overall 64x64 just trains fast so i can do more experiments
", "rawMarkdown": "Also I'm pretty sure increasing resolution gives you better score overall 64x64 just trains fast so i can do more experiments", "votes": null }, { "id": "732612", "postDate": "01/30/2020 02:01:46", "content": "Thanks for great advice!!!
", "rawMarkdown": "Thanks for great advice!!!", "votes": null }, { "id": "732613", "postDate": "01/30/2020 02:04:01", "content": "No problem!
", "rawMarkdown": "No problem!", "votes": null }, { "id": "732755", "postDate": "01/30/2020 08:09:27", "content": "resizing to 68x118 (i.e. scale by 0.5) and using serexnext50 + some modification + several augmentation can give LB score of about 0.974 for single model / single fold
", "rawMarkdown": "resizing to 68x118 (i.e. scale by 0.5) and using serexnext50 + some modification + several augmentation can give LB score of about 0.974 for single model / single fold", "votes": null }, { "id": "733086", "postDate": "01/30/2020 16:22:29", "content": "That's really magic. With image size 64 I couldn't even get LB 0.95. May I ask what gives you the boost with such a small image size.
", "rawMarkdown": "That's really magic. With image size 64 I couldn't even get LB 0.95. May I ask what gives you the boost with such a small image size.", "votes": null }, { "id": "733172", "postDate": "01/30/2020 19:05:45", "content": "I used something similar to wide resnet with dropout and cutout on 64x64 images which gave my current score. I don't use any pretrained weights and fine tune the architecture myself
", "rawMarkdown": "I used something similar to wide resnet with dropout and cutout on 64x64 images which gave my current score. I don't use any pretrained weights and fine tune the architecture myself", "votes": null }, { "id": "733179", "postDate": "01/30/2020 19:23:46", "content": "Just to confirm, Is your current score obtained from a single fold for 64x64 ?
", "rawMarkdown": "Just to confirm, Is your current score obtained from a single fold for 64x64 ?", "votes": null }, { "id": "733195", "postDate": "01/30/2020 20:00:24", "content": "Yes trained without validation after tuning
", "rawMarkdown": "Yes trained without validation after tuning", "votes": null }, { "id": "733213", "postDate": "01/30/2020 20:33:36", "content": "Hi @shujun717 If you are really limited to 64by64 pixels you can also try taking different crops from each training image. Either according to a random pattern or some fixed pattern like Center, topleft, topright etc.\nIt might give you a further increase in score. That said...with your current position you're already doing a great job.
", "rawMarkdown": "Hi @shujun717 If you are really limited to 64by64 pixels you can also try taking different crops from each training image. Either according to a random pattern or some fixed pattern like Center, topleft, topright etc.\nIt might give you a further increase in score. That said...with your current position you're already doing a great job.", "votes": null }, { "id": "733216", "postDate": "01/30/2020 20:38:03", "content": "Thanks! Will try that
", "rawMarkdown": "Thanks! Will try that", "votes": null }, { "id": "738536", "postDate": "02/06/2020 16:54:31", "content": "why my resnext50 with 68x118 lb(0.9688) so lower, but my cv is 0.9789😹 @hengck23
", "rawMarkdown": "why my resnext50 with 68x118 lb(0.9688) so lower, but my cv is 0.9789😹 @hengck23", "votes": null }, { "id": "738793", "postDate": "02/07/2020 02:56:00", "content": "Well he owns 2 gtx 2080 😅
", "rawMarkdown": "Well he owns 2 gtx 2080 😅", "votes": null }, { "id": "738961", "postDate": "02/07/2020 08:09:19", "content": "how your score 0.974, recall or accuracy？
", "rawMarkdown": "how your score 0.974, recall or accuracy？", "votes": null } ], "comments": [ { "id": 730944, "author_name": "quandapro", "author_url": "", "post_date": "01/28/2020 06:42:19", "content": "I use 64x64 too, since I have limited resources.
", "votes": null, "replies": [] }, { "id": 732110, "author_name": "inoueu1", "author_url": "", "post_date": "01/29/2020 13:57:49", "content": "Is your current score achieved by using 64x64 image size?\nIn my case, 224x224 image size is better than 128x128. I haven't tried 64x64 image size yet..
", "votes": null, "replies": [ { "id": 732239, "author_name": "shujun717", "author_url": "", "post_date": "01/29/2020 15:50:43", "content": "Yeah my current score is using 64x64
", "votes": null, "replies": [] }, { "id": 732240, "author_name": "shujun717", "author_url": "", "post_date": "01/29/2020 15:51:35", "content": "Also I'm pretty sure increasing resolution gives you better score overall 64x64 just trains fast so i can do more experiments
", "votes": null, "replies": [] }, { "id": 732612, "author_name": "inoueu1", "author_url": "", "post_date": "01/30/2020 02:01:46", "content": "Thanks for great advice!!!
", "votes": null, "replies": [] }, { "id": 732613, "author_name": "shujun717", "author_url": "", "post_date": "01/30/2020 02:04:01", "content": "No problem!
", "votes": null, "replies": [] }, { "id": 733086, "author_name": "syoya1997", "author_url": "", "post_date": "01/30/2020 16:22:29", "content": "That's really magic. With image size 64 I couldn't even get LB 0.95. May I ask what gives you the boost with such a small image size.
", "votes": null, "replies": [] }, { "id": 733172, "author_name": "shujun717", "author_url": "", "post_date": "01/30/2020 19:05:45", "content": "I used something similar to wide resnet with dropout and cutout on 64x64 images which gave my current score. I don't use any pretrained weights and fine tune the architecture myself
", "votes": null, "replies": [] }, { "id": 733179, "author_name": "sachinprabhu", "author_url": "", "post_date": "01/30/2020 19:23:46", "content": "Just to confirm, Is your current score obtained from a single fold for 64x64 ?
", "votes": null, "replies": [] }, { "id": 733195, "author_name": "shujun717", "author_url": "", "post_date": "01/30/2020 20:00:24", "content": "Yes trained without validation after tuning
", "votes": null, "replies": [] } ] }, { "id": 732755, "author_name": "hengck23", "author_url": "", "post_date": "01/30/2020 08:09:27", "content": "resizing to 68x118 (i.e. scale by 0.5) and using serexnext50 + some modification + several augmentation can give LB score of about 0.974 for single model / single fold
", "votes": null, "replies": [ { "id": 738536, "author_name": "cswwp347724", "author_url": "", "post_date": "02/06/2020 16:54:31", "content": "why my resnext50 with 68x118 lb(0.9688) so lower, but my cv is 0.9789😹 @hengck23
", "votes": null, "replies": [] }, { "id": 738961, "author_name": "masterzj", "author_url": "", "post_date": "02/07/2020 08:09:19", "content": "how your score 0.974, recall or accuracy？
", "votes": null, "replies": [] } ] }, { "id": 733213, "author_name": "rsmits", "author_url": "", "post_date": "01/30/2020 20:33:36", "content": "Hi @shujun717 If you are really limited to 64by64 pixels you can also try taking different crops from each training image. Either according to a random pattern or some fixed pattern like Center, topleft, topright etc.\nIt might give you a further increase in score. That said...with your current position you're already doing a great job.
", "votes": null, "replies": [ { "id": 733216, "author_name": "shujun717", "author_url": "", "post_date": "01/30/2020 20:38:03", "content": "Thanks! Will try that
", "votes": null, "replies": [] }, { "id": 738793, "author_name": "quandapro", "author_url": "", "post_date": "02/07/2020 02:56:00", "content": "Well he owns 2 gtx 2080 😅
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "730931": "Does everyone use 128*128 for their image size? I use 64*64 since it is too slow to train on 128*128. I think using a low resolution is important for people like me who have limited GPUs", "730944": "I use 64x64 too, since I have limited resources.", "732110": "Is your current score achieved by using 64x64 image size?\nIn my case, 224x224 image size is better than 128x128. I haven't tried 64x64 image size yet..", "732239": "Yeah my current score is using 64x64", "732240": "Also I'm pretty sure increasing resolution gives you better score overall 64x64 just trains fast so i can do more experiments", "732612": "Thanks for great advice!!!", "732613": "No problem!", "732755": "resizing to 68x118 (i.e. scale by 0.5) and using serexnext50 + some modification + several augmentation can give LB score of about 0.974 for single model / single fold", "733086": "That's really magic. With image size 64 I couldn't even get LB 0.95. May I ask what gives you the boost with such a small image size.", "733172": "I used something similar to wide resnet with dropout and cutout on 64x64 images which gave my current score. I don't use any pretrained weights and fine tune the architecture myself", "733179": "Just to confirm, Is your current score obtained from a single fold for 64x64 ?", "733195": "Yes trained without validation after tuning", "733213": "Hi @shujun717 If you are really limited to 64by64 pixels you can also try taking different crops from each training image. Either according to a random pattern or some fixed pattern like Center, topleft, topright etc.\nIt might give you a further increase in score. That said...with your current position you're already doing a great job.", "733216": "Thanks! Will try that", "738536": "why my resnext50 with 68x118 lb(0.9688) so lower, but my cv is 0.9789😹 @hengck23", "738793": "Well he owns 2 gtx 2080 😅", "738961": "how your score 0.974, recall or accuracy？" }, "source": "meta" }