Got no improvements for some trials. Wondering if any of you guys have tried using the full sized images and had a nice results
", "messages": [ { "id": "405148", "postDate": "10/17/2018 00:49:13", "content": "Got no improvements for some trials. Wondering if any of you guys have tried using the full sized images and had a nice results
", "rawMarkdown": "Got no improvements for some trials. Wondering if any of you guys have tried using the full sized images and had a nice results", "votes": null }, { "id": "405204", "postDate": "10/17/2018 03:51:31", "content": "I've not tried the 512x512 yet. I saw minimal changes from 384x384 down to 256x256. I expect as things get dialed in there may be some small gains from the higher resolution.
", "rawMarkdown": "I've not tried the 512x512 yet. I saw minimal changes from 384x384 down to 256x256. I expect as things get dialed in there may be some small gains from the higher resolution.", "votes": null }, { "id": "405690", "postDate": "10/17/2018 23:12:28", "content": "I’d suggest trying progressive resizing, so train first on 256x256 (or even down to 128x128), then train the same net on 384x384, then 512x512, etc.
", "rawMarkdown": "I’d suggest trying progressive resizing, so train first on 256x256 (or even down to 128x128), then train the same net on 384x384, then 512x512, etc.", "votes": null }, { "id": "406986", "postDate": "10/20/2018 05:43:20", "content": "I like to play around at the smaller resolutions to be sure things work, that I like the model, that it trains, etc. But near the end of the process I always use full size (if they fit at batch 2). While I have never had any great models in past challenges, I have always seen full size resolution perform better.
\n\nRunning my first decent model now at 256x256 - will run the same in a day or so at 512x512 and try to remember to come back here and document the size of the change.
", "rawMarkdown": "I like to play around at the smaller resolutions to be sure things work, that I like the model, that it trains, etc. But near the end of the process I always use full size (if they fit at batch 2). While I have never had any great models in past challenges, I have always seen full size resolution perform better.\n\nRunning my first decent model now at 256x256 - will run the same in a day or so at 512x512 and try to remember to come back here and document the size of the change.", "votes": null }, { "id": "407253", "postDate": "10/20/2018 17:59:30", "content": "I do this as well when trying out new models or loss functions. Scale it down to 64x64 if the network can run that low and run it quickly for several epochs to see how it is going to do.
", "rawMarkdown": "I do this as well when trying out new models or loss functions. Scale it down to 64x64 if the network can run that low and run it quickly for several epochs to see how it is going to do.", "votes": null }, { "id": "409253", "postDate": "10/24/2018 02:13:04", "content": "Did larger size a couple of different ways - cause I could not believe results. Three different models, larger size image always worse than smaller. I do need to reduce batch size to run the larger, but this was the case for lots of other challenges and models where larger resolution was always better.\nCannot decide what this is telling me. So guess i will try to go smaller than normal to see what happens. Something else different is that so far I am not fighting over fitting, Normally have seen this even on models in other challenges that score very poor.
", "rawMarkdown": "Did larger size a couple of different ways - cause I could not believe results. Three different models, larger size image always worse than smaller. I do need to reduce batch size to run the larger, but this was the case for lots of other challenges and models where larger resolution was always better.\nCannot decide what this is telling me. So guess i will try to go smaller than normal to see what happens. Something else different is that so far I am not fighting over fitting, Normally have seen this even on models in other challenges that score very poor.", "votes": null } ], "comments": [ { "id": 405204, "author_name": "ldm314", "author_url": "", "post_date": "10/17/2018 03:51:31", "content": "I've not tried the 512x512 yet. I saw minimal changes from 384x384 down to 256x256. I expect as things get dialed in there may be some small gains from the higher resolution.
", "votes": null, "replies": [] }, { "id": 405690, "author_name": "hortonhearsafoo", "author_url": "", "post_date": "10/17/2018 23:12:28", "content": "I’d suggest trying progressive resizing, so train first on 256x256 (or even down to 128x128), then train the same net on 384x384, then 512x512, etc.
", "votes": null, "replies": [] }, { "id": 406986, "author_name": "pcjimmmy", "author_url": "", "post_date": "10/20/2018 05:43:20", "content": "I like to play around at the smaller resolutions to be sure things work, that I like the model, that it trains, etc. But near the end of the process I always use full size (if they fit at batch 2). While I have never had any great models in past challenges, I have always seen full size resolution perform better.
\n\nRunning my first decent model now at 256x256 - will run the same in a day or so at 512x512 and try to remember to come back here and document the size of the change.
", "votes": null, "replies": [ { "id": 407253, "author_name": "ldm314", "author_url": "", "post_date": "10/20/2018 17:59:30", "content": "I do this as well when trying out new models or loss functions. Scale it down to 64x64 if the network can run that low and run it quickly for several epochs to see how it is going to do.
", "votes": null, "replies": [] } ] }, { "id": 409253, "author_name": "pcjimmmy", "author_url": "", "post_date": "10/24/2018 02:13:04", "content": "Did larger size a couple of different ways - cause I could not believe results. Three different models, larger size image always worse than smaller. I do need to reduce batch size to run the larger, but this was the case for lots of other challenges and models where larger resolution was always better.\nCannot decide what this is telling me. So guess i will try to go smaller than normal to see what happens. Something else different is that so far I am not fighting over fitting, Normally have seen this even on models in other challenges that score very poor.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "405148": "Got no improvements for some trials. Wondering if any of you guys have tried using the full sized images and had a nice results", "405204": "I've not tried the 512x512 yet. I saw minimal changes from 384x384 down to 256x256. I expect as things get dialed in there may be some small gains from the higher resolution.", "405690": "I’d suggest trying progressive resizing, so train first on 256x256 (or even down to 128x128), then train the same net on 384x384, then 512x512, etc.", "406986": "I like to play around at the smaller resolutions to be sure things work, that I like the model, that it trains, etc. But near the end of the process I always use full size (if they fit at batch 2). While I have never had any great models in past challenges, I have always seen full size resolution perform better.\n\nRunning my first decent model now at 256x256 - will run the same in a day or so at 512x512 and try to remember to come back here and document the size of the change.", "407253": "I do this as well when trying out new models or loss functions. Scale it down to 64x64 if the network can run that low and run it quickly for several epochs to see how it is going to do.", "409253": "Did larger size a couple of different ways - cause I could not believe results. Three different models, larger size image always worse than smaller. I do need to reduce batch size to run the larger, but this was the case for lots of other challenges and models where larger resolution was always better.\nCannot decide what this is telling me. So guess i will try to go smaller than normal to see what happens. Something else different is that so far I am not fighting over fitting, Normally have seen this even on models in other challenges that score very poor." }, "source": "meta" }