Does anybody have a problem with time amount for training your algorithm? We with my associate try to train toy Convolutional Neural Net and we spend a lot of time for even one epoches. I'm beginner in computer vision and I wonder, is it normal in this type of tasks and how I can to solve this problem? Maybe with GPU using, or maybe CNN can be training efficiantly only with server?Thank you.
", "messages": [ { "id": "380865", "postDate": "09/03/2018 15:41:28", "content": "Does anybody have a problem with time amount for training your algorithm? We with my associate try to train toy Convolutional Neural Net and we spend a lot of time for even one epoches. I'm beginner in computer vision and I wonder, is it normal in this type of tasks and how I can to solve this problem? Maybe with GPU using, or maybe CNN can be training efficiantly only with server?Thank you.
", "rawMarkdown": "Does anybody have a problem with time amount for training your algorithm? We with my associate try to train toy Convolutional Neural Net and we spend a lot of time for even one epoches. I'm beginner in computer vision and I wonder, is it normal in this type of tasks and how I can to solve this problem? Maybe with GPU using, or maybe CNN can be training efficiantly only with server?Thank you.", "votes": null }, { "id": "380979", "postDate": "09/03/2018 20:40:20", "content": "Training for cv without gpu is a very very long process. You will notice in the dataset an application for gcp credits. Use it wisely....
", "rawMarkdown": "Training for cv without gpu is a very very long process. You will notice in the dataset an application for gcp credits. Use it wisely....", "votes": null }, { "id": "380982", "postDate": "09/03/2018 20:45:20", "content": "You can use Kaggle kernals (free). There are some example CNN kernals that downsize the images to 256x256. They take about 10 minutes per epoch.
", "rawMarkdown": "You can use Kaggle kernals (free). There are some example CNN kernals that downsize the images to 256x256. They take about 10 minutes per epoch.", "votes": null }, { "id": "380991", "postDate": "09/03/2018 21:08:38", "content": "+1 to kaggle kernels, although it can be hard to set sometimes. From brief examination of the xrays, I don't think going down to 256*256 is a good idea. True, training will be much faster, but as the details are very fine, i think you will lose accuracy
", "rawMarkdown": "1 to kaggle kernels, although it can be hard to set sometimes. From brief examination of the xrays, I don't think going down to 256*256 is a good idea. True, training will be much faster, but as the details are very fine, i think you will lose accuracy", "votes": null }, { "id": "381005", "postDate": "09/03/2018 22:03:04", "content": "Maybe run first with 256*256, then threshold at a very low probability to find regions that might contain opacities, then re-run those regions at full resolution.
", "rawMarkdown": "Maybe run first with 256*256, then threshold at a very low probability to find regions that might contain opacities, then re-run those regions at full resolution.", "votes": null } ], "comments": [ { "id": 380979, "author_name": "moshel", "author_url": "", "post_date": "09/03/2018 20:40:20", "content": "Training for cv without gpu is a very very long process. You will notice in the dataset an application for gcp credits. Use it wisely....
", "votes": null, "replies": [] }, { "id": 380982, "author_name": "richardepstein", "author_url": "", "post_date": "09/03/2018 20:45:20", "content": "You can use Kaggle kernals (free). There are some example CNN kernals that downsize the images to 256x256. They take about 10 minutes per epoch.
", "votes": null, "replies": [] }, { "id": 380991, "author_name": "moshel", "author_url": "", "post_date": "09/03/2018 21:08:38", "content": "+1 to kaggle kernels, although it can be hard to set sometimes. From brief examination of the xrays, I don't think going down to 256*256 is a good idea. True, training will be much faster, but as the details are very fine, i think you will lose accuracy
", "votes": null, "replies": [ { "id": 381005, "author_name": "aharless", "author_url": "", "post_date": "09/03/2018 22:03:04", "content": "Maybe run first with 256*256, then threshold at a very low probability to find regions that might contain opacities, then re-run those regions at full resolution.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "380865": "Does anybody have a problem with time amount for training your algorithm? We with my associate try to train toy Convolutional Neural Net and we spend a lot of time for even one epoches. I'm beginner in computer vision and I wonder, is it normal in this type of tasks and how I can to solve this problem? Maybe with GPU using, or maybe CNN can be training efficiantly only with server?Thank you.", "380979": "Training for cv without gpu is a very very long process. You will notice in the dataset an application for gcp credits. Use it wisely....", "380982": "You can use Kaggle kernals (free). There are some example CNN kernals that downsize the images to 256x256. They take about 10 minutes per epoch.", "380991": "1 to kaggle kernels, although it can be hard to set sometimes. From brief examination of the xrays, I don't think going down to 256*256 is a good idea. True, training will be much faster, but as the details are very fine, i think you will lose accuracy", "381005": "Maybe run first with 256*256, then threshold at a very low probability to find regions that might contain opacities, then re-run those regions at full resolution." }, "source": "meta" }