I have i7 3820 and gtx 660ti 2gb - so even with batch size = 4 on ubuntu using Tensorflow GPU backend I was getting out of memory error after ~ 7 epochs on Unet128. And each epoch took me around 19 mins.
\n\nSo I'v discovered that my options:
\n\nHope it helps someone.
", "messages": [ { "id": "210640", "postDate": "08/06/2017 11:37:36", "content": "I have i7 3820 and gtx 660ti 2gb - so even with batch size = 4 on ubuntu using Tensorflow GPU backend I was getting out of memory error after ~ 7 epochs on Unet128. And each epoch took me around 19 mins.
\n\nSo I'v discovered that my options:
\n\nHope it helps someone.
", "rawMarkdown": "I have i7 3820 and gtx 660ti 2gb - so even with batch size = 4 on ubuntu using Tensorflow GPU backend I was getting out of memory error after ~ 7 epochs on Unet128. And each epoch took me around 19 mins. \n\nSo I'v discovered that my options:\n\n 1. switch to Theono GPU, because it's better optimized fork Keras. \n 2. increase VRAM by RAM but it will be around ~3.3 times slower. VRAM has 192 bit and 3000MHz vs 128bit RAM with 1333 MHz = 3.3. And I haven't found any working solution\n 3. Tried to turn off X-server on ubuntu and it worked! Instead of 19mins now each epoch takes 13mins and there is no memory issues. I guess it's because in 3820 there is no integrated video card (in comparison to 3770) and X-server was consuming part of this.\n\nHope it helps someone.", "votes": null }, { "id": "211922", "postDate": "08/10/2017 08:34:28", "content": "why downvote?
", "rawMarkdown": "why downvote?", "votes": null }, { "id": "215582", "postDate": "08/22/2017 07:51:31", "content": "Thanks for writeup. I have GTX750Ti, and facing the same batch size limitation. Can you tell how to follow your second instruction of 'Increase VRAM by RAM'?
", "rawMarkdown": "Thanks for writeup. I have GTX750Ti, and facing the same batch size limitation. Can you tell how to follow your second instruction of 'Increase VRAM by RAM'?", "votes": null }, { "id": "215618", "postDate": "08/22/2017 11:04:43", "content": "Unfortunately I haven't found any practical guide how to implement it.
", "rawMarkdown": "Unfortunately I haven't found any practical guide how to implement it.", "votes": null }, { "id": "215650", "postDate": "08/22/2017 14:38:16", "content": "but you did it right? or those are just idea? \nIf you could do it, how? Adding more memory for the GPU to use would help a lot of Kagglers for sure
", "rawMarkdown": "but you did it right? or those are just idea? \nIf you could do it, how? Adding more memory for the GPU to use would help a lot of Kagglers for sure", "votes": null }, { "id": "215672", "postDate": "08/22/2017 15:50:37", "content": "I didn't do it. It was an idea to which I haven't found any certain answer. (In old times though seems like it was possible for agp gpus https://www.techwalla.com/articles/how-to-use-system-ram-as-video-ram).
\n\nHowever I found that you can use vram as ram (https://github.com/Overv/vramfs). So in case someone has 4gb of RAM and 10x Titans it will certainly help:)
", "rawMarkdown": "I didn't do it. It was an idea to which I haven't found any certain answer. (In old times though seems like it was possible for agp gpus https://www.techwalla.com/articles/how-to-use-system-ram-as-video-ram). \n\nHowever I found that you can use vram as ram (https://github.com/Overv/vramfs). So in case someone has 4gb of RAM and 10x Titans it will certainly help:)", "votes": null }, { "id": "540852", "postDate": "06/01/2019 08:24:48", "content": "hello Anton .\ni installed tensorflow following the steps in below link
\n\nhttps://blog.quantinsti.com/install-tensorflow-gpu/
\n\nI got a problem, that is my jupyter notebook is not working .
\n\ndo you know why and solution for this.
\n\nthank you
", "rawMarkdown": "hello Anton .\ni installed tensorflow following the steps in below link\n\nhttps://blog.quantinsti.com/install-tensorflow-gpu/\n\nI got a problem, that is my jupyter notebook is not working .\n\ndo you know why and solution for this.\n\nthank you", "votes": null } ], "comments": [ { "id": 211922, "author_name": "heyt0ny", "author_url": "", "post_date": "08/10/2017 08:34:28", "content": "why downvote?
", "votes": null, "replies": [] }, { "id": 215582, "author_name": "sainathadapa", "author_url": "", "post_date": "08/22/2017 07:51:31", "content": "Thanks for writeup. I have GTX750Ti, and facing the same batch size limitation. Can you tell how to follow your second instruction of 'Increase VRAM by RAM'?
", "votes": null, "replies": [ { "id": 215618, "author_name": "heyt0ny", "author_url": "", "post_date": "08/22/2017 11:04:43", "content": "Unfortunately I haven't found any practical guide how to implement it.
", "votes": null, "replies": [] }, { "id": 215650, "author_name": "cjansen", "author_url": "", "post_date": "08/22/2017 14:38:16", "content": "but you did it right? or those are just idea? \nIf you could do it, how? Adding more memory for the GPU to use would help a lot of Kagglers for sure
", "votes": null, "replies": [] }, { "id": 215672, "author_name": "heyt0ny", "author_url": "", "post_date": "08/22/2017 15:50:37", "content": "I didn't do it. It was an idea to which I haven't found any certain answer. (In old times though seems like it was possible for agp gpus https://www.techwalla.com/articles/how-to-use-system-ram-as-video-ram).
\n\nHowever I found that you can use vram as ram (https://github.com/Overv/vramfs). So in case someone has 4gb of RAM and 10x Titans it will certainly help:)
", "votes": null, "replies": [] } ] }, { "id": 540852, "author_name": "krishna8ds", "author_url": "", "post_date": "06/01/2019 08:24:48", "content": "hello Anton .\ni installed tensorflow following the steps in below link
\n\nhttps://blog.quantinsti.com/install-tensorflow-gpu/
\n\nI got a problem, that is my jupyter notebook is not working .
\n\ndo you know why and solution for this.
\n\nthank you
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "210640": "I have i7 3820 and gtx 660ti 2gb - so even with batch size = 4 on ubuntu using Tensorflow GPU backend I was getting out of memory error after ~ 7 epochs on Unet128. And each epoch took me around 19 mins. \n\nSo I'v discovered that my options:\n\n 1. switch to Theono GPU, because it's better optimized fork Keras. \n 2. increase VRAM by RAM but it will be around ~3.3 times slower. VRAM has 192 bit and 3000MHz vs 128bit RAM with 1333 MHz = 3.3. And I haven't found any working solution\n 3. Tried to turn off X-server on ubuntu and it worked! Instead of 19mins now each epoch takes 13mins and there is no memory issues. I guess it's because in 3820 there is no integrated video card (in comparison to 3770) and X-server was consuming part of this.\n\nHope it helps someone.", "211922": "why downvote?", "215582": "Thanks for writeup. I have GTX750Ti, and facing the same batch size limitation. Can you tell how to follow your second instruction of 'Increase VRAM by RAM'?", "215618": "Unfortunately I haven't found any practical guide how to implement it.", "215650": "but you did it right? or those are just idea? \nIf you could do it, how? Adding more memory for the GPU to use would help a lot of Kagglers for sure", "215672": "I didn't do it. It was an idea to which I haven't found any certain answer. (In old times though seems like it was possible for agp gpus https://www.techwalla.com/articles/how-to-use-system-ram-as-video-ram). \n\nHowever I found that you can use vram as ram (https://github.com/Overv/vramfs). So in case someone has 4gb of RAM and 10x Titans it will certainly help:)", "540852": "hello Anton .\ni installed tensorflow following the steps in below link\n\nhttps://blog.quantinsti.com/install-tensorflow-gpu/\n\nI got a problem, that is my jupyter notebook is not working .\n\ndo you know why and solution for this.\n\nthank you" }, "source": "meta" }