For those who is not experienced in CV like me, could anyone tell us what level of hardware one will need to at least run state-of-art network in an reasonable time?
\n\nI think there might be many individual competitors who are engineers,students or researchers from other fields like me, with no accessing to professional GPUs. If it take days to run one epoch, maybe it's better to just learn what others are doing than participating.
", "messages": [ { "id": "555529", "postDate": "06/19/2019 02:56:15", "content": "For those who is not experienced in CV like me, could anyone tell us what level of hardware one will need to at least run state-of-art network in an reasonable time?
\n\nI think there might be many individual competitors who are engineers,students or researchers from other fields like me, with no accessing to professional GPUs. If it take days to run one epoch, maybe it's better to just learn what others are doing than participating.
", "rawMarkdown": "For those who is not experienced in CV like me, could anyone tell us what level of hardware one will need to at least run state-of-art network in an reasonable time?\n\nI think there might be many individual competitors who are engineers,students or researchers from other fields like me, with no accessing to professional GPUs. If it take days to run one epoch, maybe it's better to just learn what others are doing than participating.", "votes": null }, { "id": "555542", "postDate": "06/19/2019 03:36:30", "content": "You basically have two ways of participating without powerful hardware
\n\nEither way you probably won't end first (it's pretty hard to beat pfdet 5m$ cluster) but you will learn a lot.
", "rawMarkdown": "You basically have two ways of participating without powerful hardware\n\n1. Wait for the Google cloud coupons or use the Google cloud trial credits. Google allows you to use p100 and v100 on the trial credits through requests to support. This competition will burn easily through 500$ of credits, especially if you are not well versed with the cloud dark lore.\n2. Use pretrained models to do inference and have the added value by post processing the results. Even inference will take you about 10h... But, once you have the raw predictions, post processing can be done reasonably fast.\n\nEither way you probably won't end first (it's pretty hard to beat pfdet 5m$ cluster) but you will learn a lot.", "votes": null }, { "id": "555611", "postDate": "06/19/2019 06:37:33", "content": "So this is an 'expensive' competition and hard for person like me to do bottom-up work.\nThe second choice seems quit feasible, thanks for pointing that out.
", "rawMarkdown": "So this is an 'expensive' competition and hard for person like me to do bottom-up work.\nThe second choice seems quit feasible, thanks for pointing that out.", "votes": null }, { "id": "555618", "postDate": "06/19/2019 06:46:28", "content": "I was just at your position a year ago... had a laptop with oldish i7, no gpu. I actually managed pretty well on the second path and eventually teamed with someone who had nice hardware and we got to 22nd place, so don't be discouraged....\nIt is a very good competition to learn the basics, and it does not attract many of the gods of kaggle so you have a chance at gold or silver.
", "rawMarkdown": "I was just at your position a year ago... had a laptop with oldish i7, no gpu. I actually managed pretty well on the second path and eventually teamed with someone who had nice hardware and we got to 22nd place, so don't be discouraged....\nIt is a very good competition to learn the basics, and it does not attract many of the gods of kaggle so you have a chance at gold or silver.", "votes": null }, { "id": "555788", "postDate": "06/19/2019 12:45:47", "content": "Really appreciate you could tell me this, feeling more confident now. Thanks
", "rawMarkdown": "Really appreciate you could tell me this, feeling more confident now. Thanks", "votes": null }, { "id": "556143", "postDate": "06/19/2019 22:29:56", "content": "Hello! You can start with the hardware provided by Kaggle, which is a Tesla P100(Recenly) and it's pretty powerful unless you need more than one GPU. Also, I think you can train on a smaller data set if you are really worry about your hardware power.\nThe original Alexnet is trained by two GTX 580 and they succeed. So don't worry about the hardware issue too much. Hope this will help :)
", "rawMarkdown": "Hello! You can start with the hardware provided by Kaggle, which is a Tesla P100(Recenly) and it's pretty powerful unless you need more than one GPU. Also, I think you can train on a smaller data set if you are really worry about your hardware power.\nThe original Alexnet is trained by two GTX 580 and they succeed. So don't worry about the hardware issue too much. Hope this will help :)", "votes": null }, { "id": "557433", "postDate": "06/21/2019 07:01:10", "content": "Thanks, your advice is quit helpful!
", "rawMarkdown": "Thanks, your advice is quit helpful!", "votes": null }, { "id": "564224", "postDate": "06/29/2019 06:41:37", "content": "ya but how do i get the train data in the kernel\nand whats the size of the train data \ni went to followed the links for downloading the data but it shows train_1 ,train_2 ? which one to download ? this is my first time , dont know how this things works
", "rawMarkdown": "ya but how do i get the train data in the kernel\nand whats the size of the train data \ni went to followed the links for downloading the data but it shows train_1 ,train_2 ? which one to download ? this is my first time , dont know how this things works", "votes": null } ], "comments": [ { "id": 555542, "author_name": "moshel", "author_url": "", "post_date": "06/19/2019 03:36:30", "content": "You basically have two ways of participating without powerful hardware
\n\nEither way you probably won't end first (it's pretty hard to beat pfdet 5m$ cluster) but you will learn a lot.
", "votes": null, "replies": [ { "id": 555611, "author_name": "littlebeandog", "author_url": "", "post_date": "06/19/2019 06:37:33", "content": "So this is an 'expensive' competition and hard for person like me to do bottom-up work.\nThe second choice seems quit feasible, thanks for pointing that out.
", "votes": null, "replies": [] }, { "id": 555618, "author_name": "moshel", "author_url": "", "post_date": "06/19/2019 06:46:28", "content": "I was just at your position a year ago... had a laptop with oldish i7, no gpu. I actually managed pretty well on the second path and eventually teamed with someone who had nice hardware and we got to 22nd place, so don't be discouraged....\nIt is a very good competition to learn the basics, and it does not attract many of the gods of kaggle so you have a chance at gold or silver.
", "votes": null, "replies": [] }, { "id": 555788, "author_name": "littlebeandog", "author_url": "", "post_date": "06/19/2019 12:45:47", "content": "Really appreciate you could tell me this, feeling more confident now. Thanks
", "votes": null, "replies": [] } ] }, { "id": 556143, "author_name": "peraktong", "author_url": "", "post_date": "06/19/2019 22:29:56", "content": "Hello! You can start with the hardware provided by Kaggle, which is a Tesla P100(Recenly) and it's pretty powerful unless you need more than one GPU. Also, I think you can train on a smaller data set if you are really worry about your hardware power.\nThe original Alexnet is trained by two GTX 580 and they succeed. So don't worry about the hardware issue too much. Hope this will help :)
", "votes": null, "replies": [ { "id": 557433, "author_name": "littlebeandog", "author_url": "", "post_date": "06/21/2019 07:01:10", "content": "Thanks, your advice is quit helpful!
", "votes": null, "replies": [] }, { "id": 564224, "author_name": "clivefernandes", "author_url": "", "post_date": "06/29/2019 06:41:37", "content": "ya but how do i get the train data in the kernel\nand whats the size of the train data \ni went to followed the links for downloading the data but it shows train_1 ,train_2 ? which one to download ? this is my first time , dont know how this things works
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "555529": "For those who is not experienced in CV like me, could anyone tell us what level of hardware one will need to at least run state-of-art network in an reasonable time?\n\nI think there might be many individual competitors who are engineers,students or researchers from other fields like me, with no accessing to professional GPUs. If it take days to run one epoch, maybe it's better to just learn what others are doing than participating.", "555542": "You basically have two ways of participating without powerful hardware\n\n1. Wait for the Google cloud coupons or use the Google cloud trial credits. Google allows you to use p100 and v100 on the trial credits through requests to support. This competition will burn easily through 500$ of credits, especially if you are not well versed with the cloud dark lore.\n2. Use pretrained models to do inference and have the added value by post processing the results. Even inference will take you about 10h... But, once you have the raw predictions, post processing can be done reasonably fast.\n\nEither way you probably won't end first (it's pretty hard to beat pfdet 5m$ cluster) but you will learn a lot.", "555611": "So this is an 'expensive' competition and hard for person like me to do bottom-up work.\nThe second choice seems quit feasible, thanks for pointing that out.", "555618": "I was just at your position a year ago... had a laptop with oldish i7, no gpu. I actually managed pretty well on the second path and eventually teamed with someone who had nice hardware and we got to 22nd place, so don't be discouraged....\nIt is a very good competition to learn the basics, and it does not attract many of the gods of kaggle so you have a chance at gold or silver.", "555788": "Really appreciate you could tell me this, feeling more confident now. Thanks", "556143": "Hello! You can start with the hardware provided by Kaggle, which is a Tesla P100(Recenly) and it's pretty powerful unless you need more than one GPU. Also, I think you can train on a smaller data set if you are really worry about your hardware power.\nThe original Alexnet is trained by two GTX 580 and they succeed. So don't worry about the hardware issue too much. Hope this will help :)", "557433": "Thanks, your advice is quit helpful!", "564224": "ya but how do i get the train data in the kernel\nand whats the size of the train data \ni went to followed the links for downloading the data but it shows train_1 ,train_2 ? which one to download ? this is my first time , dont know how this things works" }, "source": "meta" }