I am trying out some experiments on gtx 1070 that I have on my desktop. Seems like it will not be enough as on larger models such as resnet 101 I can only pass a single image in batch while using a faster rcnn. Looking at some of the notebooks where ensembles are being used, I was wondering what hardware/gpu provider is being used? Are people using collab or some ec2 instance with gpu?
", "messages": [ { "id": "888433", "postDate": "06/16/2020 10:55:55", "content": "I am trying out some experiments on gtx 1070 that I have on my desktop. Seems like it will not be enough as on larger models such as resnet 101 I can only pass a single image in batch while using a faster rcnn. Looking at some of the notebooks where ensembles are being used, I was wondering what hardware/gpu provider is being used? Are people using collab or some ec2 instance with gpu?
", "rawMarkdown": "I am trying out some experiments on gtx 1070 that I have on my desktop. Seems like it will not be enough as on larger models such as resnet 101 I can only pass a single image in batch while using a faster rcnn. Looking at some of the notebooks where ensembles are being used, I was wondering what hardware/gpu provider is being used? Are people using collab or some ec2 instance with gpu?", "votes": null }, { "id": "888728", "postDate": "06/16/2020 14:31:05", "content": "Amazon, Kaggle and other collab frameworks are hosted on servers that are mainly on TPU (Tensor Process Unit) which they can customize in terms of sharing TPU VRAM so that multiple people can share the same resources, unless you buy an Nvidia DGX station, there is no way to outperform them.
\n\nAs in terms of GPU, you have to look at those with lots of GB such as : \nRTX 8000: 48 GB VRAM, ~$5,500\nRTX 6000: 24 GB VRAM, ~$4,000\nTitan RTX: 24 GB VRAM, ~$2,500.
\n\nor the cheap ones such as:\nRTX 2080 Ti: 11 GB VRAM, ~$1,150\n*GTX 1080 Ti: 11 GB VRAM, ~$800 refurbished *RTX 2080: 8 GB VRAM ~$720\nRTX 2070: 8 GB VRAM, ~$500. *
\n\nIt's all about the GB of VRAM , you can see a clear correlation between VRAM and money $\nBest of luck to you sir.
", "rawMarkdown": "Amazon, Kaggle and other collab frameworks are hosted on servers that are mainly on TPU (Tensor Process Unit) which they can customize in terms of sharing TPU VRAM so that multiple people can share the same resources, unless you buy an Nvidia DGX station, there is no way to outperform them. \n\nAs in terms of GPU, you have to look at those with lots of GB such as : \nRTX 8000: 48 GB VRAM, ~$5,500\nRTX 6000: 24 GB VRAM, ~$4,000\nTitan RTX: 24 GB VRAM, ~$2,500.\n\nor the cheap ones such as:\nRTX 2080 Ti: 11 GB VRAM, ~$1,150\n*GTX 1080 Ti: 11 GB VRAM, ~$800 refurbished *RTX 2080: 8 GB VRAM ~$720\nRTX 2070: 8 GB VRAM, ~$500. *\n\nIt's all about the GB of VRAM , you can see a clear correlation between VRAM and money $\nBest of luck to you sir.", "votes": null }, { "id": "888939", "postDate": "06/16/2020 16:57:13", "content": "Thanks for your reply. My question more pointedly is the submissions that have scored high, are they using any special hardware or the same performance can be achieved on consumer grade hardware by being smart about which architecture is used?
", "rawMarkdown": "Thanks for your reply. My question more pointedly is the submissions that have scored high, are they using any special hardware or the same performance can be achieved on consumer grade hardware by being smart about which architecture is used?", "votes": null }, { "id": "888945", "postDate": "06/16/2020 17:05:40", "content": "I would have to say Architecture > Hardware always.\nYou can also break the machine learning into K-folds / batches. \nHaving a great TPU could be a blessing or a curse, you can overfit super fast if you don't know how to train with it.
", "rawMarkdown": "I would have to say Architecture > Hardware always.\nYou can also break the machine learning into K-folds / batches. \nHaving a great TPU could be a blessing or a curse, you can overfit super fast if you don't know how to train with it.", "votes": null }, { "id": "889041", "postDate": "06/16/2020 18:27:00", "content": "You can try these to train larger models:\n- Lower image resolution.\n- Mixed Precision Training https://pytorch.org/docs/stable/amp.html#torch.cuda.amp.GradScaler\n- Gradient accumulation https://link.medium.com/zR6l7F4Cl7
\n\nAlso you can try colab/kaggle kernels, they provide gpu with 16 GB of memory or cloud providers with multiple preemptible T4 gpus for cheap.
", "rawMarkdown": "You can try these to train larger models:\n- Lower image resolution.\n- Mixed Precision Training https://pytorch.org/docs/stable/amp.html#torch.cuda.amp.GradScaler\n- Gradient accumulation https://link.medium.com/zR6l7F4Cl7\n\nAlso you can try colab/kaggle kernels, they provide gpu with 16 GB of memory or cloud providers with multiple preemptible T4 gpus for cheap.", "votes": null }, { "id": "889098", "postDate": "06/16/2020 18:56:35", "content": "Good hardware will always be an advantage in Kaggle competitions but it is not as important here. The dataset is reasonably sized and pretty simple models can perform quite well. My current highest score is just a FasterRCNN model with resnet50 backbone. My suggestion is that you use Google Colab or perhaps invest in Colab Pro to have access to stronger hardware than a 2070.
", "rawMarkdown": "Good hardware will always be an advantage in Kaggle competitions but it is not as important here. The dataset is reasonably sized and pretty simple models can perform quite well. My current highest score is just a FasterRCNN model with resnet50 backbone. My suggestion is that you use Google Colab or perhaps invest in Colab Pro to have access to stronger hardware than a 2070.", "votes": null }, { "id": "889558", "postDate": "06/17/2020 02:26:54", "content": "Thanks
", "rawMarkdown": "Thanks", "votes": null }, { "id": "889565", "postDate": "06/17/2020 02:29:49", "content": "With my 1070, which is a pascal card mixed precision won't give me any speed up. But I think your suggestion about resizing can help. Thanks
", "rawMarkdown": "With my 1070, which is a pascal card mixed precision won't give me any speed up. But I think your suggestion about resizing can help. Thanks", "votes": null } ], "comments": [ { "id": 888728, "author_name": "alincijov", "author_url": "", "post_date": "06/16/2020 14:31:05", "content": "Amazon, Kaggle and other collab frameworks are hosted on servers that are mainly on TPU (Tensor Process Unit) which they can customize in terms of sharing TPU VRAM so that multiple people can share the same resources, unless you buy an Nvidia DGX station, there is no way to outperform them.
\n\nAs in terms of GPU, you have to look at those with lots of GB such as : \nRTX 8000: 48 GB VRAM, ~$5,500\nRTX 6000: 24 GB VRAM, ~$4,000\nTitan RTX: 24 GB VRAM, ~$2,500.
\n\nor the cheap ones such as:\nRTX 2080 Ti: 11 GB VRAM, ~$1,150\n*GTX 1080 Ti: 11 GB VRAM, ~$800 refurbished *RTX 2080: 8 GB VRAM ~$720\nRTX 2070: 8 GB VRAM, ~$500. *
\n\nIt's all about the GB of VRAM , you can see a clear correlation between VRAM and money $\nBest of luck to you sir.
", "votes": null, "replies": [ { "id": 888939, "author_name": "gunnvant", "author_url": "", "post_date": "06/16/2020 16:57:13", "content": "Thanks for your reply. My question more pointedly is the submissions that have scored high, are they using any special hardware or the same performance can be achieved on consumer grade hardware by being smart about which architecture is used?
", "votes": null, "replies": [] }, { "id": 888945, "author_name": "alincijov", "author_url": "", "post_date": "06/16/2020 17:05:40", "content": "I would have to say Architecture > Hardware always.\nYou can also break the machine learning into K-folds / batches. \nHaving a great TPU could be a blessing or a curse, you can overfit super fast if you don't know how to train with it.
", "votes": null, "replies": [] }, { "id": 889098, "author_name": "matthewmasters", "author_url": "", "post_date": "06/16/2020 18:56:35", "content": "Good hardware will always be an advantage in Kaggle competitions but it is not as important here. The dataset is reasonably sized and pretty simple models can perform quite well. My current highest score is just a FasterRCNN model with resnet50 backbone. My suggestion is that you use Google Colab or perhaps invest in Colab Pro to have access to stronger hardware than a 2070.
", "votes": null, "replies": [] }, { "id": 889558, "author_name": "gunnvant", "author_url": "", "post_date": "06/17/2020 02:26:54", "content": "Thanks
", "votes": null, "replies": [] } ] }, { "id": 889041, "author_name": "utsavnandi", "author_url": "", "post_date": "06/16/2020 18:27:00", "content": "You can try these to train larger models:\n- Lower image resolution.\n- Mixed Precision Training https://pytorch.org/docs/stable/amp.html#torch.cuda.amp.GradScaler\n- Gradient accumulation https://link.medium.com/zR6l7F4Cl7
\n\nAlso you can try colab/kaggle kernels, they provide gpu with 16 GB of memory or cloud providers with multiple preemptible T4 gpus for cheap.
", "votes": null, "replies": [ { "id": 889565, "author_name": "gunnvant", "author_url": "", "post_date": "06/17/2020 02:29:49", "content": "With my 1070, which is a pascal card mixed precision won't give me any speed up. But I think your suggestion about resizing can help. Thanks
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "888433": "I am trying out some experiments on gtx 1070 that I have on my desktop. Seems like it will not be enough as on larger models such as resnet 101 I can only pass a single image in batch while using a faster rcnn. Looking at some of the notebooks where ensembles are being used, I was wondering what hardware/gpu provider is being used? Are people using collab or some ec2 instance with gpu?", "888728": "Amazon, Kaggle and other collab frameworks are hosted on servers that are mainly on TPU (Tensor Process Unit) which they can customize in terms of sharing TPU VRAM so that multiple people can share the same resources, unless you buy an Nvidia DGX station, there is no way to outperform them. \n\nAs in terms of GPU, you have to look at those with lots of GB such as : \nRTX 8000: 48 GB VRAM, ~$5,500\nRTX 6000: 24 GB VRAM, ~$4,000\nTitan RTX: 24 GB VRAM, ~$2,500.\n\nor the cheap ones such as:\nRTX 2080 Ti: 11 GB VRAM, ~$1,150\n*GTX 1080 Ti: 11 GB VRAM, ~$800 refurbished *RTX 2080: 8 GB VRAM ~$720\nRTX 2070: 8 GB VRAM, ~$500. *\n\nIt's all about the GB of VRAM , you can see a clear correlation between VRAM and money $\nBest of luck to you sir.", "888939": "Thanks for your reply. My question more pointedly is the submissions that have scored high, are they using any special hardware or the same performance can be achieved on consumer grade hardware by being smart about which architecture is used?", "888945": "I would have to say Architecture > Hardware always.\nYou can also break the machine learning into K-folds / batches. \nHaving a great TPU could be a blessing or a curse, you can overfit super fast if you don't know how to train with it.", "889041": "You can try these to train larger models:\n- Lower image resolution.\n- Mixed Precision Training https://pytorch.org/docs/stable/amp.html#torch.cuda.amp.GradScaler\n- Gradient accumulation https://link.medium.com/zR6l7F4Cl7\n\nAlso you can try colab/kaggle kernels, they provide gpu with 16 GB of memory or cloud providers with multiple preemptible T4 gpus for cheap.", "889098": "Good hardware will always be an advantage in Kaggle competitions but it is not as important here. The dataset is reasonably sized and pretty simple models can perform quite well. My current highest score is just a FasterRCNN model with resnet50 backbone. My suggestion is that you use Google Colab or perhaps invest in Colab Pro to have access to stronger hardware than a 2070.", "889558": "Thanks", "889565": "With my 1070, which is a pascal card mixed precision won't give me any speed up. But I think your suggestion about resizing can help. Thanks" }, "source": "meta" }