Is it possible to get a bit of insight on what kind of hardware specs do top teams on LB are using?
\n\nI hope others may have similar situation: I have only used kaggle kernels so far on kaggle and I am not an expert in setting up infrastructure for deep learning yet. But I have realized it is crucial skill to have as some competitions in recent times assume you to set up cloud instances and stuff to do anything meaningful.
\n\nFor someone who doesn't have a team is there any guide or set of instructions one can follow to at least get the right hardware up and running?
\n\nThanks
", "messages": [ { "id": "741450", "postDate": "02/10/2020 16:41:47", "content": "Is it possible to get a bit of insight on what kind of hardware specs do top teams on LB are using?
\n\nI hope others may have similar situation: I have only used kaggle kernels so far on kaggle and I am not an expert in setting up infrastructure for deep learning yet. But I have realized it is crucial skill to have as some competitions in recent times assume you to set up cloud instances and stuff to do anything meaningful.
\n\nFor someone who doesn't have a team is there any guide or set of instructions one can follow to at least get the right hardware up and running?
\n\nThanks
", "rawMarkdown": "Is it possible to get a bit of insight on what kind of hardware specs do top teams on LB are using?\n\nI hope others may have similar situation: I have only used kaggle kernels so far on kaggle and I am not an expert in setting up infrastructure for deep learning yet. But I have realized it is crucial skill to have as some competitions in recent times assume you to set up cloud instances and stuff to do anything meaningful. \n\nFor someone who doesn't have a team is there any guide or set of instructions one can follow to at least get the right hardware up and running?\n\nThanks", "votes": null }, { "id": "741455", "postDate": "02/10/2020 16:49:03", "content": "Try AWS sagemaker with a GPU based instance and large hard disk.
", "rawMarkdown": "Try AWS sagemaker with a GPU based instance and large hard disk.", "votes": null }, { "id": "741462", "postDate": "02/10/2020 17:12:49", "content": "I am using GCP with a free TPU, if you are interested how to set it up, I wrote an article some time ago.
", "rawMarkdown": "I am using GCP with a free TPU, if you are interested how to set it up, I wrote [an article](https://towardsdatascience.com/running-pytorch-on-tpu-a-bag-of-tricks-b6d0130bddd4) some time ago.", "votes": null }, { "id": "741987", "postDate": "02/11/2020 02:57:30", "content": "Kaggle notebook for training and inferencing(of course). AWS ec2 instance for preprocessing(e.g. cropping face).
", "rawMarkdown": "Kaggle notebook for training and inferencing(of course). AWS ec2 instance for preprocessing(e.g. cropping face).", "votes": null }, { "id": "742627", "postDate": "02/11/2020 12:26:57", "content": "Oh man! Kudos to you for being able to get that rank by just using kaggle notebook for training.
", "rawMarkdown": "Oh man! Kudos to you for being able to get that rank by just using kaggle notebook for training.", "votes": null }, { "id": "742689", "postDate": "02/11/2020 13:41:54", "content": "I use a single GPU (RTX Titan) on my home computer. Training takes ~ 6 hours.
", "rawMarkdown": "I use a single GPU (RTX Titan) on my home computer. Training takes ~ 6 hours.", "votes": null }, { "id": "742858", "postDate": "02/11/2020 15:47:16", "content": "I rely on AWS and am very grateful for their credits (wouldn't get this far without them)
", "rawMarkdown": "I rely on AWS and am very grateful for their credits (wouldn't get this far without them)", "votes": null }, { "id": "742878", "postDate": "02/11/2020 16:08:39", "content": "Colab Pro + GCP bucket for quick debugging +hyperparameter tuning ... then GCP workers over specific scenarios/experiments ... I'd recommend the Coursera GCP Advanced ML specialization on coursera (not the whole thing you'd skip to the relevant parts of course) :) .. and now TFRC TPUs thanks to Kaggle and Google :D
", "rawMarkdown": "Colab Pro + GCP bucket for quick debugging +hyperparameter tuning ... then GCP workers over specific scenarios/experiments ... I'd recommend the Coursera GCP Advanced ML specialization on coursera (not the whole thing you'd skip to the relevant parts of course) :) .. and now TFRC TPUs thanks to Kaggle and Google :D", "votes": null }, { "id": "742959", "postDate": "02/11/2020 17:03:31", "content": "Interesting ... Are you using / tried using mixed precision? the claims by Nvidia seem quite ambitious.
", "rawMarkdown": "Interesting ... Are you using / tried using mixed precision? the claims by Nvidia seem quite ambitious.", "votes": null }, { "id": "743344", "postDate": "02/12/2020 02:04:11", "content": "Yeah! I was trying to do a baseline by kaggle notebook(because their friendly UI) and turned out to be our best score!
", "rawMarkdown": "Yeah! I was trying to do a baseline by kaggle notebook(because their friendly UI) and turned out to be our best score!", "votes": null }, { "id": "748193", "postDate": "02/17/2020 09:03:39", "content": "I use my home computer with 2 Nvidia's GTX 1080Ti
", "rawMarkdown": "I use my home computer with 2 Nvidia's GTX 1080Ti", "votes": null } ], "comments": [ { "id": 741455, "author_name": "sciarrilli", "author_url": "", "post_date": "02/10/2020 16:49:03", "content": "Try AWS sagemaker with a GPU based instance and large hard disk.
", "votes": null, "replies": [] }, { "id": 741462, "author_name": "zaharch", "author_url": "", "post_date": "02/10/2020 17:12:49", "content": "I am using GCP with a free TPU, if you are interested how to set it up, I wrote an article some time ago.
", "votes": null, "replies": [] }, { "id": 741987, "author_name": "unkownhihi", "author_url": "", "post_date": "02/11/2020 02:57:30", "content": "Kaggle notebook for training and inferencing(of course). AWS ec2 instance for preprocessing(e.g. cropping face).
", "votes": null, "replies": [ { "id": 742627, "author_name": "akashnandi", "author_url": "", "post_date": "02/11/2020 12:26:57", "content": "Oh man! Kudos to you for being able to get that rank by just using kaggle notebook for training.
", "votes": null, "replies": [] }, { "id": 743344, "author_name": "unkownhihi", "author_url": "", "post_date": "02/12/2020 02:04:11", "content": "Yeah! I was trying to do a baseline by kaggle notebook(because their friendly UI) and turned out to be our best score!
", "votes": null, "replies": [] } ] }, { "id": 742689, "author_name": "jamesphoward", "author_url": "", "post_date": "02/11/2020 13:41:54", "content": "I use a single GPU (RTX Titan) on my home computer. Training takes ~ 6 hours.
", "votes": null, "replies": [ { "id": 742959, "author_name": "ma7moud", "author_url": "", "post_date": "02/11/2020 17:03:31", "content": "Interesting ... Are you using / tried using mixed precision? the claims by Nvidia seem quite ambitious.
", "votes": null, "replies": [] } ] }, { "id": 742858, "author_name": "rafiko1", "author_url": "", "post_date": "02/11/2020 15:47:16", "content": "I rely on AWS and am very grateful for their credits (wouldn't get this far without them)
", "votes": null, "replies": [] }, { "id": 742878, "author_name": "ma7moud", "author_url": "", "post_date": "02/11/2020 16:08:39", "content": "Colab Pro + GCP bucket for quick debugging +hyperparameter tuning ... then GCP workers over specific scenarios/experiments ... I'd recommend the Coursera GCP Advanced ML specialization on coursera (not the whole thing you'd skip to the relevant parts of course) :) .. and now TFRC TPUs thanks to Kaggle and Google :D
", "votes": null, "replies": [] }, { "id": 748193, "author_name": "ngcferreira", "author_url": "", "post_date": "02/17/2020 09:03:39", "content": "I use my home computer with 2 Nvidia's GTX 1080Ti
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "741450": "Is it possible to get a bit of insight on what kind of hardware specs do top teams on LB are using?\n\nI hope others may have similar situation: I have only used kaggle kernels so far on kaggle and I am not an expert in setting up infrastructure for deep learning yet. But I have realized it is crucial skill to have as some competitions in recent times assume you to set up cloud instances and stuff to do anything meaningful. \n\nFor someone who doesn't have a team is there any guide or set of instructions one can follow to at least get the right hardware up and running?\n\nThanks", "741455": "Try AWS sagemaker with a GPU based instance and large hard disk.", "741462": "I am using GCP with a free TPU, if you are interested how to set it up, I wrote [an article](https://towardsdatascience.com/running-pytorch-on-tpu-a-bag-of-tricks-b6d0130bddd4) some time ago.", "741987": "Kaggle notebook for training and inferencing(of course). AWS ec2 instance for preprocessing(e.g. cropping face).", "742627": "Oh man! Kudos to you for being able to get that rank by just using kaggle notebook for training.", "742689": "I use a single GPU (RTX Titan) on my home computer. Training takes ~ 6 hours.", "742858": "I rely on AWS and am very grateful for their credits (wouldn't get this far without them)", "742878": "Colab Pro + GCP bucket for quick debugging +hyperparameter tuning ... then GCP workers over specific scenarios/experiments ... I'd recommend the Coursera GCP Advanced ML specialization on coursera (not the whole thing you'd skip to the relevant parts of course) :) .. and now TFRC TPUs thanks to Kaggle and Google :D", "742959": "Interesting ... Are you using / tried using mixed precision? the claims by Nvidia seem quite ambitious.", "743344": "Yeah! I was trying to do a baseline by kaggle notebook(because their friendly UI) and turned out to be our best score!", "748193": "I use my home computer with 2 Nvidia's GTX 1080Ti" }, "source": "meta" }