Hello everyone, this is my first Kaggle competition and I'm really excited! A doubt that I got however is about the fairness in using external resources. I've read some other topics and the rules of this competition and can't find anything about it.
\nFor example, I know that I can create a notebook that just load a model uploaded by me and submit the results. But with this approach, I could simply use a local machine, with an unlimited running time and maybe a cluster of GPUs to get results that a simple notebook from Kaggle never would reach.
\nThere's something in the rules that prevents this? Maybe my final submission must be a notebook that is trained from zero or something like that? Thank you for your help!!
", "messages": [ { "id": "1107773", "postDate": "12/10/2020 00:19:10", "content": "Hello everyone, this is my first Kaggle competition and I'm really excited! A doubt that I got however is about the fairness in using external resources. I've read some other topics and the rules of this competition and can't find anything about it.
\nFor example, I know that I can create a notebook that just load a model uploaded by me and submit the results. But with this approach, I could simply use a local machine, with an unlimited running time and maybe a cluster of GPUs to get results that a simple notebook from Kaggle never would reach.
\nThere's something in the rules that prevents this? Maybe my final submission must be a notebook that is trained from zero or something like that? Thank you for your help!!
", "rawMarkdown": "Hello everyone, this is my first Kaggle competition and I'm really excited! A doubt that I got however is about the fairness in using external resources. I've read some other topics and the rules of this competition and can't find anything about it. \n\nFor example, I know that I can create a notebook that just load a model uploaded by me and submit the results. But with this approach, I could simply use a local machine, with an unlimited running time and maybe a cluster of GPUs to get results that a simple notebook from Kaggle never would reach. \n\nThere's something in the rules that prevents this? Maybe my final submission must be a notebook that is trained from zero or something like that? Thank you for your help!!", "votes": null }, { "id": "1107808", "postDate": "12/10/2020 01:25:46", "content": "You can use any available resources (Kaggle, paid cloud services, your own hardware) that you have access to for training.
\nKaggle gives you 30+ hours of GPU a week and 30 hours of TPU a week. You can get a lot done with just those resources.
\nIn this contest, inference (predictions) must happen within the Kaggle notebook with resource limitations.
\n-Rich
", "rawMarkdown": "You can use any available resources (Kaggle, paid cloud services, your own hardware) that you have access to for training.\n\nKaggle gives you 30+ hours of GPU a week and 30 hours of TPU a week. You can get a lot done with just those resources.\n\nIn this contest, inference (predictions) must happen within the Kaggle notebook with resource limitations.\n\n-Rich", "votes": null }, { "id": "1107811", "postDate": "12/10/2020 01:27:13", "content": "\"I could simply use a local machine, with an unlimited running time and maybe a cluster of GPUs to get results that a simple notebook from Kaggle never would reach.\"
\nthis is allowed.
\nhaving a supercomputer is a competitive advantage.
\ndata science = \"machine + algorithm + data + skill (people)\".
\nkaggle is a race in all of them. how to get a \"good machine\" is \"part of the competition\".
Thanks for your response!
", "rawMarkdown": "Thanks for your response!", "votes": null }, { "id": "1107819", "postDate": "12/10/2020 01:32:29", "content": "Got it! Thank you for your explanation.
", "rawMarkdown": "Got it! Thank you for your explanation.", "votes": null }, { "id": "1108552", "postDate": "12/10/2020 19:25:39", "content": "There's nothing preventing that.
\nI have to admit though that my main limitation is how fast I can actually implement ideas. I'm failing to keep my somewhat decent local machine (Intel i7-8700K 3.70GHz × 6 + GTX 1080 Ti) completely busy and I'm not using my whole GPU/TPU quota on Kaggle. I'm sure that for some of the amazing people I've had the pleasure of teaming up in the past resources would be the limiting factors. For me, it isn't really the main issue. Not all of us would be that much more efficient on a DGX-A100 (sure, it'd be nice if any training would finish super-fast, but I spend much more time figuring out how to get something implemented than training…). 😏
", "rawMarkdown": "There's nothing preventing that. \n\nI have to admit though that my main limitation is how fast I can actually implement ideas. I'm failing to keep my somewhat decent local machine (Intel i7-8700K 3.70GHz × 6 + GTX 1080 Ti) completely busy and I'm not using my whole GPU/TPU quota on Kaggle. I'm sure that for some of the amazing people I've had the pleasure of teaming up in the past resources would be the limiting factors. For me, it isn't really the main issue. Not all of us would be that much more efficient on a DGX-A100 (sure, it'd be nice if any training would finish super-fast, but I spend much more time figuring out how to get something implemented than training...). 😏", "votes": null }, { "id": "1108558", "postDate": "12/10/2020 19:36:18", "content": "Definitely think about solutions is not easy, but iterate and make mistakes fast is a key point to success. That's why I think having the right resources is a big advantage. Although, as you said, you can also join with a teammate that has the computational resources.
", "rawMarkdown": "Definitely think about solutions is not easy, but iterate and make mistakes fast is a key point to success. That's why I think having the right resources is a big advantage. Although, as you said, you can also join with a teammate that has the computational resources.", "votes": null }, { "id": "1108581", "postDate": "12/10/2020 19:54:40", "content": "If you can use TPU really well,
\nI can say you have more resources than me.
p.s. TPU Star prize is good option for you.
@piantic What do you mean? Know how to train the model with TPU's? I definitely think that knowledge is more important than resources!
", "rawMarkdown": "piantic What do you mean? Know how to train the model with TPU's? I definitely think that knowledge is more important than resources!", "votes": null } ], "comments": [ { "id": 1107808, "author_name": "richardepstein", "author_url": "", "post_date": "12/10/2020 01:25:46", "content": "You can use any available resources (Kaggle, paid cloud services, your own hardware) that you have access to for training.
\nKaggle gives you 30+ hours of GPU a week and 30 hours of TPU a week. You can get a lot done with just those resources.
\nIn this contest, inference (predictions) must happen within the Kaggle notebook with resource limitations.
\n-Rich
", "votes": null, "replies": [ { "id": 1107818, "author_name": "alvarole", "author_url": "", "post_date": "12/10/2020 01:31:53", "content": "Thanks for your response!
", "votes": null, "replies": [] } ] }, { "id": 1107811, "author_name": "hengck23", "author_url": "", "post_date": "12/10/2020 01:27:13", "content": "\"I could simply use a local machine, with an unlimited running time and maybe a cluster of GPUs to get results that a simple notebook from Kaggle never would reach.\"
\nthis is allowed.
\nhaving a supercomputer is a competitive advantage.
\ndata science = \"machine + algorithm + data + skill (people)\".
\nkaggle is a race in all of them. how to get a \"good machine\" is \"part of the competition\".
Got it! Thank you for your explanation.
", "votes": null, "replies": [] } ] }, { "id": 1108552, "author_name": "bjoernholzhauer", "author_url": "", "post_date": "12/10/2020 19:25:39", "content": "There's nothing preventing that.
\nI have to admit though that my main limitation is how fast I can actually implement ideas. I'm failing to keep my somewhat decent local machine (Intel i7-8700K 3.70GHz × 6 + GTX 1080 Ti) completely busy and I'm not using my whole GPU/TPU quota on Kaggle. I'm sure that for some of the amazing people I've had the pleasure of teaming up in the past resources would be the limiting factors. For me, it isn't really the main issue. Not all of us would be that much more efficient on a DGX-A100 (sure, it'd be nice if any training would finish super-fast, but I spend much more time figuring out how to get something implemented than training…). 😏
", "votes": null, "replies": [ { "id": 1108558, "author_name": "alvarole", "author_url": "", "post_date": "12/10/2020 19:36:18", "content": "Definitely think about solutions is not easy, but iterate and make mistakes fast is a key point to success. That's why I think having the right resources is a big advantage. Although, as you said, you can also join with a teammate that has the computational resources.
", "votes": null, "replies": [] } ] }, { "id": 1108581, "author_name": "piantic", "author_url": "", "post_date": "12/10/2020 19:54:40", "content": "If you can use TPU really well,
\nI can say you have more resources than me.
p.s. TPU Star prize is good option for you.
@piantic What do you mean? Know how to train the model with TPU's? I definitely think that knowledge is more important than resources!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1107773": "Hello everyone, this is my first Kaggle competition and I'm really excited! A doubt that I got however is about the fairness in using external resources. I've read some other topics and the rules of this competition and can't find anything about it. \n\nFor example, I know that I can create a notebook that just load a model uploaded by me and submit the results. But with this approach, I could simply use a local machine, with an unlimited running time and maybe a cluster of GPUs to get results that a simple notebook from Kaggle never would reach. \n\nThere's something in the rules that prevents this? Maybe my final submission must be a notebook that is trained from zero or something like that? Thank you for your help!!", "1107808": "You can use any available resources (Kaggle, paid cloud services, your own hardware) that you have access to for training.\n\nKaggle gives you 30+ hours of GPU a week and 30 hours of TPU a week. You can get a lot done with just those resources.\n\nIn this contest, inference (predictions) must happen within the Kaggle notebook with resource limitations.\n\n-Rich", "1107811": "\"I could simply use a local machine, with an unlimited running time and maybe a cluster of GPUs to get results that a simple notebook from Kaggle never would reach.\"\n\n\nthis is allowed.\n\nhaving a supercomputer is a competitive advantage.\n\ndata science = \"machine + algorithm + data + skill (people)\". \nkaggle is a race in all of them. how to get a \"good machine\" is \"part of the competition\".", "1107818": "Thanks for your response!", "1107819": "Got it! Thank you for your explanation.", "1108552": "There's nothing preventing that. \n\nI have to admit though that my main limitation is how fast I can actually implement ideas. I'm failing to keep my somewhat decent local machine (Intel i7-8700K 3.70GHz × 6 + GTX 1080 Ti) completely busy and I'm not using my whole GPU/TPU quota on Kaggle. I'm sure that for some of the amazing people I've had the pleasure of teaming up in the past resources would be the limiting factors. For me, it isn't really the main issue. Not all of us would be that much more efficient on a DGX-A100 (sure, it'd be nice if any training would finish super-fast, but I spend much more time figuring out how to get something implemented than training...). 😏", "1108558": "Definitely think about solutions is not easy, but iterate and make mistakes fast is a key point to success. That's why I think having the right resources is a big advantage. Although, as you said, you can also join with a teammate that has the computational resources.", "1108581": "If you can use TPU really well, \nI can say you have more resources than me.\n\np.s. `TPU Star` prize is good option for you.\n\n@alvarole", "1108613": "piantic What do you mean? Know how to train the model with TPU's? I definitely think that knowledge is more important than resources!" }, "source": "meta" }