It's unfair, I try to develop my idea and I find that I am restricted by my hardware, I have a modest gaming pc and I can not train my model or do transfer learning because of the low power of calculates or lack of RAM, even with the cloud that kaggle offers is insufficient.
\n\nI have the impression that competitions today require a mateirel as expensive as their price.
", "messages": [ { "id": 912223, "postDate": "2020-07-02T10:25:27.340Z", "content": "how about suggest to kaggle to implement some award point system that is convertible to more tpu/gpu kernel hours?\ne.g. if you setup public kernel or public dataset and people like your work, you will entitled to more tpu/gpu kernel hours.
\n\nAnother idea is \"transferable hours\". some kaggler has their own gpu resources and do not use the fixed hours per week allocated. may be kaggle can allowed such allocation to be freed for others who need them more?
", "rawMarkdown": "how about suggest to kaggle to implement some award point system that is convertible to more tpu/gpu kernel hours?\ne.g. if you setup public kernel or public dataset and people like your work, you will entitled to more tpu/gpu kernel hours.\n\nAnother idea is \"transferable hours\". some kaggler has their own gpu resources and do not use the fixed hours per week allocated. may be kaggle can allowed such allocation to be freed for others who need them more?", "votes": 17, "replies": [ { "id": 934195, "postDate": "2020-07-18T09:50:36.433Z", "content": "need to set up acceptance criteria such as high-quality content and most importantly, it is original and not a product of the so-called \"copy-paste method\".
", "rawMarkdown": "need to set up acceptance criteria such as high-quality content and most importantly, it is original and not a product of the so-called \"copy-paste method\"." } ] }, { "id": 904442, "postDate": "2020-06-27T16:18:13.357Z", "content": "It's unfair, I try to develop my idea and I find that I am restricted by my hardware, I have a modest gaming pc and I can not train my model or do transfer learning because of the low power of calculates or lack of RAM, even with the cloud that kaggle offers is insufficient.
\n\nI have the impression that competitions today require a mateirel as expensive as their price.
", "rawMarkdown": "It's unfair, I try to develop my idea and I find that I am restricted by my hardware, I have a modest gaming pc and I can not train my model or do transfer learning because of the low power of calculates or lack of RAM, even with the cloud that kaggle offers is insufficient.\n\nI have the impression that competitions today require a mateirel as expensive as their price.", "votes": 10 }, { "id": 904469, "postDate": "2020-06-27T16:41:43.597Z", "content": "but still you can learn a lot from kaggle. The main purpose is leraning smth new.
", "rawMarkdown": "but still you can learn a lot from kaggle. The main purpose is leraning smth new.", "votes": 3, "replies": [ { "id": 904547, "postDate": "2020-06-27T17:47:18Z", "content": "i agree with that, but still little bit frustrating to not have posibility to devlope our ideas.
", "rawMarkdown": "i agree with that, but still little bit frustrating to not have posibility to devlope our ideas." }, { "id": 904551, "postDate": "2020-06-27T17:49:59.613Z", "content": "You can use gpu and tpu quota as long with colab.
", "rawMarkdown": "You can use gpu and tpu quota as long with colab. " }, { "id": 935108, "postDate": "2020-07-19T04:56:01.723Z", "content": "totally agree with you.
\n\nIt doesn't matter what your rankings are, the most important thing is you have learnt something each time you play a competition.
\n\nFor me, the biggest lesson is Pytorch Efficient nets are super slow, and I should have trained other models earlier :(
", "rawMarkdown": "totally agree with you. \n\nIt doesn't matter what your rankings are, the most important thing is you have learnt something each time you play a competition.\n\nFor me, the biggest lesson is Pytorch Efficient nets are super slow, and I should have trained other models earlier :( ", "votes": 1 } ] }, { "id": 934619, "postDate": "2020-07-18T15:33:18.013Z", "content": "well sorry to sound a bit harsh... but life is unfair... you just have to deal with it... i saw people with less (GPU) won (gold) when people with more sunk to the bottom.
", "rawMarkdown": "well sorry to sound a bit harsh... but life is unfair... you just have to deal with it... i saw people with less (GPU) won (gold) when people with more sunk to the bottom.", "votes": 3 }, { "id": 905499, "postDate": "2020-06-28T15:43:54.460Z", "content": "", "rawMarkdown": "https://vast.ai", "votes": 3, "replies": [ { "id": 905502, "postDate": "2020-06-28T15:45:27.397Z", "content": "Tony san suggested this to me also 😁
", "rawMarkdown": "Tony san suggested this to me also 😁 ", "votes": 1 } ] }, { "id": 904468, "postDate": "2020-06-27T16:40:29.407Z", "content": "if you want to get high on the leaderboard, yes it requires a lot of computational resources. And it is not a surprise for me, i have bought 2080ti just to compete here:) like in all other competitions your equipment matters a lot:)
", "rawMarkdown": "if you want to get high on the leaderboard, yes it requires a lot of computational resources. And it is not a surprise for me, i have bought 2080ti just to compete here:) like in all other competitions your equipment matters a lot:) ", "votes": 1, "replies": [ { "id": 904660, "postDate": "2020-06-27T19:48:51.520Z", "content": "@vovanf98 \nAre you OK with this GPU? \nI've 2070, still very slow training.
", "rawMarkdown": "@vovanf98 \nAre you OK with this GPU? \nI've 2070, still very slow training. " }, { "id": 904709, "postDate": "2020-06-27T20:34:10.487Z", "content": "2080ti allows me to train with decent batch size in mixed-precision and yes, the training is smth like 2-3 hours / epoch on full dataset. 40 min for eb0
", "rawMarkdown": "2080ti allows me to train with decent batch size in mixed-precision and yes, the training is smth like 2-3 hours / epoch on full dataset. 40 min for eb0" } ] }, { "id": 904459, "postDate": "2020-06-27T16:30:49.737Z", "content": "I hope in the days to come we'll be able to more leverage the power of TPUs.
", "rawMarkdown": "I hope in the days to come we'll be able to more leverage the power of TPUs.", "votes": 1, "replies": [ { "id": 904544, "postDate": "2020-06-27T17:45:56.910Z", "content": "Hope too
", "rawMarkdown": "Hope too" } ] }, { "id": 912205, "postDate": "2020-07-02T09:58:01.300Z", "content": "Don't be too quick to give up. There are many cloud services that give you powerful GPUs that are very sufficient for a good standing on Kaggle. GCP and AWS have very generous free tiers for new users which my team and I have used to get our current standing without spending any money. :)
", "rawMarkdown": "Don't be too quick to give up. There are many cloud services that give you powerful GPUs that are very sufficient for a good standing on Kaggle. GCP and AWS have very generous free tiers for new users which my team and I have used to get our current standing without spending any money. :) ", "votes": 2 }, { "id": 933978, "postDate": "2020-07-18T07:09:09.047Z", "content": "Hmm in the deep learning arena I would be surprised to hear someone satisfied with his hardware resources...So we are all limited by something, it sounds to me that being limited by hardware and having plenty of ideas, is an ideal situation, much preferable than the converse.
", "rawMarkdown": "Hmm in the deep learning arena I would be surprised to hear someone satisfied with his hardware resources...So we are all limited by something, it sounds to me that being limited by hardware and having plenty of ideas, is an ideal situation, much preferable than the converse." }, { "id": 933880, "postDate": "2020-07-18T05:15:08.333Z", "content": "\"Money can not buy buy time\"\nWell the proverb fails in terms of having a powerful GPU 🙂🙂
", "rawMarkdown": "\"Money can not buy buy time\"\nWell the proverb fails in terms of having a powerful GPU 🙂🙂", "replies": [ { "id": 934188, "postDate": "2020-07-18T09:42:39.597Z", "content": "@zaber666 \nIt actually depends on the competition. Kaggle nowadays hosts the kinda research-level competition, I think. Small tips, if you don't have a strong GPU yet, try to manage resources from your university - of course not officially but personally requesting to prof/ass. prof/lecturer who has trust on you.
", "rawMarkdown": "@zaber666 \nIt actually depends on the competition. Kaggle nowadays hosts the kinda research-level competition, I think. Small tips, if you don't have a strong GPU yet, try to manage resources from your university - of course not officially but personally requesting to prof/ass. prof/lecturer who has trust on you. ", "votes": 1 }, { "id": 934299, "postDate": "2020-07-18T11:19:03.933Z", "content": "Yeah, I had it in my mind. But first it is necessary to get familiar among the teachers. Currently working on that.
", "rawMarkdown": "Yeah, I had it in my mind. But first it is necessary to get familiar among the teachers. Currently working on that.", "votes": 1 } ] }, { "id": 910892, "postDate": "2020-07-01T12:39:33.083Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 912223, "author_name": "hengck23", "author_url": "", "post_date": "2020-07-02T10:25:27.340000", "content": "how about suggest to kaggle to implement some award point system that is convertible to more tpu/gpu kernel hours?\ne.g. if you setup public kernel or public dataset and people like your work, you will entitled to more tpu/gpu kernel hours.
\n\nAnother idea is \"transferable hours\". some kaggler has their own gpu resources and do not use the fixed hours per week allocated. may be kaggle can allowed such allocation to be freed for others who need them more?
", "votes": 17, "replies": [ { "id": 934195, "author_name": "FGPC", "author_url": "", "post_date": "2020-07-18T09:50:36.433000", "content": "need to set up acceptance criteria such as high-quality content and most importantly, it is original and not a product of the so-called \"copy-paste method\".
", "votes": 0, "replies": [] } ] }, { "id": 904469, "author_name": "[ods.ai]Uladzimir Tumanau", "author_url": "", "post_date": "2020-06-27T16:41:43.597000", "content": "but still you can learn a lot from kaggle. The main purpose is leraning smth new.
", "votes": 3, "replies": [ { "id": 904547, "author_name": "Mehdi GASMI", "author_url": "", "post_date": "2020-06-27T17:47:18", "content": "i agree with that, but still little bit frustrating to not have posibility to devlope our ideas.
", "votes": 0, "replies": [] }, { "id": 904551, "author_name": "[ods.ai]Uladzimir Tumanau", "author_url": "", "post_date": "2020-06-27T17:49:59.613000", "content": "You can use gpu and tpu quota as long with colab.
", "votes": 0, "replies": [] }, { "id": 935108, "author_name": "Hao He", "author_url": "", "post_date": "2020-07-19T04:56:01.723000", "content": "totally agree with you.
\n\nIt doesn't matter what your rankings are, the most important thing is you have learnt something each time you play a competition.
\n\nFor me, the biggest lesson is Pytorch Efficient nets are super slow, and I should have trained other models earlier :(
", "votes": 1, "replies": [] } ] }, { "id": 934619, "author_name": "YL", "author_url": "", "post_date": "2020-07-18T15:33:18.013000", "content": "well sorry to sound a bit harsh... but life is unfair... you just have to deal with it... i saw people with less (GPU) won (gold) when people with more sunk to the bottom.
", "votes": 3, "replies": [] }, { "id": 905499, "author_name": "Qishen Ha", "author_url": "", "post_date": "2020-06-28T15:43:54.460000", "content": "", "votes": 3, "replies": [ { "id": 905502, "author_name": "Innat", "author_url": "", "post_date": "2020-06-28T15:45:27.397000", "content": "Tony san suggested this to me also 😁
", "votes": 1, "replies": [] } ] }, { "id": 904468, "author_name": "[ods.ai]Uladzimir Tumanau", "author_url": "", "post_date": "2020-06-27T16:40:29.407000", "content": "if you want to get high on the leaderboard, yes it requires a lot of computational resources. And it is not a surprise for me, i have bought 2080ti just to compete here:) like in all other competitions your equipment matters a lot:)
", "votes": 1, "replies": [ { "id": 904660, "author_name": "Innat", "author_url": "", "post_date": "2020-06-27T19:48:51.520000", "content": "@vovanf98 \nAre you OK with this GPU? \nI've 2070, still very slow training.
", "votes": 0, "replies": [] }, { "id": 904709, "author_name": "[ods.ai]Uladzimir Tumanau", "author_url": "", "post_date": "2020-06-27T20:34:10.487000", "content": "2080ti allows me to train with decent batch size in mixed-precision and yes, the training is smth like 2-3 hours / epoch on full dataset. 40 min for eb0
", "votes": 0, "replies": [] } ] }, { "id": 904459, "author_name": "Mighty Rains", "author_url": "", "post_date": "2020-06-27T16:30:49.737000", "content": "I hope in the days to come we'll be able to more leverage the power of TPUs.
", "votes": 1, "replies": [ { "id": 904544, "author_name": "Mehdi GASMI", "author_url": "", "post_date": "2020-06-27T17:45:56.910000", "content": "Hope too
", "votes": 0, "replies": [] } ] }, { "id": 912205, "author_name": "JunShern Chan", "author_url": "", "post_date": "2020-07-02T09:58:01.300000", "content": "Don't be too quick to give up. There are many cloud services that give you powerful GPUs that are very sufficient for a good standing on Kaggle. GCP and AWS have very generous free tiers for new users which my team and I have used to get our current standing without spending any money. :)
", "votes": 2, "replies": [] }, { "id": 933978, "author_name": "Seba", "author_url": "", "post_date": "2020-07-18T07:09:09.047000", "content": "Hmm in the deep learning arena I would be surprised to hear someone satisfied with his hardware resources...So we are all limited by something, it sounds to me that being limited by hardware and having plenty of ideas, is an ideal situation, much preferable than the converse.
", "votes": 0, "replies": [] }, { "id": 933880, "author_name": "Zaber Ibn Abdul Hakim", "author_url": "", "post_date": "2020-07-18T05:15:08.333000", "content": "\"Money can not buy buy time\"\nWell the proverb fails in terms of having a powerful GPU 🙂🙂
", "votes": 0, "replies": [ { "id": 934188, "author_name": "Innat", "author_url": "", "post_date": "2020-07-18T09:42:39.597000", "content": "@zaber666 \nIt actually depends on the competition. Kaggle nowadays hosts the kinda research-level competition, I think. Small tips, if you don't have a strong GPU yet, try to manage resources from your university - of course not officially but personally requesting to prof/ass. prof/lecturer who has trust on you.
", "votes": 1, "replies": [] }, { "id": 934299, "author_name": "Zaber Ibn Abdul Hakim", "author_url": "", "post_date": "2020-07-18T11:19:03.933000", "content": "Yeah, I had it in my mind. But first it is necessary to get familiar among the teachers. Currently working on that.
", "votes": 1, "replies": [] } ] }, { "id": 910892, "author_name": "", "author_url": "", "post_date": "2020-07-01T12:39:33.083000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "912223": "how about suggest to kaggle to implement some award point system that is convertible to more tpu/gpu kernel hours?\ne.g. if you setup public kernel or public dataset and people like your work, you will entitled to more tpu/gpu kernel hours.\n\nAnother idea is \"transferable hours\". some kaggler has their own gpu resources and do not use the fixed hours per week allocated. may be kaggle can allowed such allocation to be freed for others who need them more?", "904442": "It's unfair, I try to develop my idea and I find that I am restricted by my hardware, I have a modest gaming pc and I can not train my model or do transfer learning because of the low power of calculates or lack of RAM, even with the cloud that kaggle offers is insufficient.\n\nI have the impression that competitions today require a mateirel as expensive as their price.", "904469": "but still you can learn a lot from kaggle. The main purpose is leraning smth new.", "934619": "well sorry to sound a bit harsh... but life is unfair... you just have to deal with it... i saw people with less (GPU) won (gold) when people with more sunk to the bottom.", "905499": "https://vast.ai", "904468": "if you want to get high on the leaderboard, yes it requires a lot of computational resources. And it is not a surprise for me, i have bought 2080ti just to compete here:) like in all other competitions your equipment matters a lot:) ", "904459": "I hope in the days to come we'll be able to more leverage the power of TPUs.", "912205": "Don't be too quick to give up. There are many cloud services that give you powerful GPUs that are very sufficient for a good standing on Kaggle. GCP and AWS have very generous free tiers for new users which my team and I have used to get our current standing without spending any money. :) ", "933978": "Hmm in the deep learning arena I would be surprised to hear someone satisfied with his hardware resources...So we are all limited by something, it sounds to me that being limited by hardware and having plenty of ideas, is an ideal situation, much preferable than the converse.", "933880": "\"Money can not buy buy time\"\nWell the proverb fails in terms of having a powerful GPU 🙂🙂", "910892": "" } }