Hi all,\nI never compete in this type of competition, my question is why people split their work into a train notebook and an inference one? (this helps somehow to better use the restrictions on the GPU & CPU?, it's only matter of organize the code?)
", "messages": [ { "id": "851986", "postDate": "05/18/2020 05:06:01", "content": "Hi all,\nI never compete in this type of competition, my question is why people split their work into a train notebook and an inference one? (this helps somehow to better use the restrictions on the GPU & CPU?, it's only matter of organize the code?)
", "rawMarkdown": "Hi all,\nI never compete in this type of competition, my question is why people split their work into a train notebook and an inference one? (this helps somehow to better use the restrictions on the GPU & CPU?, it's only matter of organize the code?)", "votes": null }, { "id": "852212", "postDate": "05/18/2020 08:32:51", "content": "Yes, because you have the restrictions regarding memory, gpu usage etc , usually you preprocess your data and save them as an inference kernel and then you create another notebook by loading and training the data.
", "rawMarkdown": "Yes, because you have the restrictions regarding memory, gpu usage etc , usually you preprocess your data and save them as an inference kernel and then you create another notebook by loading and training the data.", "votes": null }, { "id": "852252", "postDate": "05/18/2020 09:22:03", "content": "So I can just train the model on one script (and access the internet, more CPU & GPU time, etc) - for let's say 30 hours\nand then just do the prediction on another script and it will be ok? (because on the inference time will be short)
", "rawMarkdown": "So I can just train the model on one script (and access the internet, more CPU & GPU time, etc) - for let's say 30 hours\nand then just do the prediction on another script and it will be ok? (because on the inference time will be short)", "votes": null }, { "id": "852285", "postDate": "05/18/2020 10:05:19", "content": "Yes, as long the prediction fulfills the kernel limitations.
", "rawMarkdown": "Yes, as long the prediction fulfills the kernel limitations.", "votes": null }, { "id": "852834", "postDate": "05/18/2020 17:56:55", "content": "I am digesting this fact. Kernel only submission often have restrictions that the kernel that produces the kernel is limited to say 180 minute execution time. Training for 30 hours and making prediction/inference in short time defeats the purpose. Probably, I am missing some details.
", "rawMarkdown": "I am digesting this fact. Kernel only submission often have restrictions that the kernel that produces the kernel is limited to say 180 minute execution time. Training for 30 hours and making prediction/inference in short time defeats the purpose. Probably, I am missing some details.", "votes": null }, { "id": "852943", "postDate": "05/18/2020 20:13:54", "content": "Can I do super train on my local server, and then only upload the network (that I already trained) to a thin inference script? from what you said it's legal no?
", "rawMarkdown": "Can I do super train on my local server, and then only upload the network (that I already trained) to a thin inference script? from what you said it's legal no?", "votes": null }, { "id": "853461", "postDate": "05/19/2020 07:51:09", "content": "Check
\n\nhere\nand
\n\n", "rawMarkdown": "Check \n\n[here](https://www.kaggle.com/c/imet-2019-fgvc6/discussion/87544)\nand \n\n[here](https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification/discussion/87719)", "votes": null } ], "comments": [ { "id": 852212, "author_name": "ggeo79", "author_url": "", "post_date": "05/18/2020 08:32:51", "content": "Yes, because you have the restrictions regarding memory, gpu usage etc , usually you preprocess your data and save them as an inference kernel and then you create another notebook by loading and training the data.
", "votes": null, "replies": [] }, { "id": 852252, "author_name": "tomeramit2", "author_url": "", "post_date": "05/18/2020 09:22:03", "content": "So I can just train the model on one script (and access the internet, more CPU & GPU time, etc) - for let's say 30 hours\nand then just do the prediction on another script and it will be ok? (because on the inference time will be short)
", "votes": null, "replies": [ { "id": 852285, "author_name": "ggeo79", "author_url": "", "post_date": "05/18/2020 10:05:19", "content": "Yes, as long the prediction fulfills the kernel limitations.
", "votes": null, "replies": [] } ] }, { "id": 852834, "author_name": "bhorkar", "author_url": "", "post_date": "05/18/2020 17:56:55", "content": "I am digesting this fact. Kernel only submission often have restrictions that the kernel that produces the kernel is limited to say 180 minute execution time. Training for 30 hours and making prediction/inference in short time defeats the purpose. Probably, I am missing some details.
", "votes": null, "replies": [] }, { "id": 852943, "author_name": "tomeramit2", "author_url": "", "post_date": "05/18/2020 20:13:54", "content": "Can I do super train on my local server, and then only upload the network (that I already trained) to a thin inference script? from what you said it's legal no?
", "votes": null, "replies": [ { "id": 853461, "author_name": "ggeo79", "author_url": "", "post_date": "05/19/2020 07:51:09", "content": "Check
\n\nhere\nand
\n\n", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "851986": "Hi all,\nI never compete in this type of competition, my question is why people split their work into a train notebook and an inference one? (this helps somehow to better use the restrictions on the GPU & CPU?, it's only matter of organize the code?)", "852212": "Yes, because you have the restrictions regarding memory, gpu usage etc , usually you preprocess your data and save them as an inference kernel and then you create another notebook by loading and training the data.", "852252": "So I can just train the model on one script (and access the internet, more CPU & GPU time, etc) - for let's say 30 hours\nand then just do the prediction on another script and it will be ok? (because on the inference time will be short)", "852285": "Yes, as long the prediction fulfills the kernel limitations.", "852834": "I am digesting this fact. Kernel only submission often have restrictions that the kernel that produces the kernel is limited to say 180 minute execution time. Training for 30 hours and making prediction/inference in short time defeats the purpose. Probably, I am missing some details.", "852943": "Can I do super train on my local server, and then only upload the network (that I already trained) to a thin inference script? from what you said it's legal no?", "853461": "Check \n\n[here](https://www.kaggle.com/c/imet-2019-fgvc6/discussion/87544)\nand \n\n[here](https://www.kaggle.com/c/jigsaw-unintended-bias-in-toxicity-classification/discussion/87719)" }, "source": "meta" }