Lots of people train model in one notebook for many hours and get a private weights. Then they use the weights for the inference notebook. Only the inference notebook is counted by the time limit? How about the pre-tuned private weights? Is it limited by time? Is the pre-tuned private weights counted as external data? If we can work like this way, I don't know why this is a code competition.
", "messages": [ { "id": "918058", "postDate": "07/06/2020 23:37:52", "content": "Lots of people train model in one notebook for many hours and get a private weights. Then they use the weights for the inference notebook. Only the inference notebook is counted by the time limit? How about the pre-tuned private weights? Is it limited by time? Is the pre-tuned private weights counted as external data? If we can work like this way, I don't know why this is a code competition.
", "rawMarkdown": "Lots of people train model in one notebook for many hours and get a private weights. Then they use the weights for the inference notebook. Only the inference notebook is counted by the time limit? How about the pre-tuned private weights? Is it limited by time? Is the pre-tuned private weights counted as external data? If we can work like this way, I don't know why this is a code competition.", "votes": null }, { "id": "918066", "postDate": "07/07/2020 00:04:44", "content": "Yes, it is allowed. There is no time limit placed on how long it takes to train your model, only the inference portion.
\n\nSecondly, you do not have to share your trained weights if you trained them entirely by yourself. However, if you are using any pre-trained weights that you downloaded from somewhere, then yes these must be disclosed in the external data thread before the entry deadline.
\n\nI think the reasoning behind this being a \"code competition\" is that the test set is entirely hidden from us. Thus, instead of us submitting predictions, we are actually submitting our inference code.
", "rawMarkdown": "Yes, it is allowed. There is no time limit placed on how long it takes to train your model, only the inference portion.\n\nSecondly, you do not have to share your trained weights if you trained them entirely by yourself. However, if you are using any pre-trained weights that you downloaded from somewhere, then yes these must be disclosed in the external data thread before the entry deadline.\n\nI think the reasoning behind this being a \"code competition\" is that the test set is entirely hidden from us. Thus, instead of us submitting predictions, we are actually submitting our inference code.", "votes": null }, { "id": "918079", "postDate": "07/07/2020 01:01:46", "content": "Thanks for the answering. It's more clear to me. If the \"code competition\" is for hidden test, we can just load private weights and predict for test dataset for competition. We can even convert the training set and other preprocessing locally, which saves more time for the inference. Am I right?
", "rawMarkdown": "Thanks for the answering. It's more clear to me. If the \"code competition\" is for hidden test, we can just load private weights and predict for test dataset for competition. We can even convert the training set and other preprocessing locally, which saves more time for the inference. Am I right?", "votes": null }, { "id": "918081", "postDate": "07/07/2020 01:04:52", "content": "Yes, it is certainly advantageous to do any processing or training you want \"offline\" and simply perform the inference in your submission.
", "rawMarkdown": "Yes, it is certainly advantageous to do any processing or training you want \"offline\" and simply perform the inference in your submission.", "votes": null }, { "id": "918116", "postDate": "07/07/2020 02:10:13", "content": "To add to the excellent replies by @matthewmasters... many people use Google Colab to do their long training runs; e.g. if you are running 60-100 epochs, that can go 12 to 20 hours. Using Google Colab, you can conserve your precious 30 GPU (Kaggle) hours per week for the actual inference runs, which again may easily consume at least 2.5 hours each if you are doing an extra 10 epochs.
", "rawMarkdown": "To add to the excellent replies by @matthewmasters... many people use Google Colab to do their long training runs; e.g. if you are running 60-100 epochs, that can go 12 to 20 hours. Using Google Colab, you can conserve your precious 30 GPU (Kaggle) hours per week for the actual inference runs, which again may easily consume *at least* 2.5 hours each if you are doing an extra 10 epochs.", "votes": null }, { "id": "918264", "postDate": "07/07/2020 06:10:51", "content": "Thanks. Now I understand why Colab is a so good choice.
", "rawMarkdown": "Thanks. Now I understand why Colab is a so good choice.", "votes": null } ], "comments": [ { "id": 918066, "author_name": "matthewmasters", "author_url": "", "post_date": "07/07/2020 00:04:44", "content": "Yes, it is allowed. There is no time limit placed on how long it takes to train your model, only the inference portion.
\n\nSecondly, you do not have to share your trained weights if you trained them entirely by yourself. However, if you are using any pre-trained weights that you downloaded from somewhere, then yes these must be disclosed in the external data thread before the entry deadline.
\n\nI think the reasoning behind this being a \"code competition\" is that the test set is entirely hidden from us. Thus, instead of us submitting predictions, we are actually submitting our inference code.
", "votes": null, "replies": [ { "id": 918079, "author_name": "tian43", "author_url": "", "post_date": "07/07/2020 01:01:46", "content": "Thanks for the answering. It's more clear to me. If the \"code competition\" is for hidden test, we can just load private weights and predict for test dataset for competition. We can even convert the training set and other preprocessing locally, which saves more time for the inference. Am I right?
", "votes": null, "replies": [] }, { "id": 918081, "author_name": "matthewmasters", "author_url": "", "post_date": "07/07/2020 01:04:52", "content": "Yes, it is certainly advantageous to do any processing or training you want \"offline\" and simply perform the inference in your submission.
", "votes": null, "replies": [] }, { "id": 918116, "author_name": "mperutti", "author_url": "", "post_date": "07/07/2020 02:10:13", "content": "To add to the excellent replies by @matthewmasters... many people use Google Colab to do their long training runs; e.g. if you are running 60-100 epochs, that can go 12 to 20 hours. Using Google Colab, you can conserve your precious 30 GPU (Kaggle) hours per week for the actual inference runs, which again may easily consume at least 2.5 hours each if you are doing an extra 10 epochs.
", "votes": null, "replies": [] }, { "id": 918264, "author_name": "tian43", "author_url": "", "post_date": "07/07/2020 06:10:51", "content": "Thanks. Now I understand why Colab is a so good choice.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "918058": "Lots of people train model in one notebook for many hours and get a private weights. Then they use the weights for the inference notebook. Only the inference notebook is counted by the time limit? How about the pre-tuned private weights? Is it limited by time? Is the pre-tuned private weights counted as external data? If we can work like this way, I don't know why this is a code competition.", "918066": "Yes, it is allowed. There is no time limit placed on how long it takes to train your model, only the inference portion.\n\nSecondly, you do not have to share your trained weights if you trained them entirely by yourself. However, if you are using any pre-trained weights that you downloaded from somewhere, then yes these must be disclosed in the external data thread before the entry deadline.\n\nI think the reasoning behind this being a \"code competition\" is that the test set is entirely hidden from us. Thus, instead of us submitting predictions, we are actually submitting our inference code.", "918079": "Thanks for the answering. It's more clear to me. If the \"code competition\" is for hidden test, we can just load private weights and predict for test dataset for competition. We can even convert the training set and other preprocessing locally, which saves more time for the inference. Am I right?", "918081": "Yes, it is certainly advantageous to do any processing or training you want \"offline\" and simply perform the inference in your submission.", "918116": "To add to the excellent replies by @matthewmasters... many people use Google Colab to do their long training runs; e.g. if you are running 60-100 epochs, that can go 12 to 20 hours. Using Google Colab, you can conserve your precious 30 GPU (Kaggle) hours per week for the actual inference runs, which again may easily consume *at least* 2.5 hours each if you are doing an extra 10 epochs.", "918264": "Thanks. Now I understand why Colab is a so good choice." }, "source": "meta" }