this competition has introduced us to a vast dataset of 200 GB. Even when that is not the case, most code competitions have an inference limit , which here is 9 hours. Since we only get 30 hours of GPU per week, any notebook utilizing its maximum inference time can have at max 3 of its versions submitted successfully. However, that does not leave enough GPU for model training. i have been facing this problem for the past few weeks as I end up exhausting my GPU within first 2 days and then I am neither able to train any new model, nor make any submissions.
\nso as a beginner, I wanted to ask you guys, where and how do you train your ml models? where are you training your ml model for this contest and how ?
", "messages": [ { "id": 2353271, "postDate": "2023-07-21T14:52:41.053Z", "content": "I also have this question. It seems to me the original spirit/intention of the Code competition was to level the playing field by requiring all participants to train on Kaggle platform. Is that right?
\nHowever, I believe that during this competition we are allowed to train on Google Colab or on a personal GPU or other external resources. Saving the weights and/or model and then uploading them for a submission from Kaggle notebook. Technically, it is possible to submit this way.
\nFor example, one way to do it is to separate the code into training and inference notebooks, as done in this solution:
\nhttps://www.kaggle.com/competitions/asl-signs/discussion/406978
Perhaps for the final submission at the deadline, it is required to fully train it on Kaggle platform or at least provide a training notebook that can train the model fully on Kaggle platform, but I'm not sure.
\nWith regard to the large dataset, you can use the select columns feature of parquet or pre-process the dataset to reduce its size considerably. People shared pre-processing example kernels in the Code section of this competition.
\nWith regard to inference time, I'm not sure but I think there's a distinction between the notebook running time and the the submission process time. I believe the time it takes to process the submission in the backend after we submit the submission.zip file (max 5 hours), does not count towards the notebook quota. The front-end notebook running time at submission can be only a few minutes when you submit inference-only code with pre-trained weights as in the above-mentioned solution.
", "rawMarkdown": "I also have this question. It seems to me the original spirit/intention of the Code competition was to level the playing field by requiring all participants to train on Kaggle platform. Is that right?\n\nHowever, I believe that during this competition we are allowed to train on Google Colab or on a personal GPU or other external resources. Saving the weights and/or model and then uploading them for a submission from Kaggle notebook. Technically, it is possible to submit this way. \n\nFor example, one way to do it is to separate the code into training and inference notebooks, as done in this solution:\nhttps://www.kaggle.com/competitions/asl-signs/discussion/406978\n\nPerhaps for the final submission at the deadline, it is required to fully train it on Kaggle platform or at least provide a training notebook that can train the model fully on Kaggle platform, but I'm not sure.\n\nWith regard to the large dataset, you can use the select columns feature of parquet or pre-process the dataset to reduce its size considerably. People shared pre-processing example kernels in the Code section of this competition.\n\nWith regard to inference time, I'm not sure but I think there's a distinction between the notebook running time and the the submission process time. I believe the time it takes to process the submission in the backend after we submit the submission.zip file (max 5 hours), does not count towards the notebook quota. The front-end notebook running time at submission can be only a few minutes when you submit inference-only code with pre-trained weights as in the above-mentioned solution. \n", "votes": 3, "replies": [ { "id": 2353989, "postDate": "2023-07-22T07:06:08.727Z", "content": "Well, this is why we are given TPU. Sure, one can spend thousands of dollars on external compute, but as long as you utilize the TPU, it's about ~1000$ per month of compute (at 12.88$ / hour, granted this price is for Google's v4-8, i.e., next generation, but still). There is no requirement that I know of to train only on Kaggle, nor can such a requirement be enforcedβ¦
", "rawMarkdown": "Well, this is why we are given TPU. Sure, one can spend thousands of dollars on external compute, but as long as you utilize the TPU, it's about ~1000$ per month of compute (at 12.88$ / hour, granted this price is for Google's v4-8, i.e., next generation, but still). There is no requirement that I know of to train only on Kaggle, nor can such a requirement be enforced...", "votes": 4, "replies": [ { "id": 2355039, "postDate": "2023-07-23T04:08:19.383Z", "content": "You can read some requirements in this link: \"https://www.kaggle.com/competitions/asl-fingerspelling/overview/code-requirements\". And it said that \"Freely & publicly available external data is allowed, including pre-trained models.\". Let's enjoy the competition together. πππ
", "rawMarkdown": "You can read some requirements in this link: \"https://www.kaggle.com/competitions/asl-fingerspelling/overview/code-requirements\". And it said that \"Freely & publicly available external data is allowed, including pre-trained models.\". Let's enjoy the competition together. πππ", "votes": 3 } ] } ] }, { "id": 2347260, "postDate": "2023-07-16T21:10:20.403Z", "content": "this competition has introduced us to a vast dataset of 200 GB. Even when that is not the case, most code competitions have an inference limit , which here is 9 hours. Since we only get 30 hours of GPU per week, any notebook utilizing its maximum inference time can have at max 3 of its versions submitted successfully. However, that does not leave enough GPU for model training. i have been facing this problem for the past few weeks as I end up exhausting my GPU within first 2 days and then I am neither able to train any new model, nor make any submissions.
\nso as a beginner, I wanted to ask you guys, where and how do you train your ml models? where are you training your ml model for this contest and how ?
", "rawMarkdown": "this competition has introduced us to a vast dataset of 200 GB. Even when that is not the case, most code competitions have an inference limit , which here is 9 hours. Since we only get 30 hours of GPU per week, any notebook utilizing its maximum inference time can have at max 3 of its versions submitted successfully. However, that does not leave enough GPU for model training. i have been facing this problem for the past few weeks as I end up exhausting my GPU within first 2 days and then I am neither able to train any new model, nor make any submissions.\n\nso as a beginner, I wanted to ask you guys, where and how do you train your ml models? where are you training your ml model for this contest and how ?", "votes": 4 } ], "comments": [ { "id": 2353271, "author_name": "WalkingMoose", "author_url": "", "post_date": "2023-07-21T14:52:41.053000", "content": "I also have this question. It seems to me the original spirit/intention of the Code competition was to level the playing field by requiring all participants to train on Kaggle platform. Is that right?
\nHowever, I believe that during this competition we are allowed to train on Google Colab or on a personal GPU or other external resources. Saving the weights and/or model and then uploading them for a submission from Kaggle notebook. Technically, it is possible to submit this way.
\nFor example, one way to do it is to separate the code into training and inference notebooks, as done in this solution:
\nhttps://www.kaggle.com/competitions/asl-signs/discussion/406978
Perhaps for the final submission at the deadline, it is required to fully train it on Kaggle platform or at least provide a training notebook that can train the model fully on Kaggle platform, but I'm not sure.
\nWith regard to the large dataset, you can use the select columns feature of parquet or pre-process the dataset to reduce its size considerably. People shared pre-processing example kernels in the Code section of this competition.
\nWith regard to inference time, I'm not sure but I think there's a distinction between the notebook running time and the the submission process time. I believe the time it takes to process the submission in the backend after we submit the submission.zip file (max 5 hours), does not count towards the notebook quota. The front-end notebook running time at submission can be only a few minutes when you submit inference-only code with pre-trained weights as in the above-mentioned solution.
", "votes": 3, "replies": [ { "id": 2353989, "author_name": "greySnow", "author_url": "", "post_date": "2023-07-22T07:06:08.727000", "content": "Well, this is why we are given TPU. Sure, one can spend thousands of dollars on external compute, but as long as you utilize the TPU, it's about ~1000$ per month of compute (at 12.88$ / hour, granted this price is for Google's v4-8, i.e., next generation, but still). There is no requirement that I know of to train only on Kaggle, nor can such a requirement be enforcedβ¦
", "votes": 4, "replies": [ { "id": 2355039, "author_name": "π<->π", "author_url": "", "post_date": "2023-07-23T04:08:19.383000", "content": "You can read some requirements in this link: \"https://www.kaggle.com/competitions/asl-fingerspelling/overview/code-requirements\". And it said that \"Freely & publicly available external data is allowed, including pre-trained models.\". Let's enjoy the competition together. πππ
", "votes": 3, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2353271": "I also have this question. It seems to me the original spirit/intention of the Code competition was to level the playing field by requiring all participants to train on Kaggle platform. Is that right?\n\nHowever, I believe that during this competition we are allowed to train on Google Colab or on a personal GPU or other external resources. Saving the weights and/or model and then uploading them for a submission from Kaggle notebook. Technically, it is possible to submit this way. \n\nFor example, one way to do it is to separate the code into training and inference notebooks, as done in this solution:\nhttps://www.kaggle.com/competitions/asl-signs/discussion/406978\n\nPerhaps for the final submission at the deadline, it is required to fully train it on Kaggle platform or at least provide a training notebook that can train the model fully on Kaggle platform, but I'm not sure.\n\nWith regard to the large dataset, you can use the select columns feature of parquet or pre-process the dataset to reduce its size considerably. People shared pre-processing example kernels in the Code section of this competition.\n\nWith regard to inference time, I'm not sure but I think there's a distinction between the notebook running time and the the submission process time. I believe the time it takes to process the submission in the backend after we submit the submission.zip file (max 5 hours), does not count towards the notebook quota. The front-end notebook running time at submission can be only a few minutes when you submit inference-only code with pre-trained weights as in the above-mentioned solution. \n", "2347260": "this competition has introduced us to a vast dataset of 200 GB. Even when that is not the case, most code competitions have an inference limit , which here is 9 hours. Since we only get 30 hours of GPU per week, any notebook utilizing its maximum inference time can have at max 3 of its versions submitted successfully. However, that does not leave enough GPU for model training. i have been facing this problem for the past few weeks as I end up exhausting my GPU within first 2 days and then I am neither able to train any new model, nor make any submissions.\n\nso as a beginner, I wanted to ask you guys, where and how do you train your ml models? where are you training your ml model for this contest and how ?" } }