Why does this competition require submissions on a per-sequence basis?
\nConsidering the nature of the competition, it doesn't seem like sequential submissions should necessarily be required for this one. I wonder if there's a specific intention behind this.
\nOne thing that catches my attention is the following statement:
\nNote that there known sensor communication failure this dataset, resulting missing data sensors .\nIn the previous CMI competition, I recall that the top-ranking teams seemed to be very good at imputing NaN data. I wonder if the idea here is to interpolate this missing data continuously/sequentially. This is just an idea, though.
", "messages": [ { "id": 3213882, "postDate": "2025-05-30T16:00:11.493Z", "content": "Why does this competition require submissions on a per-sequence basis?
\nConsidering the nature of the competition, it doesn't seem like sequential submissions should necessarily be required for this one. I wonder if there's a specific intention behind this.
\nOne thing that catches my attention is the following statement:
\nNote that there known sensor communication failure this dataset, resulting missing data sensors .\nIn the previous CMI competition, I recall that the top-ranking teams seemed to be very good at imputing NaN data. I wonder if the idea here is to interpolate this missing data continuously/sequentially. This is just an idea, though.
", "rawMarkdown": "Why does this competition require submissions on a per-sequence basis?\n\nConsidering the nature of the competition, it doesn't seem like sequential submissions should necessarily be required for this one. I wonder if there's a specific intention behind this.\n\nOne thing that catches my attention is the following statement:\n```\nNote also that there is known sensor communication failure in this dataset, resulting in missing data from some sensors in some sequences.\n```\n\nIn the previous CMI competition, I recall that the top-ranking teams seemed to be very good at imputing NaN data. I wonder if the idea here is to interpolate this missing data continuously/sequentially. This is just an idea, though.", "votes": 12 }, { "id": 3213941, "postDate": "2025-05-30T17:27:00.930Z", "content": "Seems like there is only a few patients over time. They might have different behaviours. This is to prevent time leakage.
", "rawMarkdown": "Seems like there is only a few patients over time. They might have different behaviours. This is to prevent time leakage. ", "votes": 3, "replies": [ { "id": 3213976, "postDate": "2025-05-30T17:49:12.470Z", "content": "@lucasmorin Thank you for comment! It has the potential time behaviours. FYI : cv and public lb is almost matched. See here.
\nhttps://www.kaggle.com/competitions/cmi-detect-behavior-with-sensor-data/discussion/582344
It takes some time before people find and exploit leakage. The time series API prevents that. For an exemple of the alternative see the DRW competition. Leaker start to appear on the LB and the absence of leakage will be enforced by manual code review… this is not better.
", "rawMarkdown": "It takes some time before people find and exploit leakage. The time series API prevents that. For an exemple of the alternative see the [DRW competition](https://www.kaggle.com/competitions/drw-crypto-market-prediction/leaderboard). Leaker start to appear on the LB and the absence of leakage will be enforced by manual code review... this is not better. ", "votes": 1 }, { "id": 3214001, "postDate": "2025-05-30T18:24:38.613Z", "content": "@lucasmorin That's a perspective I hadn't considered. It's possible that's the case.
", "rawMarkdown": "@lucasmorin That's a perspective I hadn't considered. It's possible that's the case." } ] }, { "id": 3215540, "postDate": "2025-06-02T11:16:10.697Z", "content": "This competition's host upvoted you, so you might be onto something
", "rawMarkdown": "This competition's host upvoted you, so you might be onto something", "votes": 1 } ] }, { "id": 3215584, "postDate": "2025-06-02T12:52:02.707Z", "content": "One sequence corresponds to a subject and hence the tasks corresponds to predicting BFRB or Non BFRB for a subject there are approximately 3,500 sequences in test set each corresponding to a subject
", "rawMarkdown": "One sequence corresponds to a subject and hence the tasks corresponds to predicting BFRB or Non BFRB for a subject there are approximately 3,500 sequences in test set each corresponding to a subject", "votes": 1 }, { "id": 3214391, "postDate": "2025-05-31T12:43:32.193Z", "content": "I am facing a lot of problems in submitting my notebook, it always fails due to some error, can you help?
", "rawMarkdown": "I am facing a lot of problems in submitting my notebook, it always fails due to some error, can you help?", "replies": [ { "id": 3214429, "postDate": "2025-05-31T14:20:00.303Z", "content": "@sahib03 It might be good to make the notebook public and create some discussion topics. I'll comment if I find the time.
", "rawMarkdown": "@sahib03 It might be good to make the notebook public and create some discussion topics. I'll comment if I find the time.", "replies": [ { "id": 3215234, "postDate": "2025-06-01T20:43:53.453Z", "content": "How do I send my notebook to you? Should I attach it with a comment?
", "rawMarkdown": "How do I send my notebook to you? Should I attach it with a comment?" }, { "id": 3215430, "postDate": "2025-06-02T07:45:09.320Z", "content": "@sahib03 You must never send it by email. First, try looking up how to use Kaggle notebook and make it public using ChatGPT or other resources.
", "rawMarkdown": "@sahib03 You must never send it by email. First, try looking up how to use Kaggle notebook and make it public using ChatGPT or other resources." } ] }, { "id": 3215425, "postDate": "2025-06-02T07:37:22.277Z", "content": "had the same issue , make sure to:
\nkeep the early initialization
place your training code between the initialization and final launch
Maintain the environment checks (KAGGLE_IS_COMPETITION_RERUN) you can follow the snippet below
\n ():\n \n server = kaggle_evaluation.cmi_inference_server.CMIInferenceServer(predict)\n\n \n os.getenv():\n test_df = pd.read_csv()\n sample_seq = test_df[test_df[] == test_df[].iloc[]]\n _ = predict(pl.DataFrame(sample_seq), )\n\n server\n\n()\ninference_server = init_server()\n\n\n\n\n os.getenv():\n ()\n inference_server.serve()\n:\n ()\n :\n inference_server.run_local_gateway(\n data_paths=(\n ,\n ,\n )\n )\n Exception e:\n ()\n \nEdit: You can do the training in separate notebook (training interface) and create anothor notbook for submission in order to respect the time of the calling the interface
", "rawMarkdown": "had the same issue , make sure to:\n\n- keep the early initialization\n\n- place your training code between the initialization and final launch\n\n- Maintain the environment checks (KAGGLE_IS_COMPETITION_RERUN) you can follow the snippet below \n```python\n# ================ SERVER INITIALIZATION ================\ndef init_server():\n \"\"\"Initialize the server before heavy computations\"\"\"\n server = kaggle_evaluation.cmi_inference_server.CMIInferenceServer(predict)\n \n # Warm-up prediction to load model\n if not os.getenv('KAGGLE_IS_COMPETITION_RERUN'):\n test_df = pd.read_csv('/kaggle/input/cmi-detect-behavior-with-sensor-data/test.csv')\n sample_seq = test_df[test_df['sequence_id'] == test_df['sequence_id'].iloc[0]]\n _ = predict(pl.DataFrame(sample_seq), None)\n \n return server\n```\n\n```python\n# Initialize server EARLY (before model training)\nprint(\"Initializing server...\")\ninference_server = init_server()\n\n# [Your training and evaluation code goes here]\n\n# ================ FINAL SERVER LAUNCH ================\nif os.getenv('KAGGLE_IS_COMPETITION_RERUN'):\n print(\"Starting production server...\")\n inference_server.serve()\nelse:\n print(\"Running local test...\")\n try:\n inference_server.run_local_gateway(\n data_paths=(\n '/kaggle/input/cmi-detect-behavior-with-sensor-data/test.csv',\n '/kaggle/input/cmi-detect-behavior-with-sensor-data/test_demographics.csv',\n )\n )\n except Exception as e:\n print(f\"Local test failed: {e}\")\n raise\n```\n\nEdit: You can do the training in separate notebook (training interface) and create anothor notbook for submission in order to respect the time of the calling the interface", "votes": 1, "replies": [ { "id": 3215432, "postDate": "2025-06-02T07:47:00.043Z", "content": "@nadiatriki Thank you for your help!
", "rawMarkdown": "@nadiatriki Thank you for your help!" }, { "id": 3217540, "postDate": "2025-06-05T05:53:11.793Z", "content": "@nadiatriki I am still facing timeout error in submitting (inference only). How much time does it generally take for you?
", "rawMarkdown": "@nadiatriki I am still facing timeout error in submitting (inference only). How much time does it generally take for you? ", "votes": -1 }, { "id": 3217661, "postDate": "2025-06-05T09:13:40.370Z", "content": "for submitting it take me around 1-2 hours, some version took less than that obviously. can you screenshot so we can offer more help
", "rawMarkdown": "for submitting it take me around 1-2 hours, some version took less than that obviously. can you screenshot so we can offer more help" }, { "id": 3218358, "postDate": "2025-06-06T05:35:15.450Z", "content": "@nadiatriki Please find the inference code snippet as attached below.
", "rawMarkdown": "@nadiatriki Please find the inference code snippet as attached below.\n" }, { "id": 3222946, "postDate": "2025-06-12T17:25:34.210Z", "content": "for better time managment ,I recommend that using two separate notebooks one for training and another for submission , once trained, save your outputs, and import them as an input dataset in your submission notebook
", "rawMarkdown": "for better time managment ,I recommend that using two separate notebooks one for training and another for submission , once trained, save your outputs, and import them as an input dataset in your submission notebook " }, { "id": 3223323, "postDate": "2025-06-13T06:53:53.167Z", "content": "learnt a lot! thx!
", "rawMarkdown": "learnt a lot! thx!", "votes": 1 } ] } ] }, { "id": 3213883, "postDate": "2025-05-30T16:03:09.113Z", "content": "This one does not need any API like setup. This could have been a simple test set and it would have been a better experience @chumajin
", "rawMarkdown": "This one does not need any API like setup. This could have been a simple test set and it would have been a better experience @chumajin ", "replies": [ { "id": 3213897, "postDate": "2025-05-30T16:27:43.127Z", "content": "@ravi20076 Thanks for your comment! I totally agree!
", "rawMarkdown": "@ravi20076 Thanks for your comment! I totally agree!", "votes": 1 } ] }, { "id": 3214007, "postDate": "2025-05-30T18:31:24.157Z", "content": "Are the sequences provided by the API the same as the ones from test.csv? A little bit confused here.
@kformer No, test.csv is just a sample. This will change the unseen test dataset.
", "rawMarkdown": "@kformer No, test.csv is just a sample. This will change the unseen test dataset.", "replies": [ { "id": 3214020, "postDate": "2025-05-30T18:52:01.463Z", "content": "Ah I see, thanks!
", "rawMarkdown": "Ah I see, thanks!", "votes": 1, "isDeleted": true } ] } ] }, { "id": 3213896, "postDate": "2025-05-30T16:27:27.567Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 3213941, "author_name": "Lucas Morin", "author_url": "", "post_date": "2025-05-30T17:27:00.930000", "content": "Seems like there is only a few patients over time. They might have different behaviours. This is to prevent time leakage.
", "votes": 3, "replies": [ { "id": 3213976, "author_name": "chumajin", "author_url": "", "post_date": "2025-05-30T17:49:12.470000", "content": "@lucasmorin Thank you for comment! It has the potential time behaviours. FYI : cv and public lb is almost matched. See here.
\nhttps://www.kaggle.com/competitions/cmi-detect-behavior-with-sensor-data/discussion/582344
It takes some time before people find and exploit leakage. The time series API prevents that. For an exemple of the alternative see the DRW competition. Leaker start to appear on the LB and the absence of leakage will be enforced by manual code review… this is not better.
", "votes": 1, "replies": [] }, { "id": 3214001, "author_name": "chumajin", "author_url": "", "post_date": "2025-05-30T18:24:38.613000", "content": "@lucasmorin That's a perspective I hadn't considered. It's possible that's the case.
", "votes": 0, "replies": [] } ] }, { "id": 3215540, "author_name": "ImperfectKitto", "author_url": "", "post_date": "2025-06-02T11:16:10.697000", "content": "This competition's host upvoted you, so you might be onto something
", "votes": 1, "replies": [] } ] }, { "id": 3215584, "author_name": "Athar Sayed", "author_url": "", "post_date": "2025-06-02T12:52:02.707000", "content": "One sequence corresponds to a subject and hence the tasks corresponds to predicting BFRB or Non BFRB for a subject there are approximately 3,500 sequences in test set each corresponding to a subject
", "votes": 1, "replies": [] }, { "id": 3214391, "author_name": "TARANDEEP SINGH KHURANA", "author_url": "", "post_date": "2025-05-31T12:43:32.193000", "content": "I am facing a lot of problems in submitting my notebook, it always fails due to some error, can you help?
", "votes": 0, "replies": [ { "id": 3214429, "author_name": "chumajin", "author_url": "", "post_date": "2025-05-31T14:20:00.303000", "content": "@sahib03 It might be good to make the notebook public and create some discussion topics. I'll comment if I find the time.
", "votes": 0, "replies": [ { "id": 3215234, "author_name": "TARANDEEP SINGH KHURANA", "author_url": "", "post_date": "2025-06-01T20:43:53.453000", "content": "How do I send my notebook to you? Should I attach it with a comment?
", "votes": 0, "replies": [] }, { "id": 3215430, "author_name": "chumajin", "author_url": "", "post_date": "2025-06-02T07:45:09.320000", "content": "@sahib03 You must never send it by email. First, try looking up how to use Kaggle notebook and make it public using ChatGPT or other resources.
", "votes": 0, "replies": [] } ] }, { "id": 3215425, "author_name": "Nadia TRIKI", "author_url": "", "post_date": "2025-06-02T07:37:22.277000", "content": "had the same issue , make sure to:
\nkeep the early initialization
place your training code between the initialization and final launch
Maintain the environment checks (KAGGLE_IS_COMPETITION_RERUN) you can follow the snippet below
\n ():\n \n server = kaggle_evaluation.cmi_inference_server.CMIInferenceServer(predict)\n\n \n os.getenv():\n test_df = pd.read_csv()\n sample_seq = test_df[test_df[] == test_df[].iloc[]]\n _ = predict(pl.DataFrame(sample_seq), )\n\n server\n\n()\ninference_server = init_server()\n\n\n\n\n os.getenv():\n ()\n inference_server.serve()\n:\n ()\n :\n inference_server.run_local_gateway(\n data_paths=(\n ,\n ,\n )\n )\n Exception e:\n ()\n \nEdit: You can do the training in separate notebook (training interface) and create anothor notbook for submission in order to respect the time of the calling the interface
", "votes": 1, "replies": [ { "id": 3215432, "author_name": "chumajin", "author_url": "", "post_date": "2025-06-02T07:47:00.043000", "content": "@nadiatriki Thank you for your help!
", "votes": 0, "replies": [] }, { "id": 3217540, "author_name": "rsachin_coder", "author_url": "", "post_date": "2025-06-05T05:53:11.793000", "content": "@nadiatriki I am still facing timeout error in submitting (inference only). How much time does it generally take for you?
", "votes": -1, "replies": [] }, { "id": 3217661, "author_name": "Nadia TRIKI", "author_url": "", "post_date": "2025-06-05T09:13:40.370000", "content": "for submitting it take me around 1-2 hours, some version took less than that obviously. can you screenshot so we can offer more help
", "votes": 0, "replies": [] }, { "id": 3218358, "author_name": "rsachin_coder", "author_url": "", "post_date": "2025-06-06T05:35:15.450000", "content": "@nadiatriki Please find the inference code snippet as attached below.
", "votes": 0, "replies": [] }, { "id": 3222946, "author_name": "Nadia TRIKI", "author_url": "", "post_date": "2025-06-12T17:25:34.210000", "content": "for better time managment ,I recommend that using two separate notebooks one for training and another for submission , once trained, save your outputs, and import them as an input dataset in your submission notebook
", "votes": 0, "replies": [] }, { "id": 3223323, "author_name": "", "author_url": "", "post_date": "2025-06-13T06:53:53.167000", "content": "learnt a lot! thx!
", "votes": 1, "replies": [] } ] } ] }, { "id": 3213883, "author_name": "Ravi Ramakrishnan", "author_url": "", "post_date": "2025-05-30T16:03:09.113000", "content": "This one does not need any API like setup. This could have been a simple test set and it would have been a better experience @chumajin
", "votes": 0, "replies": [ { "id": 3213897, "author_name": "chumajin", "author_url": "", "post_date": "2025-05-30T16:27:43.127000", "content": "@ravi20076 Thanks for your comment! I totally agree!
", "votes": 1, "replies": [] } ] }, { "id": 3214007, "author_name": "", "author_url": "", "post_date": "2025-05-30T18:31:24.157000", "content": "Are the sequences provided by the API the same as the ones from test.csv? A little bit confused here.
@kformer No, test.csv is just a sample. This will change the unseen test dataset.
", "votes": 0, "replies": [ { "id": 3214020, "author_name": "", "author_url": "", "post_date": "2025-05-30T18:52:01.463000", "content": "Ah I see, thanks!
", "votes": 1, "replies": [] } ] } ] }, { "id": 3213896, "author_name": "", "author_url": "", "post_date": "2025-05-30T16:27:27.567000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "3213882": "Why does this competition require submissions on a per-sequence basis?\n\nConsidering the nature of the competition, it doesn't seem like sequential submissions should necessarily be required for this one. I wonder if there's a specific intention behind this.\n\nOne thing that catches my attention is the following statement:\n```\nNote also that there is known sensor communication failure in this dataset, resulting in missing data from some sensors in some sequences.\n```\n\nIn the previous CMI competition, I recall that the top-ranking teams seemed to be very good at imputing NaN data. I wonder if the idea here is to interpolate this missing data continuously/sequentially. This is just an idea, though.", "3213941": "Seems like there is only a few patients over time. They might have different behaviours. This is to prevent time leakage. ", "3215584": "One sequence corresponds to a subject and hence the tasks corresponds to predicting BFRB or Non BFRB for a subject there are approximately 3,500 sequences in test set each corresponding to a subject", "3214391": "I am facing a lot of problems in submitting my notebook, it always fails due to some error, can you help?", "3213883": "This one does not need any API like setup. This could have been a simple test set and it would have been a better experience @chumajin ", "3214007": "Are the sequences provided by the API the same as the ones from `test.csv`? A little bit confused here.", "3213896": "" } }