My 21st solution is based on a great public notebook here. The chosen one of my submissions is an ensemble of 5 similar models(public LB: 0.433, private LB: 0.324). My best private LB is a single model(public LB: 0.421, private LB: 0.327), I publish it here.
\nThe basic structure of my model composes of 3 different Conv1d blocks, then concatenate and flatten, finally with a 4 labels output.
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3110858%2F0662912f80e49cb1bb8a7fb91a77364c%2FWX20230609-0907112x.png?generation=1686272859248587&alt=media\")
\nThe features are Time_frac, AccV, AccML, AccAP, V_ML, V_AP, ML_AP(the last 3 means the difference value between two of AccV, AccML, and AccAP). And I know Time_frac is a powerful(improve nearly 0.1) but controversial one. I use multi-class, so the labels are StartHesitation, Turn, Walking, and Normal.
My denoising code is very simple:
\ndef wavelet_denoising_2(x, wavelet=):\n coeffs = pywt.wavedec(x, wavelet, mode=)\n coeffs[(coeffs)] *= \n coeffs[(coeffs)] *= \n = pywt.waverec(coeffs, wavelet, mode=)\n (x)%==:\n = [:]\n \n = \n = \n = \n = \nI think that's all, nothing special. Thanks. :)
\nLooking forward to top and interesting solutions！
My 21st solution is based on a great public notebook here. The chosen one of my submissions is an ensemble of 5 similar models(public LB: 0.433, private LB: 0.324). My best private LB is a single model(public LB: 0.421, private LB: 0.327), I publish it here.
\nThe basic structure of my model composes of 3 different Conv1d blocks, then concatenate and flatten, finally with a 4 labels output.
\n
\nThe features are Time_frac, AccV, AccML, AccAP, V_ML, V_AP, ML_AP(the last 3 means the difference value between two of AccV, AccML, and AccAP). And I know Time_frac is a powerful(improve nearly 0.1) but controversial one. I use multi-class, so the labels are StartHesitation, Turn, Walking, and Normal.
My denoising code is very simple:
\ndef wavelet_denoising_2(x, wavelet=):\n coeffs = pywt.wavedec(x, wavelet, mode=)\n coeffs[(coeffs)] *= \n coeffs[(coeffs)] *= \n = pywt.waverec(coeffs, wavelet, mode=)\n (x)%==:\n = [:]\n \n = \n = \n = \n = \nI think that's all, nothing special. Thanks. :)
\nLooking forward to top and interesting solutions！
What were your CV values? Did you do GroupKFold?
", "rawMarkdown": "What were your CV values? Did you do GroupKFold?", "votes": null }, { "id": "2293351", "postDate": "06/09/2023 05:56:56", "content": "Nice solution! By the way, you have two \"Conv1d_2\" on your diagram - probably one should be \"Conv1d_3\"?
", "rawMarkdown": "Nice solution! By the way, you have two \"Conv1d_2\" on your diagram - probably one should be \"Conv1d_3\"?", "votes": null }, { "id": "2293355", "postDate": "06/09/2023 05:58:06", "content": "Yes, my mistake.😮
", "rawMarkdown": "Yes, my mistake.😮", "votes": null }, { "id": "2293356", "postDate": "06/09/2023 06:00:09", "content": "I use StratifiedGroupKFold(y=labels, group=Subject). CV is around 0.32-0.34.
Thank for referencing my notebook. Great work on your solution 👍
", "rawMarkdown": "Thank for referencing my notebook. Great work on your solution 👍", "votes": null }, { "id": "2294793", "postDate": "06/10/2023 10:31:48", "content": "Thanks, your notebook is well-structured and amazing.👍
", "rawMarkdown": "Thanks, your notebook is well-structured and amazing.👍", "votes": null }, { "id": "2324011", "postDate": "06/30/2023 10:12:35", "content": "Hi Joseph, I'm really impressed by the simplicity of your solution!
\nI'm trying to understand the structure of your input windows.
\nWhat do you mean by \"_future\"? are you predecting the next 50 points? what's the use of it?
\nThank you
Hi Alberto, we can use the 'future' points to predict a 'history' one. For example, if we want to predict the time T, we can use the time from T-50 to T+50 to generate features. It seems to be a leakage, but in this comp it's fine, since we get the whole series of data.
so you mean e.g. for tdcs that you are using 200 timesteps before the target event and 50 after it?
", "rawMarkdown": "so you mean e.g. for tdcs that you are using 200 timesteps before the target event and 50 after it?", "votes": null }, { "id": "2324292", "postDate": "06/30/2023 14:07:39", "content": "That's right.
", "rawMarkdown": "That's right.", "votes": null } ], "comments": [ { "id": 2293332, "author_name": "exjustice", "author_url": "", "post_date": "06/09/2023 05:19:45", "content": "What were your CV values? Did you do GroupKFold?
", "votes": null, "replies": [ { "id": 2293356, "author_name": "takanashihumbert", "author_url": "", "post_date": "06/09/2023 06:00:09", "content": "I use StratifiedGroupKFold(y=labels, group=Subject). CV is around 0.32-0.34.
Nice solution! By the way, you have two \"Conv1d_2\" on your diagram - probably one should be \"Conv1d_3\"?
", "votes": null, "replies": [ { "id": 2293355, "author_name": "takanashihumbert", "author_url": "", "post_date": "06/09/2023 05:58:06", "content": "Yes, my mistake.😮
", "votes": null, "replies": [] } ] }, { "id": 2293667, "author_name": "coderrkj", "author_url": "", "post_date": "06/09/2023 11:51:20", "content": "Thank for referencing my notebook. Great work on your solution 👍
", "votes": null, "replies": [ { "id": 2294793, "author_name": "takanashihumbert", "author_url": "", "post_date": "06/10/2023 10:31:48", "content": "Thanks, your notebook is well-structured and amazing.👍
", "votes": null, "replies": [] } ] }, { "id": 2324011, "author_name": "albertoannoni", "author_url": "", "post_date": "06/30/2023 10:12:35", "content": "Hi Joseph, I'm really impressed by the simplicity of your solution!
\nI'm trying to understand the structure of your input windows.
\nWhat do you mean by \"_future\"? are you predecting the next 50 points? what's the use of it?
\nThank you
Hi Alberto, we can use the 'future' points to predict a 'history' one. For example, if we want to predict the time T, we can use the time from T-50 to T+50 to generate features. It seems to be a leakage, but in this comp it's fine, since we get the whole series of data.
so you mean e.g. for tdcs that you are using 200 timesteps before the target event and 50 after it?
", "votes": null, "replies": [ { "id": 2324292, "author_name": "takanashihumbert", "author_url": "", "post_date": "06/30/2023 14:07:39", "content": "That's right.
", "votes": null, "replies": [] } ] } ] } ] } ], "raw_markdown_by_id": { "2293132": "My 21st solution is based on a great public notebook [here](https://www.kaggle.com/code/coderrkj/parkinson-fog-pred-conv1d-separate-tf-model). The chosen one of my submissions is an ensemble of 5 similar models(**public LB: 0.433, private LB: 0.324**). My best private LB is a single model(**public LB: 0.421, private LB: 0.327**), I publish it [here](https://www.kaggle.com/code/takanashihumbert/gait-single-models-inference/notebook).\n## model\nThe basic structure of my model composes of 3 different `Conv1d` blocks, then `concatenate` and `flatten`, finally with a 4 labels output.\n\nThe features are **Time_frac**, **AccV**, **AccML**, **AccAP**, **V_ML**, **V_AP**, **ML_AP**(the last 3 means the difference value between two of AccV, AccML, and AccAP). And I know **Time_frac** is a powerful(improve nearly 0.1) but controversial one. I use multi-class, so the labels are **StartHesitation**, **Turn**, **Walking**, and **Normal**.\n## denoising\nMy denoising code is very simple:\n```\ndef wavelet_denoising_2(x, wavelet='db4'):\n coeffs = pywt.wavedec(x, wavelet, mode=\"per\")\n coeffs[len(coeffs)-1] *= 0\n coeffs[len(coeffs)-2] *= 0\n result = pywt.waverec(coeffs, wavelet, mode='per')\n if len(x)%2==1:\n result = result[:-1]\n return result\n```\n## window size\n```\ndefog_window_size = 200\ndefog_window_future = 50\ntdcsfog_window_size = 256\ntdcsfog_window_future = 64\n```\nI think that's all, nothing special. Thanks. :)\nLooking forward to top and interesting solutions！", "2293332": "What were your CV values? Did you do GroupKFold?", "2293351": "Nice solution! By the way, you have two \"Conv1d_2\" on your diagram - probably one should be \"Conv1d_3\"?", "2293355": "Yes, my mistake.😮", "2293356": "I use `StratifiedGroupKFold`(y=labels, group=Subject). CV is around 0.32-0.34.", "2293667": "Thank for referencing my notebook. Great work on your solution 👍", "2294793": "Thanks, your notebook is well-structured and amazing.👍", "2324011": "Hi Joseph, I'm really impressed by the simplicity of your solution!\nI'm trying to understand the structure of your input windows.\nWhat do you mean by \"_future\"? are you predecting the next 50 points? what's the use of it?\nThank you", "2324032": "Hi Alberto, we can use the 'future' points to predict a 'history' one. For example, if we want to predict the time `T`, we can use the time from `T-50` to `T+50` to generate features. It seems to be a leakage, but in this comp it's fine, since we get the whole series of data.", "2324127": "so you mean e.g. for tdcs that you are using 200 timesteps before the target event and 50 after it?", "2324292": "That's right." }, "source": "meta" }