![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3197853%2F5561d2b2c2638ed84ccf08fa1059b08a%2FScreenshot%202025-06-05%20at%2010.09.35.png?generation=1749130065965897&alt=media\")
I am sharing my WaveNet approach using IMU-only data using PyTorch:
\n\nThis approach's model is a 2-stream WaveNet model with Temporal attention and 2 classification heads, 1 for the multiclass classification task and another one for the binary classification task. For the multiclass problem, Non-Target gestures were collapsed into a single class, thus the last last linear layer contains 9 classes instead of 18. The acceleration data and the rotation data are handled independently and are later concatenated.
Here is an overview diagram of the architecture:
\n
This approach achieves CV: ~0.66 LB: 0.65 with a StratifiedGroupKFold validation taking subject as groups.
Please, leave your thoughts, comments and improvements! Thanks!
\nHappy Kaggling
\nEdit: enlarging the WaveNet improved scrore CV and LB: 0.65 -> 0.68
", "messages": [ { "id": "3217881", "postDate": "06/05/2025 13:30:58", "content": "I am sharing my WaveNet approach using IMU-only data using PyTorch:
\n\nThis approach's model is a 2-stream WaveNet model with Temporal attention and 2 classification heads, 1 for the multiclass classification task and another one for the binary classification task. For the multiclass problem, Non-Target gestures were collapsed into a single class, thus the last last linear layer contains 9 classes instead of 18. The acceleration data and the rotation data are handled independently and are later concatenated.
Here is an overview diagram of the architecture:
\n
This approach achieves CV: ~0.66 LB: 0.65 with a StratifiedGroupKFold validation taking subject as groups.
Please, leave your thoughts, comments and improvements! Thanks!
\nHappy Kaggling
\nEdit: enlarging the WaveNet improved scrore CV and LB: 0.65 -> 0.68
", "rawMarkdown": "I am sharing my WaveNet approach using IMU-only data using PyTorch:\n- [CMI | WaveNet PyTorch IMU-Only [Train]][1]\n- [CMI | WaveNet PyTorch IMU-Only [Inference]][2]\n\nThis approach's model is a 2-stream WaveNet model with Temporal attention and 2 classification heads, 1 for the multiclass classification task and another one for the binary classification task. For the multiclass problem, `Non-Target` gestures were collapsed into a single class, thus the last last linear layer contains 9 classes instead of 18. The acceleration data and the rotation data are handled independently and are later concatenated.\n\nHere is an overview diagram of the architecture:\n\n\n\nThis approach achieves CV: ~0.66 LB: 0.65 with a `StratifiedGroupKFold` validation taking `subject` as groups.\n\nPlease, leave your thoughts, comments and improvements! Thanks!\n\nHappy Kaggling\n\nEdit: enlarging the WaveNet improved scrore CV and LB: 0.65 -> 0.68\n\n[1]: https://www.kaggle.com/code/alejopaullier/cmi-wavenet-pytorch-imu-only-train\n[2]: https://www.kaggle.com/code/alejopaullier/cmi-wavenet-pytorch-imu-only-inference", "votes": null }, { "id": "3217883", "postDate": "06/05/2025 13:33:48", "content": "Thank you for sharing! I noticed that you designed two parallel classification heads. In your experiments, did this design lead to better performance?
", "rawMarkdown": "Thank you for sharing! I noticed that you designed two parallel classification heads. In your experiments, did this design lead to better performance?", "votes": null }, { "id": "3217913", "postDate": "06/05/2025 14:09:27", "content": "Not much tbh. Most perfomance boosts came from trying different configurations of batch size, LR, epochs.
", "rawMarkdown": "Not much tbh. Most perfomance boosts came from trying different configurations of batch size, LR, epochs.", "votes": null }, { "id": "3217917", "postDate": "06/05/2025 14:16:31", "content": "Have you tried adding demographic data (numerical and categorical features)? I've tried doing that by adding another input stream that would create a dense representation out of those features, and concatenating with IMU activations. Like this:
\n = Input(shape=(pad_len_to_use, len(sensor_feature_cols)), name=)\n = Conv1D(filters=, kernel_size=, activation=)(sensor_input)\n = MaxPooling1D(pool_size=)(x)\n = Dropout()(x)\n = Conv1D(filters=, kernel_size=, activation=)(x)\n = MaxPooling1D(pool_size=)(x)\n = Dropout()(x)\n = Flatten()(x)\n\n\n = Input(shape=(len(demo_feature_cols),), name=)\n = Dense(, activation=)(participant_input)\n = Dropout()(y)\n\n\n = concatenate([x, y])\n = Dense(, activation=)(combined)\n = Dropout()(z)\n = Dense(, activation=)(z)\n = Dropout()(z)\n = Dense(num_classes, activation=)(z)\nI found it improves my CV by a little, have some difficulties with actual inference though
", "rawMarkdown": "Have you tried adding demographic data (numerical and categorical features)? I've tried doing that by adding another input stream that would create a dense representation out of those features, and concatenating with IMU activations. Like this:\n```\nsensor_input = Input(shape=(pad_len_to_use, len(sensor_feature_cols)), name='sensor_input')\nx = Conv1D(filters=128, kernel_size=5, activation='relu')(sensor_input)\nx = MaxPooling1D(pool_size=2)(x)\nx = Dropout(0.4)(x)\nx = Conv1D(filters=256, kernel_size=2, activation='relu')(x)\nx = MaxPooling1D(pool_size=2)(x)\nx = Dropout(0.4)(x)\nx = Flatten()(x)\n\n# Demo\nparticipant_input = Input(shape=(len(demo_feature_cols),), name='demographic_data')\ny = Dense(64, activation='relu')(participant_input)\ny = Dropout(0.3)(y)\n\n# Combine, then dense\ncombined = concatenate([x, y])\nz = Dense(256, activation='relu')(combined)\nz = Dropout(0.5)(z)\nz = Dense(128, activation='relu')(z)\nz = Dropout(0.3)(z)\noutput = Dense(num_classes, activation='softmax')(z)\n```\nI found it improves my CV by a little, have some difficulties with actual inference though", "votes": null }, { "id": "3217944", "postDate": "06/05/2025 15:05:49", "content": "I wanted to try an IMU-only approach. What you can do is train and infer with your model and create OOF predictions. Then use these leak-less predictions as new columns together with demographic data and train a GBM model instead of directly inputting the demographics data in the neural network.
", "rawMarkdown": "I wanted to try an IMU-only approach. What you can do is train and infer with your model and create OOF predictions. Then use these leak-less predictions as new columns together with demographic data and train a GBM model instead of directly inputting the demographics data in the neural network.", "votes": null }, { "id": "3218055", "postDate": "06/05/2025 18:05:26", "content": "Hey Alejo! Cool write up, I have mixed feelings about that model though ;).
", "rawMarkdown": "Hey Alejo! Cool write up, I have mixed feelings about that model though ;).", "votes": null }, { "id": "3218057", "postDate": "06/05/2025 18:10:08", "content": "Hi Jan! Long time no see! :)
", "rawMarkdown": "Hi Jan! Long time no see! :)", "votes": null }, { "id": "3219920", "postDate": "06/08/2025 14:05:10", "content": "Curious if you tried a one wavenet version where you concat acc and rot and pass everything in together. My initiation tells me that since there's not a ton of data here, creating separate models for each sensor type makes more sense than a pass it all in at once approach.
", "rawMarkdown": "Curious if you tried a one wavenet version where you concat acc and rot and pass everything in together. My initiation tells me that since there's not a ton of data here, creating separate models for each sensor type makes more sense than a pass it all in at once approach.", "votes": null }, { "id": "3220036", "postDate": "06/08/2025 17:26:20", "content": "Haha @janbrederecke type of comp.. be nice model reminds of your old models :)
", "rawMarkdown": "Haha @janbrederecke type of comp.. be nice model reminds of your old models :)", "votes": null }, { "id": "3220038", "postDate": "06/08/2025 17:28:12", "content": "Yeah, there is definitely some stuff that could be recycled - hit me up if you want to team up again! 😊
", "rawMarkdown": "Yeah, there is definitely some stuff that could be recycled - hit me up if you want to team up again! 😊", "votes": null }, { "id": "3220043", "postDate": "06/08/2025 17:37:31", "content": "Ya this reminded me of h$# :) came back from a hiatus (finally some break from office work).. just trying few ideas then surely gang up 😄
", "rawMarkdown": "Ya this reminded me of h$# :) came back from a hiatus (finally some break from office work).. just trying few ideas then surely gang up 😄", "votes": null } ], "comments": [ { "id": 3217883, "author_name": "ryenhails", "author_url": "", "post_date": "06/05/2025 13:33:48", "content": "Thank you for sharing! I noticed that you designed two parallel classification heads. In your experiments, did this design lead to better performance?
", "votes": null, "replies": [ { "id": 3217913, "author_name": "alejopaullier", "author_url": "", "post_date": "06/05/2025 14:09:27", "content": "Not much tbh. Most perfomance boosts came from trying different configurations of batch size, LR, epochs.
", "votes": null, "replies": [] } ] }, { "id": 3217917, "author_name": "defdet", "author_url": "", "post_date": "06/05/2025 14:16:31", "content": "Have you tried adding demographic data (numerical and categorical features)? I've tried doing that by adding another input stream that would create a dense representation out of those features, and concatenating with IMU activations. Like this:
\n = Input(shape=(pad_len_to_use, len(sensor_feature_cols)), name=)\n = Conv1D(filters=, kernel_size=, activation=)(sensor_input)\n = MaxPooling1D(pool_size=)(x)\n = Dropout()(x)\n = Conv1D(filters=, kernel_size=, activation=)(x)\n = MaxPooling1D(pool_size=)(x)\n = Dropout()(x)\n = Flatten()(x)\n\n\n = Input(shape=(len(demo_feature_cols),), name=)\n = Dense(, activation=)(participant_input)\n = Dropout()(y)\n\n\n = concatenate([x, y])\n = Dense(, activation=)(combined)\n = Dropout()(z)\n = Dense(, activation=)(z)\n = Dropout()(z)\n = Dense(num_classes, activation=)(z)\nI found it improves my CV by a little, have some difficulties with actual inference though
", "votes": null, "replies": [ { "id": 3217944, "author_name": "alejopaullier", "author_url": "", "post_date": "06/05/2025 15:05:49", "content": "I wanted to try an IMU-only approach. What you can do is train and infer with your model and create OOF predictions. Then use these leak-less predictions as new columns together with demographic data and train a GBM model instead of directly inputting the demographics data in the neural network.
", "votes": null, "replies": [] } ] }, { "id": 3218055, "author_name": "janbrederecke", "author_url": "", "post_date": "06/05/2025 18:05:26", "content": "Hey Alejo! Cool write up, I have mixed feelings about that model though ;).
", "votes": null, "replies": [ { "id": 3218057, "author_name": "alejopaullier", "author_url": "", "post_date": "06/05/2025 18:10:08", "content": "Hi Jan! Long time no see! :)
", "votes": null, "replies": [] }, { "id": 3220036, "author_name": "gauravbrills", "author_url": "", "post_date": "06/08/2025 17:26:20", "content": "Haha @janbrederecke type of comp.. be nice model reminds of your old models :)
", "votes": null, "replies": [ { "id": 3220038, "author_name": "janbrederecke", "author_url": "", "post_date": "06/08/2025 17:28:12", "content": "Yeah, there is definitely some stuff that could be recycled - hit me up if you want to team up again! 😊
", "votes": null, "replies": [ { "id": 3220043, "author_name": "gauravbrills", "author_url": "", "post_date": "06/08/2025 17:37:31", "content": "Ya this reminded me of h$# :) came back from a hiatus (finally some break from office work).. just trying few ideas then surely gang up 😄
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 3219920, "author_name": "msthil", "author_url": "", "post_date": "06/08/2025 14:05:10", "content": "Curious if you tried a one wavenet version where you concat acc and rot and pass everything in together. My initiation tells me that since there's not a ton of data here, creating separate models for each sensor type makes more sense than a pass it all in at once approach.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3217881": "I am sharing my WaveNet approach using IMU-only data using PyTorch:\n- [CMI | WaveNet PyTorch IMU-Only [Train]][1]\n- [CMI | WaveNet PyTorch IMU-Only [Inference]][2]\n\nThis approach's model is a 2-stream WaveNet model with Temporal attention and 2 classification heads, 1 for the multiclass classification task and another one for the binary classification task. For the multiclass problem, `Non-Target` gestures were collapsed into a single class, thus the last last linear layer contains 9 classes instead of 18. The acceleration data and the rotation data are handled independently and are later concatenated.\n\nHere is an overview diagram of the architecture:\n\n\n\nThis approach achieves CV: ~0.66 LB: 0.65 with a `StratifiedGroupKFold` validation taking `subject` as groups.\n\nPlease, leave your thoughts, comments and improvements! Thanks!\n\nHappy Kaggling\n\nEdit: enlarging the WaveNet improved scrore CV and LB: 0.65 -> 0.68\n\n[1]: https://www.kaggle.com/code/alejopaullier/cmi-wavenet-pytorch-imu-only-train\n[2]: https://www.kaggle.com/code/alejopaullier/cmi-wavenet-pytorch-imu-only-inference", "3217883": "Thank you for sharing! I noticed that you designed two parallel classification heads. In your experiments, did this design lead to better performance?", "3217913": "Not much tbh. Most perfomance boosts came from trying different configurations of batch size, LR, epochs.", "3217917": "Have you tried adding demographic data (numerical and categorical features)? I've tried doing that by adding another input stream that would create a dense representation out of those features, and concatenating with IMU activations. Like this:\n```\nsensor_input = Input(shape=(pad_len_to_use, len(sensor_feature_cols)), name='sensor_input')\nx = Conv1D(filters=128, kernel_size=5, activation='relu')(sensor_input)\nx = MaxPooling1D(pool_size=2)(x)\nx = Dropout(0.4)(x)\nx = Conv1D(filters=256, kernel_size=2, activation='relu')(x)\nx = MaxPooling1D(pool_size=2)(x)\nx = Dropout(0.4)(x)\nx = Flatten()(x)\n\n# Demo\nparticipant_input = Input(shape=(len(demo_feature_cols),), name='demographic_data')\ny = Dense(64, activation='relu')(participant_input)\ny = Dropout(0.3)(y)\n\n# Combine, then dense\ncombined = concatenate([x, y])\nz = Dense(256, activation='relu')(combined)\nz = Dropout(0.5)(z)\nz = Dense(128, activation='relu')(z)\nz = Dropout(0.3)(z)\noutput = Dense(num_classes, activation='softmax')(z)\n```\nI found it improves my CV by a little, have some difficulties with actual inference though", "3217944": "I wanted to try an IMU-only approach. What you can do is train and infer with your model and create OOF predictions. Then use these leak-less predictions as new columns together with demographic data and train a GBM model instead of directly inputting the demographics data in the neural network.", "3218055": "Hey Alejo! Cool write up, I have mixed feelings about that model though ;).", "3218057": "Hi Jan! Long time no see! :)", "3219920": "Curious if you tried a one wavenet version where you concat acc and rot and pass everything in together. My initiation tells me that since there's not a ton of data here, creating separate models for each sensor type makes more sense than a pass it all in at once approach.", "3220036": "Haha @janbrederecke type of comp.. be nice model reminds of your old models :)", "3220038": "Yeah, there is definitely some stuff that could be recycled - hit me up if you want to team up again! 😊", "3220043": "Ya this reminded me of h$# :) came back from a hiatus (finally some break from office work).. just trying few ideas then surely gang up 😄" }, "source": "meta" }