First of all, congratulations to all the winners and big respect to the organizers for creating such a challenging competition.
\nI planned to publish my solution as a paper, but my student's income didn't allow it. Also, this is my first write-up, if something isn’t clear, feel free to ask.
\nLet me give a quick summary of my solution and share the code.
This year's WEAR challenge emphasizes robustness and generalization on data from new and augmented participants. To address this, my solution is built on specialization.
\nEach sub-model uses a hybrid of Patch Embedding, a Bidirectional LSTM to capture temporal dependencies, and Attention Pooling to extract salient information from the time series.
\n1. Sliding Window and Normalization
\nThe data is segmented into 1-second windows for processing.
window_size = 50: Each data window is 50 timesteps longstride = 25: Windows have a 50% overlap to increase the number of training samples and ensure activity continuity is captured.2. Data Augmentation
\nThe following augmentations are Randomly applied during training; on each __getitem__ call, an independent draw determines whether each augmentation is applied, teaching the model to handle uncertainty:
This is the central architecture used for all four sub-models.
\nLocationSpecificModel (
\n (patch): PatchEmbed1D (...)
\n (lstm): LSTM(..., bidirectional=True)
\n (attn_pool): Sequential (...)
\n (head): Sequential (...)
\n)
Observed per-location validation
\nNo worries about the name. I call every Kaggle notebook “Testing”. ;)
", "messages": [ { "id": 3300776, "postDate": "2025-10-11T13:17:20.193Z", "content": "First of all, congratulations to all the winners and big respect to the organizers for creating such a challenging competition.
\nI planned to publish my solution as a paper, but my student's income didn't allow it. Also, this is my first write-up, if something isn’t clear, feel free to ask.
\nLet me give a quick summary of my solution and share the code.
This year's WEAR challenge emphasizes robustness and generalization on data from new and augmented participants. To address this, my solution is built on specialization.
\nEach sub-model uses a hybrid of Patch Embedding, a Bidirectional LSTM to capture temporal dependencies, and Attention Pooling to extract salient information from the time series.
\n1. Sliding Window and Normalization
\nThe data is segmented into 1-second windows for processing.
window_size = 50: Each data window is 50 timesteps longstride = 25: Windows have a 50% overlap to increase the number of training samples and ensure activity continuity is captured.2. Data Augmentation
\nThe following augmentations are Randomly applied during training; on each __getitem__ call, an independent draw determines whether each augmentation is applied, teaching the model to handle uncertainty:
This is the central architecture used for all four sub-models.
\nLocationSpecificModel (
\n (patch): PatchEmbed1D (...)
\n (lstm): LSTM(..., bidirectional=True)
\n (attn_pool): Sequential (...)
\n (head): Sequential (...)
\n)
Observed per-location validation
\nNo worries about the name. I call every Kaggle notebook “Testing”. ;)
", "rawMarkdown": "First of all, congratulations to all the winners and big respect to the organizers for creating such a challenging competition.\nI planned to publish my solution as a paper, but my student's income didn't allow it. Also, this is my first write-up, if something isn’t clear, feel free to ask.\nLet me give a quick summary of my solution and share the code.\n\n###Overview\nThis year's WEAR challenge emphasizes ****robustness**** and ****generalization**** on data from new and augmented participants. To address this, my solution is built on **specialization**.\n\nEach sub-model uses a hybrid of Patch Embedding, a Bidirectional LSTM to capture temporal dependencies, and Attention Pooling to extract salient information from the time series.\n\n###Preprocessing and Data Preparation\n\n**1. Sliding Window and Normalization**\nThe data is segmented into 1-second windows for processing.\n- `window_size = 50`: Each data window is 50 timesteps long\n- `stride = 25`: Windows have a 50% overlap to increase the number of training samples and ensure activity continuity is captured.\nNormalization uses the mean and standard deviation computed per sensor location over the entire training set.\n\n**2. Data Augmentation**\nThe following augmentations are **Randomly** applied during training; on each `__getitem__` call, an independent draw determines whether each augmentation is applied, teaching the model to handle uncertainty:\n - **Jittering**: Adds small Gaussian noise.\n - **Scaling**: Randomly scales the signal's amplitude.\n - **Rotation**: Applies a random 3D rotation.\n - **Time warping**: Stretches or compresses the signal along the time axis.\n\n###Model Architecture\nThis is the central architecture used for all four sub-models.\n`LocationSpecificModel (`\n `(patch): PatchEmbed1D (...)`\n `(lstm): LSTM(..., bidirectional=True)`\n `(attn_pool): Sequential (...)`\n `(head): Sequential (...)`\n`)`\n1. Input: 1-second window, shape (3, 50).\n2. PatchEmbed1D: split into 10 patches (0.1 s each). A Conv1d encodes each patch to a 256-dimensional vector; add positional information.\n3. Bidirectional LSTM: 4-layer Bi-LSTM processes the 10-patch sequence to learn temporal relationships in both directions.\n4. Attention Pooling: learn weights over patches and aggregate to focus on the most informative parts within the window.\n5. Classification Head: feed the pooled representation into a classification block to predict one of 19 activity classes.\n\n###Training\n- Split for Generalization: split by sbj_id. Validation uses participants unseen during training to better simulate the final test scenario with new users.\n- Optimization: Adam optimizer (lr = 1e-3) with Cross-Entropy Loss and early stopping.\n- To improve the model, just change the random score in the data arguments and hope for better luck with the RNG.\n\n**Observed per-location validation**\n- Right Arm: macro F1 ~ 0.60\n- Left Arm: macro F1 ~ 0.59\n- Right Leg: macro F1 ~ 0.61\n- Left Leg: macro F1 ~ 0.58\n\n###Link to code: [Kaggle Notebook](https://www.kaggle.com/code/schw1eg0/testing-v0-535)\nNo worries about the name. I call every Kaggle notebook “Testing”. ;)", "votes": 4 } ], "comments": [], "raw_markdown_by_id": { "3300776": "First of all, congratulations to all the winners and big respect to the organizers for creating such a challenging competition.\nI planned to publish my solution as a paper, but my student's income didn't allow it. Also, this is my first write-up, if something isn’t clear, feel free to ask.\nLet me give a quick summary of my solution and share the code.\n\n###Overview\nThis year's WEAR challenge emphasizes ****robustness**** and ****generalization**** on data from new and augmented participants. To address this, my solution is built on **specialization**.\n\nEach sub-model uses a hybrid of Patch Embedding, a Bidirectional LSTM to capture temporal dependencies, and Attention Pooling to extract salient information from the time series.\n\n###Preprocessing and Data Preparation\n\n**1. Sliding Window and Normalization**\nThe data is segmented into 1-second windows for processing.\n- `window_size = 50`: Each data window is 50 timesteps long\n- `stride = 25`: Windows have a 50% overlap to increase the number of training samples and ensure activity continuity is captured.\nNormalization uses the mean and standard deviation computed per sensor location over the entire training set.\n\n**2. Data Augmentation**\nThe following augmentations are **Randomly** applied during training; on each `__getitem__` call, an independent draw determines whether each augmentation is applied, teaching the model to handle uncertainty:\n - **Jittering**: Adds small Gaussian noise.\n - **Scaling**: Randomly scales the signal's amplitude.\n - **Rotation**: Applies a random 3D rotation.\n - **Time warping**: Stretches or compresses the signal along the time axis.\n\n###Model Architecture\nThis is the central architecture used for all four sub-models.\n`LocationSpecificModel (`\n `(patch): PatchEmbed1D (...)`\n `(lstm): LSTM(..., bidirectional=True)`\n `(attn_pool): Sequential (...)`\n `(head): Sequential (...)`\n`)`\n1. Input: 1-second window, shape (3, 50).\n2. PatchEmbed1D: split into 10 patches (0.1 s each). A Conv1d encodes each patch to a 256-dimensional vector; add positional information.\n3. Bidirectional LSTM: 4-layer Bi-LSTM processes the 10-patch sequence to learn temporal relationships in both directions.\n4. Attention Pooling: learn weights over patches and aggregate to focus on the most informative parts within the window.\n5. Classification Head: feed the pooled representation into a classification block to predict one of 19 activity classes.\n\n###Training\n- Split for Generalization: split by sbj_id. Validation uses participants unseen during training to better simulate the final test scenario with new users.\n- Optimization: Adam optimizer (lr = 1e-3) with Cross-Entropy Loss and early stopping.\n- To improve the model, just change the random score in the data arguments and hope for better luck with the RNG.\n\n**Observed per-location validation**\n- Right Arm: macro F1 ~ 0.60\n- Left Arm: macro F1 ~ 0.59\n- Right Leg: macro F1 ~ 0.61\n- Left Leg: macro F1 ~ 0.58\n\n###Link to code: [Kaggle Notebook](https://www.kaggle.com/code/schw1eg0/testing-v0-535)\nNo worries about the name. I call every Kaggle notebook “Testing”. ;)" } }