![](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F13474157%2F6746c1774c63e26d52986939d718f401%2FFOG%20CNN%20Architecture.jpeg?generation=1686624959127555&alt=media\")
Thanks to the hosts for providing such an interesting/worthwhile problem to work on, and to everyone who shared their code and insights throughout the competition! This is my first completed Kaggle competition; I feel like I spent as much time learning about the Kaggle ecosystem as I did about the FOG data 😂 but overall it was a fun experience.
\nMy solution was a fairly straightforward 1D CNN ensemble, but I still wanted to share it!
\nMy code: https://www.kaggle.com/code/abandura/fog-1d-cnn
\nI referenced this public notebook when I began the comp, and wanted to thank the author for sharing their work.
\nTS window dimensions:
\nwindow_size = \nwindow_future = \nwindow_past = window_size - window_future\nTS Features (standardized for each file):
\nRaw 3D accelerometer data\nSeperate 3D acc data into high freq low freq components MA \nSpectrogram of 3D accelerometer data (NFFT = )\nTemporal location of sequence**: time_index/total_time\n**Some people feel that using temporal features reduces model usability, but I disagree. These models could be useful for auto-labeling more data from the same FOG-inducing protocols, in which case temporal features would be consistent to the data we used in this competition. On the flip side, I doubt these models would generalize well to something like the daily living dataset regardless of whether temporal features were leveraged. Just my two cents.
\nCV strategy: 6 fold on patients, separating patients so each fold has roughly the same amount of data
\n
Public LB: 0.362
\nPrivate LB: 0.305
Thanks to the hosts for providing such an interesting/worthwhile problem to work on, and to everyone who shared their code and insights throughout the competition! This is my first completed Kaggle competition; I feel like I spent as much time learning about the Kaggle ecosystem as I did about the FOG data 😂 but overall it was a fun experience.
\nMy solution was a fairly straightforward 1D CNN ensemble, but I still wanted to share it!
\nMy code: https://www.kaggle.com/code/abandura/fog-1d-cnn
\nI referenced this public notebook when I began the comp, and wanted to thank the author for sharing their work.
\nTS window dimensions:
\nwindow_size = \nwindow_future = \nwindow_past = window_size - window_future\nTS Features (standardized for each file):
\nRaw 3D accelerometer data\nSeperate 3D acc data into high freq low freq components MA \nSpectrogram of 3D accelerometer data (NFFT = )\nTemporal location of sequence**: time_index/total_time\n**Some people feel that using temporal features reduces model usability, but I disagree. These models could be useful for auto-labeling more data from the same FOG-inducing protocols, in which case temporal features would be consistent to the data we used in this competition. On the flip side, I doubt these models would generalize well to something like the daily living dataset regardless of whether temporal features were leveraged. Just my two cents.
\nCV strategy: 6 fold on patients, separating patients so each fold has roughly the same amount of data
\n
Public LB: 0.362
\nPrivate LB: 0.305
\n\n###Training Parameters\n- batch_size = 1024\n- learning_rate = 1e-4\n- num_epochs = 1\n- max_batches_per_epoch = 12000\n- max_val_batches = 500\n- val_freq = 500\n- loss function:\n\tFor tdcsfog I used cross entropy loss with inverse class frequency weights\n\tFor defog I used focal loss with gamma = 5\n\n**Public LB**: 0.362\n**Private LB**: 0.305\n\n#### Other Ideas I tried that didn't improve accuracy:\n- Combined models: single CNN model for both tdcsfog and defog after unifying sampling rates and units. \n- Multi-head CNN model: Add a second loss function for presence/absence of FOG event so that the notype data could be leveraged.\n- CNN features as inputs for a Light GBM model: Use the output of the convolutional blocks as input to a GBM model, along with other temporal and patient features.",
"votes": null
},
{
"id": "2303298",
"postDate": "06/15/2023 08:16:38",
"content": "Nice work! I noticed you didnt use GroupKFold, but manually defined folds at the patient level, which I think is the same? Did you observe any systematic differences between LB and CV?
", "rawMarkdown": "Nice work! I noticed you didnt use GroupKFold, but manually defined folds at the patient level, which I think is the same? Did you observe any systematic differences between LB and CV?", "votes": null }, { "id": "2305393", "postDate": "06/16/2023 17:03:35", "content": "Thanks! I think you're right that GroupKFold would be similar to the way I manually split up the folds--I did that so I'd have a little more control over the process. My CV tended to be lower than LB, but they were well correlated so still useful for selecting models. I think I saw discussions of other folks seeing a similar trend?
", "rawMarkdown": "Thanks! I think you're right that GroupKFold would be similar to the way I manually split up the folds--I did that so I'd have a little more control over the process. My CV tended to be lower than LB, but they were well correlated so still useful for selecting models. I think I saw discussions of other folks seeing a similar trend?", "votes": null }, { "id": "2479910", "postDate": "10/13/2023 02:00:44", "content": "thanks for sharing，really learned a lot. simple solution also leads to a good result!😃
", "rawMarkdown": "thanks for sharing，really learned a lot. simple solution also leads to a good result!😃", "votes": null } ], "comments": [ { "id": 2303298, "author_name": "exjustice", "author_url": "", "post_date": "06/15/2023 08:16:38", "content": "Nice work! I noticed you didnt use GroupKFold, but manually defined folds at the patient level, which I think is the same? Did you observe any systematic differences between LB and CV?
", "votes": null, "replies": [ { "id": 2305393, "author_name": "abandura", "author_url": "", "post_date": "06/16/2023 17:03:35", "content": "Thanks! I think you're right that GroupKFold would be similar to the way I manually split up the folds--I did that so I'd have a little more control over the process. My CV tended to be lower than LB, but they were well correlated so still useful for selecting models. I think I saw discussions of other folks seeing a similar trend?
", "votes": null, "replies": [] } ] }, { "id": 2479910, "author_name": "roger92", "author_url": "", "post_date": "10/13/2023 02:00:44", "content": "thanks for sharing，really learned a lot. simple solution also leads to a good result!😃
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2300104": "Thanks to the hosts for providing such an interesting/worthwhile problem to work on, and to everyone who shared their code and insights throughout the competition! This is my first completed Kaggle competition; I feel like I spent as much time learning about the Kaggle ecosystem as I did about the FOG data 😂 but overall it was a fun experience.\n\nMy solution was a fairly straightforward 1D CNN ensemble, but I still wanted to share it!\n\nMy code: https://www.kaggle.com/code/abandura/fog-1d-cnn\n\nI referenced [this](https://www.kaggle.com/code/coderrkj/parkinson-fog-pred-conv1d-separate-tf-model) public notebook when I began the comp, and wanted to thank the author for sharing their work.\n\n## Solution Overview\n\n### Preprocessing\n\nTS window dimensions:\n```python\nwindow_size = 200\nwindow_future = 75\nwindow_past = window_size - window_future\n```\n\nTS Features (standardized for each file):\n```python\nRaw 3D accelerometer data\nSeperate 3D acc data into high freq and low freq components with MA filter\nSpectrogram of 3D accelerometer data (NFFT = 8)\nTemporal location of sequence**: time_index/total_time\n```\n\t\n**Some people feel that using temporal features reduces model usability, but I disagree. These models could be useful for auto-labeling more data from the same FOG-inducing protocols, in which case temporal features would be consistent to the data we used in this competition. On the flip side, I doubt these models would generalize well to something like the daily living dataset regardless of whether temporal features were leveraged. Just my two cents.\n\nCV strategy: 6 fold on patients, separating patients so each fold has roughly the same amount of data\n\n###Model Architecture\n- The best models based on CV for each fold were ensembled using a weighted average.\n- There are batch normalization layers between each convolutional layer even though they aren't included in the diagram.\n\n\n\n###Training Parameters\n- batch_size = 1024\n- learning_rate = 1e-4\n- num_epochs = 1\n- max_batches_per_epoch = 12000\n- max_val_batches = 500\n- val_freq = 500\n- loss function:\n\tFor tdcsfog I used cross entropy loss with inverse class frequency weights\n\tFor defog I used focal loss with gamma = 5\n\n**Public LB**: 0.362\n**Private LB**: 0.305\n\n#### Other Ideas I tried that didn't improve accuracy:\n- Combined models: single CNN model for both tdcsfog and defog after unifying sampling rates and units. \n- Multi-head CNN model: Add a second loss function for presence/absence of FOG event so that the notype data could be leveraged.\n- CNN features as inputs for a Light GBM model: Use the output of the convolutional blocks as input to a GBM model, along with other temporal and patient features.",
"2303298": "Nice work! I noticed you didnt use GroupKFold, but manually defined folds at the patient level, which I think is the same? Did you observe any systematic differences between LB and CV?",
"2305393": "Thanks! I think you're right that GroupKFold would be similar to the way I manually split up the folds--I did that so I'd have a little more control over the process. My CV tended to be lower than LB, but they were well correlated so still useful for selecting models. I think I saw discussions of other folks seeing a similar trend?",
"2479910": "thanks for sharing，really learned a lot. simple solution also leads to a good result!😃"
},
"source": "meta"
}