| language: en | |
| tags: | |
| - eeg | |
| - meg | |
| - pytorch | |
| - neuroimaging | |
| license: mit | |
| datasets: | |
| - gabrycina/eeg2meg-tiny | |
| metrics: | |
| - mse | |
| # EEG to MEG Prediction Model | |
| This model was trained to predict MEG signals from EEG recordings. | |
| ## Training Configuration | |
| - Dataset: gabrycina/eeg2meg-tiny | |
| - Batch Size: 32 | |
| - Learning Rate: 0.0001 | |
| - Device: mps | |
| - Training Date: 20250104_185119 | |
| ## Performance | |
| - Best Validation Loss: 0.171059 | |
| - Best Epoch: 100 | |
| ## Model Description | |
| This model uses a deep learning architecture to predict MEG signals from EEG recordings. The architecture includes: | |
| - Frequency and temporal convolutions for feature extraction | |
| - Multi-head attention mechanisms for sensor relationships | |
| - Residual connections for better gradient flow | |
| - Separate prediction heads for magnetometers and gradiometers | |
| ## Usage | |
| ```python | |
| import torch | |
| # Load the model | |
| model = torch.load('best_model.pth') | |
| # Prepare your EEG data (shape: [batch_size, channels, time_points]) | |
| # Make predictions | |
| with torch.no_grad(): | |
| meg_predictions = model(eeg_data) | |
| ``` | |