SkinVestigatorAI Model V0.1.5 Release Notes
This release brings numerous updates and improvements to the SkinVestigatorAI model, along with a variety of bug fixes.
Major Changes:
- Model Architecture: The model has been updated to be transformer-based, providing a robust and efficient architecture for sequence understanding and transduction models.
- Performance Metrics: Added various metrics including F1-score, Specificity, and AUC-ROC to the model evaluation process, providing a comprehensive view of the model's performance.
- Training Parameters: The training parameters were updated for better optimization and quicker training.
- Quantization: The model now supports quantization, enhancing model efficiency without significant loss in model performance.
Improvements:
- The GPU memory usage has been optimized by reducing the batch size and image size, and implementing mixed precision training.
- The data scraper has been optimized to balance the data before training.
- The predict view has been improved to automatically select the latest model available in the models folder, and to automatically load in the TFLite model if the system is resource-limited.
Bug Fixes:
- Fixed a bug that was causing the test data to not load correctly.
- Fixed a bug related to the model parameters.
- Fixed a bug that occurred when loading the model with custom metrics.
Other Changes:
- Adjusted the hyperparameters to mitigate overfitting and enhance the learning rate adaptation. A cooldown period was also introduced to the ReduceLROnPlateau callback for more stable training.
- Replaced custom precision and recall metrics with TensorFlow's Precision and Recall, while keeping F1 score and specificity as custom metrics due to the absence of direct TensorFlow functions.
- Refactored the SkinCancerDetector class for improved readability and modularity, and to fix a TensorFlow variable error.
- Added a print of the model summary on the load of the predict view.
- Updated the README.md to include the latest model information.
Model Performance:
Test Accuracy: 84.04%
Test Loss: 0.23201
Test Sensitivity: 84.04%
Test Precision: 84.04%
Test F1-score: 84.47%
Test Specificity: 84.02%
Test AUC-ROC: 91.69%
Thank you for your continued support. We look forward to your feedback on this release.