LesNet (Lesion Net) is an open-source project for AI-based skin lesion detection. It aims to create a reliable tool and foster community involvement in critical AI problems. Contributions are welcome!
Mozilla Public License 2.0
12
stars
2
forks
source link
M-2 Model Update: Fix Tensor Shape Mismatch & Enhance Data Processing #31
This pull request marks a significant upgrade to the M-2 model from its predecessor, transitioning away from reliance on pre-trained base models towards a fully custom architecture. This shift aims to tailor the model more closely to our specific requirements for skin cancer detection, enhancing performance and adaptability. Additionally, we outline plans for upcoming UI changes and refinements to improve user experience.
Highlights of This Upgrade
Custom Model Architecture: Developed a custom model architecture, optimized for our dataset and specific task, enhancing model performance and interpretability.
Performance Enhancements and Bug Fixes: Addressed critical issues, including a tensor shape mismatch during training, and optimized the data processing pipeline for efficiency.
Metric System Overhaul: Replaced custom metrics with TensorFlow's built-in metrics for more reliable and standardized evaluation.
Data Handling Improvements: Streamlined data augmentation and ensured consistency in data shapes and batch sizes, leveraging ImageDataGenerator.
Model Debugging and Optimization: Simplified the model for easier debugging and faster iteration, setting the stage for continuous improvements.
Impact
By moving to a custom model architecture, we expect to achieve better control over model behavior, resulting in improved accuracy and efficiency. These updates lay the groundwork for further model and UI enhancements, aiming to deliver an intuitive and effective tool for skin cancer detection.
Testing & Validation
Successfully resolved the tensor shape mismatch issue, enabling stable and efficient training.
Preliminary tests show promising improvements in model performance compared to the previous iteration.
Next Phase
UI Changes and Refinements: Plan to introduce significant UI updates to enhance user interaction and accessibility, based on user feedback and usability studies.
Continuous Model Refinement: Focus on further optimizing the model, exploring additional architectural adjustments, and improving prediction accuracy.
This pull request marks a significant upgrade to the M-2 model from its predecessor, transitioning away from reliance on pre-trained base models towards a fully custom architecture. This shift aims to tailor the model more closely to our specific requirements for skin cancer detection, enhancing performance and adaptability. Additionally, we outline plans for upcoming UI changes and refinements to improve user experience.
Highlights of This Upgrade
ImageDataGenerator
.Impact
By moving to a custom model architecture, we expect to achieve better control over model behavior, resulting in improved accuracy and efficiency. These updates lay the groundwork for further model and UI enhancements, aiming to deliver an intuitive and effective tool for skin cancer detection.
Testing & Validation
Next Phase