v2.3.4

v2.3.4 Release Notes

Major Changes

• Replaced KerasSurgeon with TensorFlow Model Optimization (TFMOT) for filter pruning
• Migrated codebase to TensorFlow 2.x API for better compatibility
• Implemented proper layer-by-layer pruning sequence from lower to higher layers

Improvements

• Fixed reward function calculation to better match paper's formula
• Added filter tracking system to monitor pruning decisions across episodes
• Improved validation metrics and performance stability
• Added GPU memory management to prevent OOM errors
• Corrected data preprocessing (fixed rescaling from 1./225 to 1./255)

Bug Fixes

• Fixed action representation for binary decisions in policy gradient
• Resolved issues with loss function selection for model training
• Added proper validation split for early stopping
• Fixed directory structure creation for saved models

Documentation

• Enhanced comments explaining the paper's methodology implementation
• Updated requirements.txt with appropriate versions

This release significantly improves pruning performance and stability while bringing the implementation closer to the original paper's approach using modern TensorFlow capabilities.