How samtitar/asterx works
Key Relationships to Highlight
• Data flow from raw comics to processed HDF5 files
• Feature extraction pipeline using ViT
• Demonstrate the two-branch optimization process
Technical Details
• Input formats: JPG images, CSV annotations
• Intermediate formats: HDF5 files
• Model architecture details:
• ViT backbone
• Sequential encoder
• Sampling strategies (mix, comic_level)
• Optimization branches
Implementation Notes
• scripts/data/* for preprocessing
• src/vision_transformer.py for feature extraction
• src/models.py for the main ASTERX architecture
• src/engine.py for training logic
• src/data.py for data handling