How idealab-isu/slice-100k works
Main Components
• Dataset & Data Sources: Represent the raw inputs such as G-code files, CAD models, and LVIS category definitions (evident from the LVIS_categories_final.txt in Category_Generation).
• Category Generation Module: (Directory: Category_Generation) – Processes LVIS data and potentially assigns or organizes category labels.
• Slicing Module: (Directory: slicing) – Uses configuration files (INI files and llm.yml) and scripts (e.g., slice_binary_gcode.py) to "slice" the G-code files. This likely transforms or segments the G-code into meaningful parts.
• Blender Rendering Module: (Directory: blender_rendering) – Leverages Blender scripts (blender_script.py, custom_main.py) and configuration (blender.yml) to generate STL renderings. It even includes a SLURM batch file for scheduling on clusters.
• G-code Rendering Module: (Directory: gcode_rendering) – Contains scripts (gcode_render.py) to produce visualizations or alternative renderings specifically from G-code.
• Translation & Experimentation Module: (Directory: translation_experiments) – Contains submodules:
• gcode_preprocessing: Applies transformations like chunking, contour flipping, and various preprocessing steps.
• Model Evaluation: (model_eval) – Scripts for evaluating generated data or models (e.g., eval_model.py).
• Model Training: (model_training) – Implements training strategies (e.g., sft.py).
Relationships and Data Flow
• The Slicing Module takes raw G-code files as input and produces sliced segments.
• The Blender Rendering Module uses 3D models or STL files (perhaps generated from sliced data) to produce visual renderings.
• The G-code Rendering Module may take the same or processed G-code files to produce an alternate representation.
• The Category Generation module appears to specialize in mapping or assigning LVIS categories to the dataset.
• Translation experiments work on the processed G-code data, indicating a loop that further enhances and analyzes the outputs with custom preprocessing, evaluation, and training routines.
• These modules ultimately contribute to assembling the final multimodal dataset described in the README (Slice-100K).
Architectural Patterns and Technologies
• The architecture is modular and loosely coupled—each module performs a distinct part of the data preparation and processing pipeline.
• The use of configuration files (INI, YAML) indicates a design that supports parameterization and possibly different processing scenarios.
• There are clear batch processing/automation aspects (e.g., SLURM batch file, Blender YAML) which suggest the solution may be deployed at scale.
• Each module may be developed and run independently but eventually integrates to create the unified dataset.
Detailed Elements
• Component block includes its primary files or configuration details where relevant (for instance, mention that the Blender Rendering block uses blender.yml and the execution of blender_script.py and custom_main.py).
• That the Translation Experiments contains multiple submodules (gcode preprocessing, model evaluation, and model training) to emphasize its experimental and iterative role.
• Including external integration points such as the dataset explorer and access link provided in the README, which may represent how the final dataset is exposed to users.
Final Advice
• Aim to capture every key architectural element by breaking down the project into as many logical components as possible without oversimplifying.