How kchen92/text2shape works
Main Components to Include:
Data Processing Pipeline
• Input Sources:
• ShapeNet voxelizations
• Primitives dataset
• Text descriptions/captions
• Processing Components:
• NRRD file handling (nrrd_rw.py)
• Data preprocessing (preprocess.py)
• Various data processing modules (data_process*.py files)
Core Architecture Components
a. Text Encoder:
• Text embedding generation
• Learning joint embeddings
• Located in models/encoder_component.py
• Learning-by-association (LBA) components
• Located in models/lba_component.py and models/lba_models.py
c. Conditional Wasserstein GAN (CWGAN):
• Generator (t2s_generator_component.py)
• Critic (t2s_critic_component.py)
• WGAN training logic (cwgan.py)
• Classification models (classifier_models.py)
• Training and evaluation components
Training Pipeline
• Loss functions (losses.py)
• Optimizers (optimizers.py)
• Solvers for different components (solver*.py files)
• Network components (network.py, net_component.py)
Use distinct sections with clear boundaries
• Data Input & Processing (blue section)
• Model Architecture (green section)
• Training Pipeline (orange section)
• Output Generation (purple section)
Flow Direction
• Main data flow should move left to right
• Feedback loops should be shown for GAN training
Component Relationships
• Where embeddings merge
Specific Elements
• Embedding space visualization
• GAN architecture with generator and critic
• Classification pathway
• Output visualization component
Technical Details to Note
• Configuration parameters from improved_wgan.yaml
• Use of NRRD format for voxel data
• Integration points with SmartScenes Toolkit
• Batch sizes and learning rates
Visualization Guidelines:
Use standardized symbols
• Rectangles: Processing modules
• Diamonds: Decision points
Layer Organization
• Top layer: Data input and preprocessing
• Middle layer: Core neural network architecture
• Bottom layer: Training and optimization
• Side panel: Configuration and utilities