How rahulrajln1111/dgcnn works
This repository combines an edge-deployable DGCNN classifier with a separate, experimental hardware-aware architecture-search workflow. The main user path is to obtain ModelNet10 data, train a configurable classifier, then evaluate or benchmark a checkpoint on edge hardware. Additional capabilities include model profiling, ablation sweeps, and result visualization. The nouse/ tree contains the HGNAS search system. Some of its source was not sampled; its documented module relationships are included only where the SOURCE INDEX or excerpts support them.
Main entry points: main.py Search Orchestrator · search.py Evolutionary Search · dataset.py Point-Cloud Dataset · train_ablation.py Ablation Sweeps
Components 22
DGCNN Classifier
- Point-Cloud Dataset
dataset.py - DGCNN Model
dgcnn_model.py - Training Loop
train.py - Data Fetcher
download_data.py
Edge Evaluation
- Test Inference
inference.py - Hardware Benchmark
benchmark.py - Layer Profiling
profile_model.py
Analysis
- Ablation Sweeps
train_ablation.py - Experiment Configurations
experiments.py - Result Visualization
visualize.py
HGNAS Search
- Search Orchestrator
nouse/main.py - Search Configuration
nouse/config.py - NAS Data Loading
nouse/data_loader.py - Architecture Space
nouse/design_space.py - GNN Supernet
nouse/supernet.py - Supernet Training
nouse/trainer.py - Evolutionary Search
nouse/search.py - Hardware Predictor
nouse/hw_predictor.py - Memory Estimator
nouse/memory_estimator.py - Search Result Plots
nouse/plot_res.py
Other components
- User
- ModelNet10 Files
Connections 33
- User → Data Fetcher: fetches data
README.md - Data Fetcher → ModelNet10 Files: downloads
README.md - ModelNet10 Files → Point-Cloud Dataset: provides files
dataset.py - User → Training Loop: trains model
README.md - Training Loop → Point-Cloud Dataset: loads samples
train.py - Training Loop → DGCNN Model: optimizes
train.py - User → Test Inference: evaluates checkpoint
README.md - Test Inference → Point-Cloud Dataset: loads test data
inference.py - Test Inference → DGCNN Model: runs inference
inference.py - User → Hardware Benchmark: benchmarks device
README.md - User → Layer Profiling: profiles layers
README.md - Layer Profiling → DGCNN Model: profiles
profile_model.py - Layer Profiling → Point-Cloud Dataset: loads data
profile_model.py - User → Ablation Sweeps: runs sweeps
README.md - Ablation Sweeps → DGCNN Model: trains variants
train_ablation.py - Ablation Sweeps → Point-Cloud Dataset: loads samples
train_ablation.py - Ablation Sweeps → Experiment Configurations: uses settings
train_ablation.py - User → Result Visualization: visualizes results
README.md - Search Orchestrator → NAS Data Loading: loads data
nouse/main.py - Search Orchestrator → GNN Supernet: initializes
nouse/main.py - Search Orchestrator → Architecture Space: sets search space
nouse/main.py - Search Orchestrator → Hardware Predictor: trains predictor
nouse/main.py - Search Orchestrator → Supernet Training: trains supernet
nouse/main.py - Search Orchestrator → Evolutionary Search: runs search
nouse/main.py - Search Orchestrator → Search Result Plots: plots results
nouse/main.py - Evolutionary Search → Architecture Space: mutates architectures
nouse/search.py - Evolutionary Search → Hardware Predictor: estimates hardware cost
nouse/search.py - Evolutionary Search → Supernet Training: evaluates candidates
nouse/search.py - Evolutionary Search → GNN Supernet: scores candidates
nouse/search.py - Supernet Training → GNN Supernet: trains paths
nouse/trainer.py - Supernet Training → Architecture Space: samples architectures
nouse/trainer.py - Hardware Predictor → Architecture Space: encodes architectures
nouse/hw_predictor.py - Hardware Predictor → Memory Estimator: uses estimates
nouse/hw_predictor.py