How ai-unife/edge-dnntuner works
EDGE-DNN Tuner is a hardware-aware neural architecture search framework: users configure a dataset, backend, search strategy, and constraints, then the tuner proposes and evaluates candidate networks. The controller coordinates search, model building, training, and constraint checks; optional loss modules add FLOPs or hardware costs. The repository also includes NVDLA model replacement, profiling, and quantization capabilities. Relationships involving unread entry points and subsystems are omitted or left unverified; the graph relies on the README and cited source-index references and excerpts.
Main entry points: controller.py Search controller · components/neural_network.py Network builder · hardware_module.py Hardware costs · pytorch_implementation/neural_network.py PyTorch backend
Components 23
Search and orchestration
- Search controller
components/controller.py - Search space
components/search_space.py - Bayesian search
components/bo/optimizer.py - Random search
components/random_search.py - Symbolic rules
components/tuning_rules_symbolic.py - Symbolic tuner
symbolic_tuner.py - Search experience
components/storing_experience.py
Data and training
- Dataset handling
components/dataset.py - Gesture datasets
components/gesture_dataset.py - Network builder
components/neural_network.py - TensorFlow backend
tensorflow_implementation/neural_network.py - PyTorch backend
pytorch_implementation/neural_network.py - PyTorch models
pytorch_implementation/model.py - Model zoo
nvdla/models/zoo.py
Evaluation constraints
- Loss modules
modules/module.py - FLOPs constraints
modules/loss/flops_module.py - Hardware costs
modules/loss/hardware_module.py
NVDLA deployment
- NVDLA profiler
nvdla/profiler.py - Torch replacer
nvdla/wrapper/parser/torch.py - NVDLA wrapper
nvdla/wrapper/wrapper.py - Quantization routines
nvdla/models/quantize/qlib/numpy.py - Quantizer implementations
quantizer/quantizer_POTQ.py
Other components
- Tuner user
Connections 24
- Tuner user → Search controller: configures and starts
- Search controller → Search space: uses
components/controller.py - Search controller → Symbolic rules: uses
components/controller.py - Search controller → Symbolic tuner: integrates
components/controller.py - Search controller → Search experience: uses
components/controller.py - Search controller → Loss modules: loads
components/controller.py - Search controller → Network builder: builds network
components/controller.py - Search controller → Dataset handling: reads dataset
- Network builder → Search space: uses dimensions
components/neural_network.py - Network builder → Dataset handling: uses data
components/neural_network.py - Dataset handling → Gesture datasets: imports
components/dataset.py - Network builder → TensorFlow backend: backend implementation
tensorflow_implementation/neural_network.py - Network builder → PyTorch backend: backend implementation
pytorch_implementation/neural_network.py - PyTorch backend → PyTorch models: uses model
pytorch_implementation/neural_network.py - PyTorch backend → Hardware costs: uses constraints
pytorch_implementation/neural_network.py - Loss modules → FLOPs constraints: loads by name
modules/module.py - Loss modules → Hardware costs: loads by name
modules/module.py - Hardware costs → NVDLA profiler: profiles layers
modules/loss/hardware_module.py - Hardware costs → PyTorch models: reads layer specs
modules/loss/hardware_module.py - Model zoo → Torch replacer: imports
nvdla/models/zoo.py - Torch replacer → NVDLA wrapper: constructs wrapper
nvdla/wrapper/parser/torch.py - NVDLA wrapper → Quantization routines: uses quantization
nvdla/wrapper/wrapper.py - Quantizer implementations → Quantization routines: imports routines
quantizer/quantizer_POTQ.py - Bayesian search → Quantization routines: references
components/bo/optimizer.py