How jaryp/adaptivesemantictokencommunication works
This repository implements research experiments for adaptive semantic-token communication and edge inference. The primary workflow configures and trains or loads a vision model, evaluates it, and can insert a JSCC encoder/channel/decoder path before communication evaluation. A separate optimization entry point analyzes communication results for resource allocation. The graph preserves the adaptive ViT, channel, baseline evaluation, and optimization branches; some config-driven wiring is only documented or referenced in the source index, so uncertain calls are omitted.
Main entry points: main.py Train and evaluate · main_opt.py Optimize allocation · evaluation.py Communication evaluation · channel.py Noisy channel
Components 16
Experiment entry points
- Train and evaluate
main.py - Optimize allocation
main_opt.py - Experiment configs
configs
Adaptive inference
- Adaptive ViT
methods/proposal/base.py - Token regularization
methods/proposal/loss.py
Communication
- JSCC pipeline
utils.py - Noisy channel
comm/channel.py - Pipeline construction
comm/pipeline.py
Evaluation and optimization
- Semantic evaluation
methods/proposal/evaluation.py - Compute cost
methods/flops_count.py - Baseline evaluation
methods/base.py - Communication evaluation
comm/evaluation.py
Runtime support
- Run naming and serialization
serialization.py - Training helpers
utils.py
Other components
- Experimenter
- Vision model provider
Connections 22
- Experimenter → Train and evaluate: runs
README.md - Experimenter → Optimize allocation: runs
README.md - Experiment configs → Train and evaluate: configures
main.py - Experiment configs → Optimize allocation: configures
main_opt.py - Train and evaluate → Adaptive ViT: uses
main.py - Train and evaluate → JSCC pipeline: constructs
main.py - Train and evaluate → Noisy channel: uses
main.py - Train and evaluate → Communication evaluation: runs
main.py - Train and evaluate → Baseline evaluation: uses
main.py - Train and evaluate → Run naming and serialization: uses
main.py - Train and evaluate → Training helpers: uses
main.py - Train and evaluate → Vision model provider: wraps
README.md - Adaptive ViT → Token regularization: regularized by
methods/proposal/__init__.py - Semantic evaluation → Adaptive ViT: evaluates
methods/proposal/evaluation.py - Semantic evaluation → Compute cost: measures with
methods/proposal/evaluation.py - Baseline evaluation → Noisy channel: reads symbols
methods/base.py - Communication evaluation → JSCC pipeline: inspects
comm/evaluation.py - Communication evaluation → Noisy channel: reads
comm/evaluation.py - JSCC pipeline → Noisy channel: transmits through
utils.py - Train and evaluate → Pipeline construction: uses
- Optimize allocation → Communication evaluation: uses results from
main_opt.py - Optimize allocation → Run naming and serialization: uses
main_opt.py