How sup3rus3r/tensor-optix works
tensor-optix is a reinforcement-learning training-loop framework, with a dataset-driven ML path and optional neuroevolution. Users select or construct an agent, pair it with a pipeline, and run episodes through a controller that learns, evaluates validation performance, adapts hyperparameters, and checkpoints or rolls back weights. The README also documents distributed actor-learning and trial-level Optuna tuning. Excerpts are partial, so links among unsampled components are included only where the README or sampled code establishes their roles; uncertain call wiring is omitted.
Main entry points: loop_controller.py Loop controller · base_pipeline.py Pipeline contract · base_agent.py Agent interface · factory.py Agent factory
Components 31
Setup
- Agent factory
tensor_optix/factory.py - Optimizer facade
tensor_optix/optimizer.py
Training loop
- Loop controller
tensor_optix/core/loop_controller.py - Pipeline contract
tensor_optix/core/base_pipeline.py - Batch pipeline
tensor_optix/pipeline/batch_pipeline.py - Vector pipeline
tensor_optix/pipeline/vector_pipeline.py - Live pipeline
tensor_optix/pipeline/live_pipeline.py - Dataset pipeline
tensor_optix/ml/dataset_pipeline.py - Episode data
tensor_optix/core/types.py - Evaluator
tensor_optix/core/base_evaluator.py
Agents and algorithms
- Agent interface
tensor_optix/core/base_agent.py - Framework adapters
tensor_optix/adapters - RL algorithms
tensor_optix/algorithms - ML agent
tensor_optix/ml/ml_agent.py
Adaptation and state
- Online optimizer
tensor_optix/core/base_optimizer.py - Backoff scheduler
tensor_optix/core/backoff_scheduler.py - Checkpoint registry
tensor_optix/core/checkpoint_registry.py - Diagnostic controller
tensor_optix/core/diagnostic_controller.py - Policy manager
tensor_optix/core/policy_manager.py - Meta controller
tensor_optix/core/meta_controller.py
Extended capabilities
- Training callbacks
tensor_optix/callbacks - Graph agent
tensor_optix/neuroevo/agent/graph_agent.py - Neuron graph
tensor_optix/neuroevo/graph/neuron_graph.py - Topology controller
tensor_optix/neuroevo/controller/topology_controller.py - Brain network
tensor_optix/neuroevo/brain_network.py - Hebbian hook
tensor_optix/neuroevo/hebbian.py - Async learner
tensor_optix/distributed/async_learner.py - Trial orchestrator
tensor_optix/orchestrator.py
Other components
- Training user
- Environment or stream
- Optuna study
Connections 39
- Training user → Agent factory: selects agent
- Training user → Optimizer facade: starts training
- Agent factory → RL algorithms: constructs
- RL algorithms → Framework adapters: uses framework
- RL algorithms → Agent interface: implements
- Optimizer facade → Loop controller: configures and runs
- Training user → Batch pipeline: provides environment
- Environment or stream → Batch pipeline: steps
- Environment or stream → Vector pipeline: parallel environments
- Environment or stream → Live pipeline: streams items
- Environment or stream → Dataset pipeline: provides batches
- Batch pipeline → Pipeline contract: implements
- Vector pipeline → Pipeline contract: implements
- Live pipeline → Pipeline contract: implements
- Dataset pipeline → Pipeline contract: implements
- Pipeline contract → Episode data: yields
- Loop controller → Pipeline contract: consumes episodes
- Loop controller → Agent interface: calls learn
- Loop controller → Evaluator: evaluates performance
- Evaluator → Pipeline contract: reads validation episodes
- Loop controller → Online optimizer: updates hyperparameters
- Loop controller → Backoff scheduler: checks convergence state
- Loop controller → Checkpoint registry: saves and restores
- Loop controller → Diagnostic controller: passes training signals
- Diagnostic controller → Agent interface: adjusts hyperparameters
- Loop controller → Training callbacks: dispatches lifecycle events
- Loop controller → Policy manager: handles dormant policy
- Policy manager → Meta controller: requests decision
- Policy manager → Checkpoint registry: loads best weights
- Agent factory → Graph agent: optional graph policy
- Graph agent → Neuron graph: uses graph policy
- Graph agent → Hebbian hook: applies learning hook
- Loop controller → Topology controller: dispatches callback
- Topology controller → Neuron graph: mutates topology
- Neuron graph → Brain network: composes regions
- Training user → Trial orchestrator: requests tuning
- Trial orchestrator → Loop controller: runs trial training
- Trial orchestrator → Optuna study: reports trial results
- Async learner → Agent interface: trains asynchronously