How texasinstruments/tinyml-tensorlab works
Tiny ML Tensorlab is TI’s MCU AI toolchain. The principal workflow starts with a user’s natural-language or configured task, loads and analyzes a modality-specific dataset, trains or selects a model, optionally searches and quantizes it, compiles it for a supported MCU, and deploys or validates the embedded result. The tree confirms task-specific runners, dataset loaders, model zoo metadata, NAS, quantization, compilation, and device deployment. Several cross-package calls are inferred only from documented capabilities and are marked optional or omitted where source excerpts do not show ownership.
Main entry points: runner.py Workflow Agent · train_base.py Training References · timeseries_dataset.py Dataset Loaders · run_tinyml_modelmaker.py Modelmaker Entry
Components 19
Agent Workflow
- Agent Setup
tinyml-agent-skills/plugins/tinyml-agent-skills/skills/setup/scripts/runner.py - Workflow Agent
tinyml-agent-skills/plugins/tinyml-agent-skills/skills/tinyml-workflow-agent/scripts/runner.py
Modelmaker Flows
- Modelmaker Entry
tinyml-modelmaker/tinyml_modelmaker/run_tinyml_modelmaker.py - Task Runners
tinyml-modelmaker/tinyml_modelmaker/ai_modules - Dataset Analysis
tinyml-agent-skills/plugins/tinyml-agent-skills/skills/tinyml-workflow-agent/scripts/dataset_analysis.py
Tinyverse Runtime
- Dataset Loaders
tinyml-tinyverse/tinyml_tinyverse/common/datasets/timeseries_dataset.py - Feature Pipeline
tinyml-tinyverse/tinyml_tinyverse/common/transforms/basic_transforms.py - Training References
tinyml-tinyverse/tinyml_tinyverse/references/common/train_base.py - On Device Learning
tinyml-tinyverse/tinyml_tinyverse/common/utils/ondevice_training.py
Optimization Deployment
- Architecture Search
tinyml-modeloptimization/torchmodelopt/tinyml_torchmodelopt/nas/train_cnn_search.py - Quantization
tinyml-modeloptimization/torchmodelopt/tinyml_torchmodelopt/quantization/generic/quant_fx.py - MCU Compilation
tinyml-tinyverse/tinyml_tinyverse/references/common/compilation.py - Device Deployment
tinyml-agent-skills/plugins/tinyml-agent-skills/skills/tinyml-workflow-agent/scripts/device_deployment.py
Model Zoo
- Model Registry
tinyml-modelzoo/tinyml_modelzoo/models/classification.py - Model Descriptions
tinyml-modelzoo/tinyml_modelzoo/model_descriptions/classification.py - Device Metrics
tinyml-modelzoo/tinyml_modelzoo/device_info/run_info.py
Other components
- ML User
- TI MCU
- Embedded Model
Connections 23
- ML User → Agent Setup: installs
- ML User → Workflow Agent: requests workflow
- Workflow Agent → Modelmaker Entry: starts modeling
- Workflow Agent → Dataset Analysis: analyzes data
- Workflow Agent → Device Deployment: dispatches deployment
- Modelmaker Entry → Task Runners: dispatches tasks
- Task Runners → Dataset Loaders: loads datasets
- Dataset Analysis → Dataset Loaders: profiles data
- Dataset Loaders → Feature Pipeline: provides samples
- Training References → Model Registry: creates model
- Dataset Loaders → Training References: feeds loaders
- Feature Pipeline → Training References: prepares inputs
- Workflow Agent → Training References: orchestrates training
- Training References → Architecture Search: optionally searches
- Architecture Search → Model Registry: selects architecture
- Model Registry → Model Descriptions: uses specifications
- Modelmaker Entry → Quantization: optionally quantizes
- Quantization → MCU Compilation: passes model
- MCU Compilation → Device Deployment: produces binary
- Device Deployment → TI MCU: deploys firmware
- TI MCU → Embedded Model: runs inference
- Device Metrics → Workflow Agent: informs selection
- On Device Learning → Embedded Model: updates model