3.4. Understanding Config

Every run is driven by one YAML file with up to six sections. Keys you leave out use their defaults, so testing: {} and compilation: {} are valid as written.

Section	Purpose
`common`	`task_type`, `target_device`; optionally `projects_path`, `run_name`
`dataset`	`dataset_name` and where the data comes from (`input_data_path`: URL, `.zip` or folder)
`data_processing_feature_extraction`	Windowing and feature extraction: a preset (`feature_extraction_name`) or custom transforms
`training`	`model_name`, `batch_size`, `training_epochs`, early stopping, quantization, `num_gpus`
`testing`	Evaluation of the trained model
`compilation`	Compilation with TI’s Neural Network Compiler (`compile_preset_name`)

3.4.1. Minimal Config

common:
  task_type: generic_timeseries_classification
  target_device: F28P55
dataset:
  dataset_name: generic_timeseries_classification
data_processing_feature_extraction:
  feature_extraction_name: Generic_1024Input_FFTBIN_64Feature_8Frame
  variables: 1
training:
  model_name: CLS_1k_NPU
  training_epochs: 20
testing: {}
compilation: {}

3.4.2. Common Keys

common.target_device: the MCU to compile for, e.g. F28P55, MSPM0G5187, F29H85, CC2755, AM263
common.projects_path: where data/projects is created (default ./data/projects)
training.early_stopping and training.early_stopping_patience: stop when validation stops improving
training.num_gpus: 0 for CPU, 1 or more for CUDA (or MPS on Apple Silicon)
<stage>.enable: set False to skip a stage (for example compilation: {enable: False})

The full walkthrough (4-step recipe, feature extraction presets, custom transforms, useful knobs and pitfalls) is in the How to Create a ModelZoo Config YAML. There is also an interactive slide deck.

3.4.3. Next Steps