13.1. Configuration Reference

Complete reference for all YAML configuration options in Tiny ML Tensorlab.

13.1.1. Configuration File Structure

common:
  # General settings
dataset:
  # Dataset configuration
data_processing_feature_extraction:
  # Feature extraction settings
training:
  # Training parameters
testing:
  # Testing configuration
compilation:
  # Compilation settings

13.1.2. Common Section

General project settings.

common:
  target_module: 'timeseries'    # 'timeseries' or 'image'
  task_type: 'generic_timeseries_classification'
  target_device: 'F28P55'
  run_name: '{date-time}/{model_name}'

Option	Required	Description
target_module	Yes	Module type: timeseries or image
task_type	Yes	Task type (see below)
target_device	Yes	Target MCU device name
run_name	No	Output directory name pattern

Task Types:

• generic_timeseries_classification

• generic_timeseries_regression

• generic_timeseries_forecasting

• generic_timeseries_anomalydetection

• generic_image_classification

• byom_compilation

Target Devices:

• C2000: F28P55, F28P65, F29H85, F29P58, F29P32, F2837, F28004, F28003, F280013, F280015

• MSPM0: MSPM0G3507, MSPM0G3519, MSPM0G5187

• MSPM33C: MSPM33C32, MSPM33C34, AM13E2

• AM26x: AM263, AM263P, AM261

• Connectivity: CC2755, CC1352

13.1.3. Dataset Section

dataset:
  enable: True
  dataset_name: 'my_dataset'
  input_data_path: '/path/to/dataset'
  data_split_type: 'random'
  data_split_ratio: [0.8, 0.1, 0.1]

Option	Default	Description
enable	True	Enable dataset processing
dataset_name	Required	Dataset identifier
input_data_path	None	Path to custom dataset
data_split_type	‘random’	Split method: ‘random’, ‘sequential’, ‘predefined’
data_split_ratio	[0.8, 0.1, 0.1]	Train/val/test split ratios
input_data_split_type	‘amongst_files’	‘amongst_files’ or ‘within_files’

13.1.4. Feature Extraction Section

data_processing_feature_extraction:
  feature_extraction_name: 'Generic_1024Input_FFTBIN_64Feature_8Frame'
  variables: 1
  gof_test: False

Option	Default	Description
feature_extraction_name	None	Preset name (e.g., 'Generic_1024Input_FFTBIN_64Feature_8Frame') or custom name starting with 'Custom_'
data_proc_transforms	None	List of data processing transforms: ['SimpleWindow'], ['Downsample'], ['SimpleWindow', 'Downsample'], or []
feat_ext_transform	None	List of feature extraction transforms (e.g., ['FFT_FE', 'FFT_POS_HALF', 'ABS', 'BINNING', 'LOG_DB', 'CONCAT'])
variables	1	Number of input channels, or list of column indices/names
frame_size	None	Samples per frame
feature_size_per_frame	None	Output features per frame after transform
num_frame_concat	None	Number of frames to concatenate
stride_size	None	Stride between frames as a fraction
sampling_rate	None	Original sampling rate (used with Downsample)
new_sr	None	Target sampling rate (used with Downsample)
scale	None	Scaling factor applied to input data
offset	None	Offset added to input data
frame_skip	None	Frames to skip between selected frames
normalize_bin	None	Enable bin normalization
stacking	None	Feature stacking mode: '2D1' or '1D'
gof_test	False	Run Goodness of Fit test
gain_variations	None	Dict of class-to-gain-range for data augmentation
store_feat_ext_data	False	Store extracted feature data to disk
nn_for_feature_extraction	False	Use neural network for feature extraction

Forecasting-Specific:

data_processing_feature_extraction:
  data_proc_transforms:
  - SimpleWindow
  frame_size: 32
  stride_size: 0.1
  forecast_horizon: 2
  variables: 1
  target_variables:
  - 0

13.1.5. Training Section

training:
  enable: True
  model_name: 'CLS_4k_NPU'
  training_epochs: 30
  batch_size: 256
  learning_rate: 0.001
  num_gpus: 0

Option	Default	Description
enable	True	Enable training
model_name	Required	Model name from registry
training_epochs	20	Number of training epochs
batch_size	256	Batch size for training
learning_rate	0.001	Initial learning rate
optimizer	‘adam’	‘adam’, ‘sgd’, ‘adamw’
weight_decay	0.0001	Weight decay (L2 regularization)
num_gpus	0	Number of GPUs (0 for CPU)
num_workers	4	Data loader workers
seed	42	Random seed

Quantization Options:

training:
  quantization: 2
  quantization_method: 'QAT'
  quantization_weight_bitwidth: 8
  quantization_activation_bitwidth: 8

Option	Default	Description
quantization	0	Quantization mode: 0 = floating point training, 1 = standard PyTorch Quantization, 2 = TI style optimised Quantization
quantization_method	None	'PTQ' or 'QAT'. Only applicable when quantization is 1 or 2
quantization_weight_bitwidth	None	Weight bit width: 8, 4, or 2. Only applicable when quantization is 1 or 2
quantization_activation_bitwidth	None	Activation bit width: 8, 4, or 2. Only applicable when quantization is 1 or 2

Learning Rate Scheduler:

training:
  lr_scheduler: 'cosine'
  lr_warmup_epochs: 5

Option	Default	Description
lr_scheduler	None	‘cosine’, ‘step’, ‘exponential’
lr_warmup_epochs	0	Warmup epochs
lr_step_size	10	Epochs per step (step scheduler)
lr_gamma	0.1	LR decay factor

13.1.6. Testing Section

testing:
  enable: True
  test_float: True
  test_quantized: True

Option	Default	Description
enable	True	Enable testing
test_float	True	Test float32 model
test_quantized	True	Test quantized model
save_predictions	False	Save prediction results
error_analysis	False	Save misclassified samples

13.1.7. NAS Section

nas:
  enable: True
  search_type: 'multi_trial'
  num_trials: 20
  param_range: [500, 5000]
  accuracy_target: 0.95

Option	Default	Description
enable	False	Enable NAS
search_type	‘multi_trial’	‘single_trial’, ‘multi_trial’
num_trials	20	Architectures to evaluate
param_range	[500, 5000]	[min, max] parameters
accuracy_target	0.9	Minimum accuracy target
npu_compatible	True	Enforce NPU constraints

13.1.8. Compilation Section

compilation:
  enable: True
  preset_name: 'compress_npu_layer_data'

Option	Default	Description
enable	True	Enable compilation
preset_name	‘default_preset’	Compilation preset
optimize_memory	True	Memory optimization
debug_info	False	Include debug symbols

Compilation Presets:

• default_preset - Standard compilation

• compress_npu_layer_data - NPU-optimized

13.1.9. BYOM Section

For compilation-only mode:

byom:
  enable: True
  onnx_model_path: '/path/to/model.onnx'
  input_shape: [1, 1, 512, 1]
  already_quantized: False

Option	Default	Description
enable	False	Enable BYOM mode
onnx_model_path	Required	Path to ONNX model
input_shape	Required	Model input shape
already_quantized	False	True if model is quantized

13.1.10. Complete Example

common:
  target_module: 'timeseries'
  task_type: 'generic_timeseries_classification'
  target_device: 'F28P55'
  run_name: '{date-time}/{model_name}'

dataset:
  enable: True
  dataset_name: 'dc_arc_fault_example_dsk'
  input_data_path: 'https://software-dl.ti.com/...'
  data_split_type: 'random'
  data_split_ratio: [0.8, 0.1, 0.1]

data_processing_feature_extraction:
  enable: True
  feature_extraction_name: 'FFT1024Input_256Feature_1Frame_Full_Bandwidth'
  variables: 1
  gof_test: False

training:
  enable: True
  model_name: 'ArcFault_model_400_t'
  training_epochs: 30
  batch_size: 256
  learning_rate: 0.001
  num_gpus: 0
  quantization: 2
  quantization_method: 'QAT'
  quantization_weight_bitwidth: 8
  quantization_activation_bitwidth: 8

testing:
  enable: True
  test_float: True
  test_quantized: True

compilation:
  enable: True
  preset_name: 'compress_npu_layer_data'