13.1. Configuration Reference

Reference for the YAML configuration options of Tiny ML ModelZoo 1.5.0. For a step-by-step guide on writing or adapting a config, see How to Create a ModelZoo Config YAML.

Note

The defaults below are read from the init_params() functions of ModelMaker (package tinyml_modelmaker, file tinyml_modelmaker/ai_modules/<module>/params.py). They are the values used when a key is missing from your YAML. The tables are written for the timeseries module (target_module: timeseries); the audio, radar and vision modules add or drop a few keys, listed at the end of each section. Keys that are filled in at run time (paths, model ids) are not meant to be set by the user and are listed as “internal”.

13.1.1. How a YAML is merged

The user YAML is loaded with yaml.safe_load and merged recursively on top of the defaults (ConfigDict in tinyml_modelmaker/utils/config_dict.py), so you only need to write the keys you want to change. Then, in the ModelMaker runner:

• the training defaults of the chosen model are taken from its model description (model_name must be in the enabled_models_list of the model descriptions; see Model Zoo Reference),

• a feature_extraction_name preset, if given, fills the data_processing_feature_extraction keys it defines,

• a compilation preset is chosen from compile_preset_name and target_device.

If CUDA is available, training.compile_model and training.native_amp are switched on automatically unless you set them in the YAML or set training.num_gpus: 0 (utils/hardware_defaults.py).

13.1.2. File structure

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

There is also a download key in the defaults (default None). It is not documented in the examples and is not used in the shipped configs.

13.1.3. Common section

Key	Default	Description
target_module	'timeseries'	timeseries, audio, radar or image (vision module).
task_type	None (required)	See the task type list below.
task_category	None	Filled in from task_type.
target_device	None (required)	Target MCU, for example F28P55. See the device list below.
run_name	{date-time}/{model_name}	Run folder pattern under projects_path/<project>/run. {date-time} becomes %Y%m%d-%H%M%S and {model_name} the model name.
download_path	./data/downloads	Where datasets are downloaded.
projects_path	./data/projects	Parent folder of project outputs.
project_path, project_run_path	None	Internal, derived from projects_path and run_name.
generic_model	False	Set from the model description.

Task types (common.task_type):

• Timeseries: generic_timeseries_classification, generic_timeseries_regression, generic_timeseries_forecasting, generic_timeseries_anomalydetection

• Timeseries application aliases mapped to classification: arc_fault, motor_fault, ecg_classification, blower_imbalance, pir_detection

• Other modules: audio_classification, radar_classification, image_classification

The byom_compilation task type and the BYOM section of the 1.4.0 reference are not present in the 1.5.0 code (no match in tinyml_modelmaker). To use your own model see Bring Your Own Model (BYOM).

Target devices (tinyml_modelmaker/ai_modules/timeseries/constants.py, TARGET_DEVICES; the exact devices valid for a task depend on the model, see Model Zoo Reference):

• C2000 F28x: F280013, F280015, F28003, F28004, F2837, F28P55, F28P65, F28E12, F2807x, F28002x, F28P551x, F2837xS, F2838x

• C2000 F29x: F29H85, F29P58, F29P32

• MSPM0 and MSPM33: MSPM0G3507, MSPM0G3519, MSPM0G5187, MSPM33C32

• AM13: AM13E2

• AM26x: AM263, AM263P, AM261

• Connectivity: CC2755, CC2745, CC1352, CC1312, CC1354, CC1314, CC35X1

13.1.4. Dataset section

Key	Default	Description
enable	True	Enable dataset handling.
dataset_name	None	Name of a built-in dataset, or your own name when input_data_path is set.
input_data_path	None	Path or URL of the dataset (folder, zip or URL). Used for custom data.
input_annotation_path	None	Annotation file, if the format needs one.
dataset_path, extract_path	None	Internal; where the split dataset is created.
split_factor	(0.6, 0.3, 0.1)	Train/val/test ratios.
split_names	('train', 'val', 'test')	Names of the splits.
split_type	'amongst_files'	amongst_files (whole files go to one split) or within_files (each file is split). Other values are not handled in ai_modules/common/datasets/__init__.py.
max_num_files	10000	Maximum number of files used.
data_dir	None	None auto-detects the folder based on the task category.
annotation_format	'univ_ts_json'	univ_ts_json or coco_json (any other value raises Unsupported annotation_format).
annotation_dir	'annotations'	Annotation folder name.
annotation_prefix	'instances'	Annotation file prefix.
input_data_path splits: data_path_splits, annotation_path_splits	None	Internal.

13.1.5. Data processing and feature extraction section

The preset in feature_extraction_name (see Feature Extraction) sets most of these keys for you. Values you write in the YAML override the preset.

Key	Default	Description
feature_extraction_name	None	Preset name, for example Generic_1024Input_FFTBIN_64Feature_8Frame, or Custom_<name> for your own transform chain.
data_proc_transforms	[]	Data processing transforms: SimpleWindow, Downsample (combined as ['Downsample', 'SimpleWindow']).
feat_ext_transform	[]	Feature extraction chain. Names used by the 1.5.0 presets: FFT_FE, FFT_POS_HALF, WINDOWING, BINNING, NORMALIZE, ABS, LOG_DB, CONCAT, DC_REMOVE, RAW_FE, FFT_Q15, Q15_MAG, Q15_SCALE, BIN_Q15, FE_Q15, FE_DECIMATE, FFT_COL, L2_NORM_ROW, PIR_FE, PIR_FE_Q15, ROUND_OFF.
variables	1	Input columns after the time column(s): an int (first n columns), a list of ints, or a list of column names.
target_variables	[]	Forecasting: columns to predict.
sampling_rate	1	Original sampling rate.
new_sr	1	Target sampling rate (used with Downsample).
resampling_factor	1	Resampling factor.
frame_size	1	Samples per frame.
stride_size	0.01	Stride between frames as a fraction of frame_size.
forecast_horizon	1	Forecasting: number of future time steps to predict.
feature_size_per_frame	None	Output features per frame.
num_frame_concat	1	Frames concatenated into one input.
frame_skip	1	Frames skipped between selected frames.
min_bin	1	First FFT bin kept.
analysis_bandwidth	1	Bandwidth factor for binning.
normalize_bin	0	Bin normalization.
dc_remove	True	Remove the DC component.
q15_scale_factor	4	Q15 scaling for the fixed-point chain.
log_base, log_mul, log_threshold	None	Parameters of LOG_DB; the presets set log_mul=10, log_base='e', log_threshold=1e-12.
stacking	'2D1'	Feature stacking mode (2D1 or 1D).
offset, scale	0, None	Offset and scale applied to the input.
gain_variations	None	Dict of class-to-gain range for augmentation.
fft_size, window_count, chunk_size	1	Used by some presets.
store_feat_ext_data, feat_ext_store_dir	False, None	Store extracted features to disk.
dont_train_just_feat_ext	False	Only run feature extraction.
nn_for_feature_extraction	False	Run feature extraction in the NN (for devices without the FE library).
gof_test	False	Goodness-of-fit test.

Audio module (target_module: audio) replaces the timeseries keys with: sampling_rate (16000), audio_duration_ms (1000), audio_feature ("MFCC"), n_mfcc (10), n_mels (40), frame_length_ms (30), frame_step_ms (20), nlpc (70), lpc_order (10), the learned filterbank keys fb_conv_kernel (64), fb_output_channel (64), fb_bitwidth (8), fb_conv_stride (4), fb_context_ms (20), input_bit_depth (16), frame_size (512), normalize_audio (True), mono (True) and augmentation_transform ([]). The allowed values of audio_feature were not listed here; see ai_modules/audio/constants.py.

Radar module (target_module: radar) has the timeseries keys sampling_rate (1), frame_size (8), variables (1), feat_ext_transform and data_proc_transforms, and no feature extraction presets.

Vision module (target_module: image): image_height (28), image_width (28), image_num_channel (1), image_mean (0.1307), image_scale (0.3081), feat_ext_transform, data_proc_transforms, augmentation_transform and the transform parameters pad_value (0), binary_threshold (128), clahe_clip_limit (2.0), clahe_tile_grid_size ((8, 8)), sobel_mode ("magnitude"), sobel_ksize (3), laplacian_ksize (3), horizontal_flip_prob (0.5), vertical_flip_prob (0.5), random_rotation_deg (15) and color_jitter_brightness (0.10), _contrast (0.10), _saturation (0.05), _hue (0.01).

13.1.6. Training section

Training defaults are overridden per model by the model description (for example learning_rate and batch_size), then by your YAML.

Key	Default	Description
enable	True	Run the training stage.
model_name	None (required)	Must be listed in the enabled_models_list of the model descriptions. See Model Zoo Reference.
model_spec, model_config	None	Model specification or config (for spec-based models).
dataset_loader	'GenericTSDataset'	GenericAudioDataset, GenericImageDataset and GenericRadarDataset for the other modules.
augment_config	None	Augmentation config. Usage not verified in this reference.
pretrained_checkpoint_path	None	Start from this checkpoint.
load_saved_model	None	Load a saved model. Usage not verified in this reference.
training_epochs	10	Number of epochs.
batch_size	8	Batch size (model description default overrides).
learning_rate	2e-3	Learning rate.
optimizer	'sgd'	sgd (sgd_nesterov style names starting with sgd are accepted), rmsprop, adam, adamw. Unknown values log a warning and fall back to Adam.
weight_decay	1e-4	Weight decay.
lr_scheduler	'cosineannealinglr'	cosineannealinglr, steplr, exponentiallr, none. Other values raise Invalid lr scheduler.
lambda_reg	0	Regularization weight.
early_stopping	True	New in 1.5.0. Stop when the validation metric stops improving.
early_stopping_patience	5	New in 1.5.0. Epochs without improvement before stopping.
early_stopping_min_delta	0.0	New in 1.5.0. Minimum change that counts as improvement.
training_device	'cuda'	cpu, cuda or mps. Falls back if unavailable.
num_gpus	1	0 or 1. 0 also disables the automatic compile_model/native_amp.
distributed	True	Use distributed training when launched for it.
compile_model	0	1 enables torch.compile. Turned on automatically when CUDA is available and the key is not set.
native_amp	False	PyTorch native AMP. Turned on automatically when CUDA is available and the key is not set.
quantization	0 (NO_QUANTIZATION)	0 none, 1 generic, 2 TINPU (NPU style). Many model descriptions default to 2.
quantization_method	'QAT'	QAT or PTQ.
quantization_weight_bitwidth	8	Choices in the training script: 16, 8, 4, 2.
quantization_activation_bitwidth	8	Activation bitwidth.
auto_quantization	True	New in 1.5.0. Automatic bitwidth selection (binary search, Hessian-based mixed precision). Set False to use the fixed bitwidths above.
autoquant_tolerance_classification	0.05	Tolerated accuracy drop (fraction) versus the float model.
autoquant_tolerance_regression	0.05	Tolerated R2 drop (fraction).
autoquant_tolerance_forecasting	2.0	Tolerated SMAPE increase: float SMAPE x (1 + 2.0).
autoquant_tolerance_anomaly	2.0	Tolerated MSE increase: float MSE x (1 + 2.0).
output_int	None	None decides automatically from the task category and quantization (integer output for classification when quantized, float output for regression, forecasting and anomaly detection). True/False forces it and overrides the compile target string.
dual_op	False	Passed to the radar training script as --dual-op; unused by the other timeseries tasks in this reference.
run_quant_train_only	False	Run only the quantization training step.
ondevice_training	False	Mark the model as trainable on device. See On-Device Training Documentation.
trainable_layers_from_last	1	Number of trainable layers from the end for on-device training.
target_device_flash_kb	100	Flash size of the target in KB.
export_samples_per_class	(10, 10, 10)	Samples exported per class for (train, val, test), used for the golden vectors.
nas_enabled	False	Neural architecture search.
nas_optimization_mode	'Memory'	Memory or Compute.
nas_epochs	10	Search epochs.
nas_model_size	None	Preset mode: s, m, l, xl.
nas_nodes_per_layer, nas_layers, nas_init_channels, nas_init_channel_multiplier, nas_fanout_concat	4, 3, 1, 3, 4	Customization mode. nas_layers minimum is 3; nas_fanout_concat must not exceed nas_nodes_per_layer.
model_training_id, training_backend, target_devices, properties	various	Filled from the model description (internal / GUI metadata).
project_path, dataset_path, training_path, training_path_quantization, train_output_path, log_file_path, file_level_classification_log_path, log_summary_regex	None	Internal.

The radar module also has momentum (default 0).

Memory estimation (pre-flight) is not a YAML key. The training script has a --skip_memory_preflight flag, and the standalone estimate_memory command takes a config file; see Memory Estimation.

13.1.7. Testing section

Key	Default	Description
enable	True	Run the test stage.
skip_train	False	Skip training and test an existing model (set model_path).
model_path	None	Model to test when training is skipped.
test_data	None	Test data location.
device_inference	False	Run inference on the device. Usage not verified in this reference.

13.1.8. Compilation section

Key	Default	Description
enable	True	Run the compilation stage (needs the compiler tools, see Common Errors).
compile_preset_name	None, then default_preset	default_preset, compress_npu_layer_data (NPU, optimizes layer data for space) or forced_soft_npu_preset (use the software NPU on an NPU device such as F28P55).
keep_libc_files	False	Keep intermediate library files.
model_path	None	Model to compile (normally set by the pipeline).
compile_output_path	None	Internal.

See Supported Devices and NPU Configuration Guidelines for TI Neural Network Compiler for which preset suits which device.

13.1.9. Example

This is examples/dc_arc_fault/config_dsk.yaml without comments:

common:
  task_type: arc_fault
  target_device: F28P55
dataset:
  dataset_name: dc_arc_fault_example_dsk
data_processing_feature_extraction:
  feature_extraction_name: FFT1024Input_256Feature_1Frame_Full_Bandwidth
  variables: 1
training:
  model_name: CLS_1k_NPU
  model_config: ''
  batch_size: 32
  training_epochs: 10
  num_gpus: 0
  early_stopping: true
  early_stopping_patience: 5
testing: {}
compilation: {}