3.3. Understanding Config

This guide explains the YAML configuration file structure used by Tiny ML Tensorlab.

3.3.1. Configuration Overview

Every training run is controlled by a YAML configuration file with these sections:

common:        # General settings
dataset:       # Data source and splitting
data_processing_feature_extraction:  # Preprocessing
training:      # Model and training parameters
testing:       # Test evaluation settings
compilation:   # Device compilation settings

3.3.2. Common Section

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

Parameter	Description
`target_module`	`'timeseries'` or `'image'`
`task_type`	Task type (see valid values below)
`target_device`	Target MCU (e.g., `F28P55`, `MSPM0G3507`)
`run_name`	Output directory name. Supports placeholders: `{date-time}`, `{model_name}`

Valid task_type Values:

`generic_timeseries_classification`	Time series classification
`generic_timeseries_regression`	Time series regression
`generic_timeseries_forecasting`	Time series forecasting
`generic_timeseries_anomalydetection`	Anomaly detection (autoencoder)
`image_classification`	Image classification

3.3.3. Dataset Section

dataset:
  enable: True
  dataset_name: 'my_dataset'
  input_data_path: '/path/to/data'  # or URL
  split_type: 'amongst_files'
  split_factor: [0.6, 0.3, 0.1]

Parameter	Description
`enable`	Set `False` to skip dataset loading (for BYOM compilation)
`dataset_name`	Name used for output directory
`input_data_path`	Local path, zip file, or URL
`split_type`	`'amongst_files'` (files split) or `'within_files'` (rows split)
`split_factor`	[train, val, test] ratios (must sum to 1.0)

3.3.4. Data Processing & Feature Extraction Section

data_processing_feature_extraction:
  # Data processing
  data_proc_transforms: ['DownSample', 'SimpleWindow']
  sampling_rate: 100000
  new_sr: 10000
  frame_size: 1024
  stride_size: 0.5

  # Feature extraction
  feature_extraction_name: 'Generic_1024Input_FFTBIN_64Feature_8Frame'
  variables: 1

Data Processing Parameters:

Parameter	Description
`data_proc_transforms`	List of transforms: `'DownSample'`, `'SimpleWindow'`, `'GainVariation'`
`sampling_rate`	Original sample rate (for DownSample)
`new_sr`	Target sample rate after downsampling
`frame_size`	Window size in samples
`stride_size`	Overlap fraction (0.5 = 50% overlap)

Feature Extraction Parameters:

Parameter	Description
`feature_extraction_name`	Preset name (see Feature Extraction)
`variables`	Number of input channels
`feat_ext_transform`	Custom transform list if not using preset

3.3.5. Training Section

training:
  enable: True
  model_name: 'CLS_1k_NPU'
  batch_size: 256
  training_epochs: 20
  num_gpus: 0
  learning_rate: 0.001

Parameter	Description
`enable`	Set `False` to skip training
`model_name`	Model architecture name from ModelZoo
`batch_size`	Samples per training batch
`training_epochs`	Number of training passes
`num_gpus`	`0` for CPU, `1` for single GPU
`learning_rate`	Initial learning rate (default varies by task)

NAS Parameters (Optional):

training:
  nas_enabled: True
  nas_epochs: 10
  nas_optimization_mode: 'Memory'  # or 'Compute'
  nas_model_size: 'm'  # s, m, l, xl, xxl

3.3.6. Testing Section

testing:
  enable: True

Usually left empty ({}) to use defaults. Testing automatically runs on the test set after training.

3.3.7. Compilation Section

compilation:
  enable: True
  preset_name: 'default_preset'  # or 'compress_npu_layer_data'

Parameter	Description
`enable`	Set `False` to skip compilation
`preset_name`	Compilation options preset
`model_path`	Path to external ONNX model (for BYOM)

3.3.8. 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: 'my_vibration_data'
  input_data_path: '/data/vibration_dataset'
  split_factor: [0.7, 0.2, 0.1]

data_processing_feature_extraction:
  data_proc_transforms: ['SimpleWindow']
  frame_size: 512
  stride_size: 0.25
  feature_extraction_name: 'Generic_512Input_FFTBIN_32Feature_8Frame'
  variables: 3

training:
  enable: True
  model_name: 'CLS_4k_NPU'
  batch_size: 128
  training_epochs: 50
  num_gpus: 1

testing:
  enable: True

compilation:
  enable: True

3.3.9. Tips

Start with a preset feature_extraction_name before customizing
Use smaller batch_size if you run out of memory
Set num_gpus: 0 to force CPU training
Use enable: False to skip stages for debugging

3.3.10. See Also

Feature Extraction - Feature extraction presets
Configuration Reference - Complete parameter reference