3.5. How to Create a ModelZoo Config YAML

Every example under examples/ is driven by a single YAML file. This guide explains the structure of that file and how to adapt an existing example to your own dataset and target device.

1. Config Skeleton and 4-Step Recipe
2. Feature Extraction: Preset vs. Custom
3. Useful Knobs
Common Pitfalls

3.5.1. 1. Config Skeleton and 4-Step Recipe

Almost every config follows the same 6-section skeleton (based on generic_timeseries_classification/config.yaml):

common:                          # WHAT task, WHICH device
  task_type: generic_timeseries_classification
  target_device: F28P55          # F28P55, MSPM0G5187, F29H85, CC2755, AM263 ...

dataset:                         # WHERE the data is
  dataset_name: my_dataset
  input_data_path: <URL | .zip | local folder>

data_processing_feature_extraction:   # HOW raw signal becomes model input
  feature_extraction_name: Generic_1024Input_FFTBIN_64Feature_8Frame  # preset
  variables: 1                   # no. of input channels

training:                        # WHICH model, HOW to train
  model_name: CLS_1k_NPU         # from ModelZoo
  batch_size: 256
  training_epochs: 20
  quantization: 2                # 0 = float, 1 = generic INT, 2 = TI-NPU INT

testing: {}                      # {} = use defaults
compilation: {}                  # {} = default_preset

Section	Purpose
`common`	Task type, target device, (optionally) output location and run name
`dataset`	Dataset name and where to load it from
`data_processing_feature_extraction`	Windowing, transforms and feature extraction applied to raw data
`training`	Model architecture, training hyperparameters and quantization
`testing`	Evaluation of the trained model
`compilation`	Compilation into device-ready code with TI’s Neural Network Compiler

3.5.1.1. 4-Step Recipe

Copy the closest example. Use the use-case folder if one matches your problem (e.g. motor_bearing_fault). If none does, use the generic example for your task: generic_timeseries_classification, generic_timeseries_regression, generic_timeseries_forecasting or generic_timeseries_anomalydetection.
Set common. Pick task_type and target_device. Many examples already ship device-specific variants such as config_MSPM0.yaml, config_CC2755.yaml or config_CC1352.yaml. Start from those when available.
Point dataset at your data. Set input_data_path (URL, .zip or local folder) and set variables to match the number of input channels in your data.

Pick feature extraction and a model, then run:

./run_tinyml_modelzoo.sh examples/<name>/config.yaml       # Linux
run_tinyml_modelzoo.bat examples\<name>\config.yaml        # Windows

Tip: Any key you leave out uses its default value, so testing: {} and compilation: {} are valid as written.

3.5.2. 2. Feature Extraction: Preset vs. Custom

Approach	When	Example
Preset	Classification, audio, image	`feature_extraction_name: Generic_512Input_FFTBIN_32Feature_8Frame`
Custom transforms	Fine-tuning a preset	`frame_size`, `feature_size_per_frame`, `num_frame_concat`, `feat_ext_transform: [FFT_FE, FFT_POS_HALF, DC_REMOVE, ABS, BINNING, LOG_DB, CONCAT]`
Windowing only	Regression, forecasting, anomaly detection	`data_proc_transforms: [SimpleWindow, Downsample]`, `frame_size`, `stride_size`, `sampling_rate`, `new_sr`

3.5.2.1. Preset example (classification)

data_processing_feature_extraction:
  feature_extraction_name: Generic_1024Input_FFTBIN_64Feature_8Frame
  variables: 1

Other generic presets available for timeseries classification:

Preset	Transform
`Generic_1024Input_FFTBIN_64Feature_8Frame`	FFT, binned, 8 frames concatenated
`Generic_512Input_FFTBIN_32Feature_8Frame`	FFT, binned, 8 frames concatenated
`Generic_256Input_FFTBIN_16Feature_8Frame`	FFT, binned, 8 frames concatenated
`Generic_1024Input_FFT_512Feature_1Frame`	FFT, single frame
`Generic_512Input_FFT_256Feature_1Frame`	FFT, single frame
`Generic_256Input_FFT_128Feature_1Frame`	FFT, single frame
`Generic_512Input_RAW_512Feature_1Frame`	Raw samples, no transform
`Generic_256Input_RAW_256Feature_1Frame`	Raw samples, no transform
`Generic_128Input_RAW_128Feature_1Frame`	Raw samples, no transform

3.5.2.2. Custom transforms example

data_processing_feature_extraction:
  frame_size: 1024
  feature_size_per_frame: 64
  num_frame_concat: 4
  feat_ext_transform: [ 'FFT_FE', 'FFT_POS_HALF', 'DC_REMOVE', 'ABS', 'BINNING', 'LOG_DB', 'CONCAT' ]
  variables: 1

3.5.2.3. Windowing-only example (anomaly detection)

data_processing_feature_extraction:
  data_proc_transforms:
  - SimpleWindow
  - Downsample
  frame_size: 128
  sampling_rate: 100
  new_sr: 10
  variables: 1

3.5.2.4. How the task type changes the config

Task	What is different
Forecasting	`SimpleWindow` is mandatory; add `forecast_horizon` (future timesteps to predict) and `target_variables` (column indices to predict). See `generic_timeseries_forecasting`.
Anomaly detection	Uses `dataset.split_factor: [0.5, 0.1, 0.4]` and an `AD_*` autoencoder model. See `generic_timeseries_anomalydetection`.
Regression	Usually windowing only (`SimpleWindow`) with a `REGR_*` model. See `generic_timeseries_regression`.
Image	Set `common.target_module: vision` and use a preset such as `Mnist_Default`. See `MNIST_image_classification`.

3.5.2.5. Selecting input channels with `variables`

variables supports three formats (column indices are 0-indexed, counted after time columns are removed):

variables: 3                           # first N columns
variables: [0, 2, 4]                   # specific column indices
variables: ['accel_x', 'accel_y']      # column names (requires CSV header)

3.5.3. 3. Useful Knobs

3.5.3.1. Training

Key	Description
`early_stopping`, `early_stopping_patience`	Stop training when validation stops improving
`optimizer`	e.g. `adam`
`lr_scheduler`	e.g. `cosineannealinglr`
`learning_rate`	Initial learning rate
`num_gpus`	`0` for CPU training

3.5.3.2. Quantization

Key	Description
`quantization`	`0` = float, `1` = generic integer, `2` = TI-NPU integer
`quantization_method`	`QAT` (quantization-aware training) or `PTQ` (post-training quantization)
`quantization_weight_bitwidth`	`8`, `4` or `2`
`quantization_activation_bitwidth`	`8`
`auto_quantization`	Automatically select per-layer precision

3.5.3.3. Compilation

Key	Description
`compile_preset_name`	`default_preset`, `compress_npu_layer_data` or `forced_soft_npu_preset`. The last two apply only to NPU devices.

See NPU Configuration Guidelines for designing models that run efficiently on the TI NPU.

3.5.3.4. Output location

Key	Description
`common.projects_path`	Where artifacts are written. Default: `./data/projects/<project_name>/`, relative to the directory you run from
`common.run_name`	Run sub-folder name, e.g. `'{date-time}/{model_name}'`

3.5.4. Common Pitfalls

Be consistent with indentation. The examples use either 2-space indents with bare values or 4-space indents with quoted values. Both are valid YAML, but every key within one file must use the same indentation.
Match variables to your data. A mismatch between variables and the number of channels in your CSV files is a common source of errors.
NPU-only compile presets. compress_npu_layer_data and forced_soft_npu_preset apply only to devices with the TinyEngine NPU (e.g. F28P55, MSPM0G5187, AM13E2).