How to Create a ModelZoo Config YAML

ModelZoo · Config YAML

Every example in examples/ follows the same 6-section skeleton

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

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

data_processing_feature_extraction:  # HOW signal becomes input
  feature_extraction_name: Generic_1024Input_FFTBIN_64Feature_8Frame
  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

4-Step Recipe

1 · Copy the closest example

Use the use-case folder if one matches (e.g. motor_bearing_fault). Otherwise use generic_timeseries_<task>.

2 · Set common

Pick task_type and target_device. Device variants exist: config_MSPM0.yaml, config_CC2755.yaml, ...

3 · Point dataset at your data

Set input_data_path (URL, .zip or local folder) and match variables to your input channels.

4 · Pick features + model, then run

./run_tinyml_modelzoo.sh examples/<name>/config.yaml

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

ModelZoo · Config YAML

Feature Extraction: Preset vs. Custom

The data_processing_feature_extraction section decides how raw signals become model input

Approach	When	Example keys
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`

How the task type changes the config

Forecasting: SimpleWindow is mandatory; add forecast_horizon and target_variables.

Anomaly detection: split_factor: [0.5, 0.1, 0.4] and an AD_* model.

Regression: usually SimpleWindow only, with a REGR_* model.

Image: target_module: vision with a preset such as Mnist_Default.

Selecting input channels with variables

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

Column indices are 0-indexed, counted after time columns are removed.

ModelZoo · Config YAML

Useful Knobs

Optional keys you will reach for most often. Omit them to keep the defaults

training

early_stopping, early_stopping_patience: stop when validation stops improving

optimizer (e.g. adam), lr_scheduler (e.g. cosineannealinglr)

learning_rate: initial learning rate

num_gpus: 0 for CPU training

training · quantization

quantization: 0 float · 1 generic integer · 2 TI-NPU integer

quantization_method: QAT (quantization-aware training) or PTQ (post-training)

quantization_weight_bitwidth: 8 | 4 | 2 · quantization_activation_bitwidth: 8

auto_quantization: automatically select per-layer precision

compilation

compile_preset_name:

any device default_preset

NPU only compress_npu_layer_data, forced_soft_npu_preset

See docs/NPU_CONFIGURATION_GUIDELINES.md for NPU-friendly models

common · output

projects_path: where artifacts are written

Default: ./data/projects/<project_name>/, relative to the directory you run from

run_name: run sub-folder, e.g. '{date-time}/{model_name}'

ModelZoo · Config YAML

Common Pitfalls

Three things to check before you run a modified config

Inconsistent indentation

What: examples use either 2-space indents with bare values or 4-space indents with quoted values.

Why it matters: both are valid YAML, but keys that are not aligned within one file break parsing.

Fix: pick one style per file and keep every key under a section at the same indent.

variables mismatch

What: variables does not match the number of channels in your CSV files.

Why it matters: a common source of errors when moving from an example dataset to your own.

Fix: count channels after time columns are removed, or list indices / column names explicitly.

NPU-only compile presets

What: compress_npu_layer_data and forced_soft_npu_preset are set for a device without an NPU.

Why it matters: these presets apply only to TinyEngine NPU devices (e.g. F28P55, MSPM0G5187, AM13E2).

Fix: use default_preset, or leave compilation: {}.

Full reference: tinyml-modelzoo/docs/CONFIG_YAML_GUIDE.md

Adithya Thonse · October 2026