7.9.1. Generic Time Series Classification

This example is the “Hello World” of Tiny ML ModelZoo time series classification. It classifies synthetic sine, square and sawtooth waveforms and shows the complete flow (data, feature extraction, training, quantization, compilation) with the smallest possible configuration.

Example folder: examples/generic_timeseries_classification (GitHub)

7.9.1.1. Overview

• Task: generic_timeseries_classification (3 classes)

• Dataset: synthetic waveforms (sine, square, sawtooth), 1 variable

• Model: CLS_1k_NPU (about 1,000 parameters)

• Default device: F28P55 (NPU)

• Training: 20 epochs, batch size 256, early stopping enabled (patience 5)

7.9.1.2. Available Configurations

One configuration file is provided per device family.

Config file	Target device	Model	Feature extraction preset
config.yaml	F28P55	CLS_1k_NPU	Generic_1024Input_FFTBIN_64Feature_8Frame
config_MSPM0.yaml	MSPM0G5187	CLS_13k_NPU	FFT256Input_FE_RFFT_16Feature_8Frame_removeDC_Full_Bandwidth
config_CC1352.yaml	CC1352	CLS_1k_NPU	FFT256Input_FE_RFFT_16Feature_8Frame_removeDC_Full_Bandwidth
config_CC1354.yaml	CC1354	CLS_1k_NPU	FFT256Input_FE_RFFT_16Feature_8Frame_removeDC_Full_Bandwidth
config_CC2755.yaml	CC2755	CLS_1k_NPU	FFT256Input_FE_RFFT_16Feature_8Frame_removeDC_Full_Bandwidth
config_CC35X1.yaml	CC35X1	CLS_1k_NPU	FFT256Input_FE_RFFT_16Feature_8Frame_removeDC_Full_Bandwidth

The non-C2000 configs use the dataset name hello_world_example_dsg with the input data path https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/generic_timeseries_classification.zip.

7.9.1.3. Running the Example

Install Tiny ML ModelZoo first (see Installation and Quickstart), then run from the tinyml-modelzoo folder:

Linux / macOSWindows

./run_tinyml_modelzoo.sh examples/generic_timeseries_classification/config.yaml

To target another device, pass its config file, for example examples/generic_timeseries_classification/config_MSPM0.yaml.

7.9.1.4. Dataset Format

For classification tasks, Tiny ML ModelZoo expects the dataset in this structure:

{dataset_name}.zip/
|
|-- classes/
      |-- class_1/
      |     |-- file1.csv
      |     |-- file2.csv
      |
      |-- class_N/
            |-- ...

Note

The folder names under classes/ become the class labels. No annotation files are required. See Classification Dataset Format for details.

7.9.1.5. Configuration (config.yaml)

common:
  task_type: generic_timeseries_classification
  target_device: F28P55
dataset:
  dataset_name: generic_timeseries_classification
data_processing_feature_extraction:
  feature_extraction_name: Generic_1024Input_FFTBIN_64Feature_8Frame
  variables: 1
training:
  model_name: CLS_1k_NPU
  batch_size: 256
  training_epochs: 20
  early_stopping: true
  early_stopping_patience: 5
  num_gpus: 0
testing: {}
compilation: {}

The Generic_1024Input_FFTBIN_64Feature_8Frame preset uses frames of 1024 samples, computes the FFT, bins it into 64 features per frame and concatenates 8 frames (64 x 8 = 512 features). The config file also lists, as comments, the other generic presets (..._FFT_..., ..._RAW_...) you can try. For all keys see How to Create a ModelZoo Config YAML.

7.9.1.6. Output Location

Results are written to ./data/projects/generic_timeseries_classification/run/<date-time>/CLS_1k_NPU/ (relative to where you ran the script; configurable with common.projects_path). The most important artifacts are the trained model under training/, the quantized ONNX model, and the compiled artifacts under compilation/artifacts. See Understanding the Run Output.

7.9.1.7. Expected Results

The example readme.md does not publish accuracy numbers. The waveforms are well separated, so training converges quickly; check the printed test accuracy and confusion matrix at the end of your own run.

7.9.1.8. Variations to Try

• Raw time-domain input: feature_extraction_name: Generic_512Input_RAW_512Feature_1Frame

• Other model sizes: CLS_100_NPU (smaller) or CLS_4k_NPU (larger)

• Another device: use one of the per-device configs above

7.9.1.9. Next Steps

1. Try DC Arc Fault Detection for a real-world arc fault detection example.

2. Use your own data: Classification Dataset Format.

3. Learn all configuration options: How to Create a ModelZoo Config YAML.