3.1. Quickstart

From zero to a trained and compiled model in a few minutes.

3.1.1. Prerequisites

  • Python 3.10 (needed for the compile step; see Prerequisites)

  • Tiny ML ModelZoo installed (see Installation)

  • For the compile step: the compiler for your target device (see Environment Variables). The hello world example targets the F28P55, so it needs C2000_CG_ROOT (default ~/ti/ti-cgt-c2000_25.11.1.LTS).

3.1.2. Step 1: Get the Example Config

With the clone install the example is already in examples/:

cd tinyml-modelzoo

With the wheel install the examples/ folder is not part of the package. Download examples/generic_timeseries_classification/config.yaml from the GitHub repository into any folder.

3.1.3. Step 2: Run Hello World

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

Tip

For MSPM0 devices use config_MSPM0.yaml where it exists, for example examples/generic_timeseries_classification/config_MSPM0.yaml. Connectivity devices have config_CC1352.yaml, config_CC1354.yaml, config_CC2755.yaml and config_CC35X1.yaml.

The run:

  1. Downloads the waveform dataset (sine, square, sawtooth)

  2. Applies FFT-based feature extraction (Generic_1024Input_FFTBIN_64Feature_8Frame)

  3. Trains CLS_1k_NPU (early stopping is on, with patience 5)

  4. Quantizes the model for the MCU

  5. Compiles for the F28P55 (NPU device)

3.1.4. Step 3: View Results

Results are in:

./data/projects/generic_timeseries_classification/run/<date-time>/CLS_1k_NPU/

relative to the folder you ran the command from (change with common.projects_path). The layout is:

<date-time>/CLS_1k_NPU/
├── run.yaml, status.yaml              # resolved parameters and stage status
├── training/
│   ├── base/                          # float model
│   └── quantization/                  # quantized model (the one you deploy)
└── compilation/
    ├── artifacts/mod.a, tvmgen_default.h
    └── <date-time>_CLS_1k_NPU_F28P55.zip

Understanding the Run Output explains every file.

3.1.5. Step 4: Read the Config

examples/generic_timeseries_classification/config.yaml (comments omitted):

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: {}

  • task_type: the ML task

  • target_device: the MCU to compile for

  • dataset_name: a built-in dataset preset, downloaded automatically

  • feature_extraction_name: how raw data is preprocessed

  • model_name: the model architecture from the model zoo

3.1.6. Step 5: Try Another Example

./run_tinyml_modelzoo.sh examples/dc_arc_fault/config_dsk.yaml

Note that some example folders have no plain config.yaml; look at the files in the folder (for dc_arc_fault there are config_dsk.yaml, config_dsi.yaml and others).

3.1.7. Next Steps