9. Model Compile and Evaluate

The EdgeAI Model Zoo runs pre-compiled artifacts on the AM62Dx target. This section covers the x86 host workflow that produces those artifacts: how to compile an audio model for the C7™ NPU, run host-emulation inference, and measure accuracy against a data set - using edgeai-tidlrunner. It runs entirely on an Ubuntu x86 PC; nothing here touches the target.

This is the workflow for any audio model, not just the reference models. The four reference models below are worked examples. To bring your own model you point the same tidlrunner-cli at your ONNX graph and a per-model YAML config, using a reference config as a starting template. Which operators offload to the C7™ NPU (versus fall back to the Arm core) is determined by TIDL - see edgeai-tidl-tools for the authoritative list of supported operators.

9.1. edgeai-tidlrunner

edgeai-tidlrunner is a bring-your-own-model (BYOM) command-line wrapper over the core edgeai-tidl-tools compilation stack. A single YAML config per model drives every stage, and one CLI - tidlrunner-cli - exposes the stages this section uses:

Command

What it does

compile

Imports the ONNX graph, quantizes and partitions it, and writes the deployable TIDL/TVM-RT artifacts.

infer

Runs the compiled model on the x86 host (PC emulation) to check output correctness before deploying to the target.

evaluate

Runs the compiled model over a data set and reports accuracy or speech-enhancement metrics.

inspect

One-shot compile + inference + reference comparison that emits an interactive HTML analysis report (see Model Inspect).

All commands are run from the repository root (edgeai-tidlrunner/) and take --config_path <cfg> pointing at the per-model YAML. The same config file is reused across compile, infer, and evaluate.

9.2. Setup

Model compilation runs on an x86 PC (Ubuntu Linux recommended). The AM62D audio flow is self-contained: a single setup script installs the whole TVM toolchain into a dedicated Python 3.10 environment. The TIDL tools target Python 3.10. These steps assume pyenv.

  1. Clone the repository (https://github.com/TexasInstruments/edgeai-tidlrunner):

    $ git clone https://github.com/TexasInstruments/edgeai-tidlrunner.git
$ cd edgeai-tidlrunner

  2. Create and activate a dedicated Python 3.10 environment. The AM62D flow uses a TI TVM wheel; keep it in its own venv (tidlrunner-am62d) so it never clobbers a standard tidlrunner setup:

    $ pyenv install 3.10
$ pyenv virtualenv 3.10 tidlrunner-am62d
$ pyenv activate tidlrunner-am62d

  3. Run the AM62D setup script. With the venv active, one script installs everything this flow needs:

    $ ./devices/setup_am62d.sh

    It downloads the ARM GCC 15.2 and C7000 CGT 5.0.0 LTS cross-toolchains into tools/tidl_tools_package/bin/ (skipped if already present), installs the x86 TI TVM wheel - which bundles the AM62D x86 TIDL tools inside the package - and installs tidlrunner[pc,audio], tools, onnxruntime, and tidl_onnx_model_optimizer. The [audio] extra pulls in the audio packages (librosa, soundfile, scipy, pesq, pystoi, scikit-learn). No separate pip install is needed.

  4. Source the env script once per shell. TVM dlopens the TIDL runtime (libvx_tidl_rt.so) before Python’s own import runs, so the tool paths must be live in the shell before tidlrunner-cli starts. Source (do not execute) devices/am62d_env.sh once; every tidlrunner-cli command run afterward in that shell picks the vars up:

    $ source devices/am62d_env.sh

    It exports:

    Variable

    Value

    TIDL_TOOLS_PATH

    <tvm-package-dir>/3rdparty/x86_tidl_tools/AM62D (resolved from the active venv’s installed tvm)

    LD_LIBRARY_PATH

    prepends TIDL_TOOLS_PATH so libvx_tidl_rt.so resolves

    ARM64_GCC_PATH

    the aarch64 GNU toolchain installed by setup_am62d.sh under tools/tidl_tools_package/bin/ (override by exporting before sourcing)

    CGT7X_ROOT

    the C7000 code-generation tools installed by setup_am62d.sh under tools/tidl_tools_package/bin/ (override by exporting before sourcing)

    SOC

    am62d

9.2.1. Data Sets

Accuracy evaluation needs the reference data sets. Compilation alone can run on random data (for latency-only measurement), but evaluate requires the real sets. Download them from the repository root:

$ # UrbanSound8K (~5.6 GB) - sound classification (VGGish11, YAMNet)
$ bash examples/audio/scripts/download_urbansound8k.sh

$ # VoiceBank-DEMAND-16k (~2 GB) - speech enhancement (GCRN)
$ python3 examples/audio/scripts/download_voicebank_demand.py

They land under data/datasets/UrbanSound8K/ and data/datasets/VoiceBank-DEMAND-16k/ - the paths the model configs expect. The classification configs use UrbanSound8K fold 10 (837 samples) as the test set; the speech-enhancement configs use the VoiceBank-DEMAND-16k test set (824 files).

The reference ONNX models auto-download at compile time via .link files, so fetching them ahead of time is optional. To pre-download all reference ONNX models:

$ bash examples/audio/scripts/download_audio_models.sh

9.3. Compile and evaluate

Each model has one config that drives all stages. All four run under the tidlrunner-am62d environment with devices/am62d_env.sh already sourced (see Setup). Three models compile to the C7™ NPU through the TVM runtime. The AM62D config paths are:

Model

Model ID

Runtime

Tensor Bits

Config path (relative to edgeai-tidlrunner/)

VGGish11

ac-0101

TVM-RT

8

data/configs/samples/models/audio/audio_classification/urbansound8k/vggish11_tvmrt_config.yaml

YAMNet

ac-0201

TVM-RT

8

data/configs/samples/models/audio/audio_classification/urbansound8k/yamnet_tvmrt_config.yaml

GCRN

ase-0201

TVM-RT

16 (LSTM), FP32 (Enc/Dec)

data/configs/samples/models/audio/speech_enhancement/voicebank_demand_16k/gcrn_fixed_4sec_tvmrt_config.yaml

9.3.1. VGGish11 (TVM-RT, 8-bit)

Runs on the C7™ NPU (tidl_offload: true):

$ CFG=data/configs/samples/models/audio/audio_classification/urbansound8k/vggish11_tvmrt_config.yaml
$ tidlrunner-cli compile  --config_path $CFG
$ tidlrunner-cli infer    --config_path $CFG
$ tidlrunner-cli evaluate --config_path $CFG

evaluate reports top-1 / top-5 / macro-F1 over UrbanSound8K fold 10.

9.3.2. YAMNet (TVM-RT, 8-bit)

$ CFG=data/configs/samples/models/audio/audio_classification/urbansound8k/yamnet_tvmrt_config.yaml
$ tidlrunner-cli compile  --config_path $CFG
$ tidlrunner-cli infer    --config_path $CFG
$ tidlrunner-cli evaluate --config_path $CFG

9.3.3. GCRN (TVM-RT, 16-bit)

GCRN is C7™ NPU-offloaded at 16-bit (tidl_offload: true, fixed 4-second input):

$ CFG=data/configs/samples/models/audio/speech_enhancement/voicebank_demand_16k/gcrn_fixed_4sec_tvmrt_config.yaml
$ tidlrunner-cli compile  --config_path $CFG
$ tidlrunner-cli infer    --config_path $CFG
$ tidlrunner-cli evaluate --config_path $CFG

evaluate reports PESQ / STOI / SI-SDR over the VoiceBank-DEMAND-16k test set.