model-inspect-tool 0.1.0

Inspect input and output metadata for common AI model formats.

model-inspect

A unified command-line tool for inspecting model input/output metadata.

Supported backends:

• ONNX: native onnx parser
• TFLite: tflite-runtime or TensorFlow Lite Interpreter
• PyTorch: TorchScript by default; checkpoint loading only with explicit opt-in
• TensorRT Engine: TensorRT Python API or trtexec
• CVIModel/BModel: external model_tool --info
• RKNN: best-effort external RKNN toolkit integration

Install

Core CLI:

pip install model-inspect-tool

Install common open-format backends:

pip install "model-inspect-tool[onnx,torch]"

On Linux, TFLite can often be installed with:

pip install "model-inspect-tool[tflite]"

Alternatively install TensorFlow:

pip install "model-inspect-tool[tensorflow]"

Vendor formats require their matching SDK/runtime:

• TensorRT: NVIDIA TensorRT and/or trtexec
• CVIModel: TPU-MLIR model_tool
• RKNN: RKNN Toolkit2

Usage

Human-readable output:

model-inspect model.onnx

JSON output:

model-inspect model.onnx --json

Write JSON to a file:

model-inspect model.onnx --json --output report.json

Force a backend:

model-inspect unknown.bin --format onnx

Inspect an ordinary PyTorch checkpoint:

model-inspect model.pt --allow-unsafe-pickle

Warning: ordinary PyTorch checkpoints may use Python Pickle. Never enable --allow-unsafe-pickle for untrusted model files.

Output schema

{
  "format": "onnx",
  "path": "/absolute/path/model.onnx",
  "backend": "onnx",
  "inputs": [
    {
      "name": "images",
      "shape": [1, 3, 640, 640],
      "dtype": "float32"
    }
  ],
  "outputs": [],
  "dynamic": false,
  "metadata": {},
  "warnings": []
}

Development

python -m venv .venv
source .venv/bin/activate
python -m pip install -e ".[dev,onnx,torch]"
pytest
model-inspect tests/assets/example.onnx --json

Build

python -m build
python -m twine check dist/*

Security

The tool does not load arbitrary PyTorch Pickle checkpoints unless the user explicitly supplies --allow-unsafe-pickle. Vendor command execution uses argument arrays and does not invoke a shell.