---

name: "legacy-models" description: "Operate the repository's legacy TensorRT Caffe demos: build GoogLeNet classification and the three-stage MTCNN face detector, compile the pytrt Cython bridge, validate model/label assets, and diagnose version-bound runtime failures." disable-model-invocation: true metadata: disco-role: operating license: MIT

Legacy Caffe models: GoogLeNet and MTCNN

Use this sub-skill when a task concerns the googlenet/ or mtcnn/ demos, pytrt.pyx/pytrt.pxd, trt_googlenet*.py, trt_mtcnn.py, or utils/mtcnn.py. It describes the repository snapshot's legacy Caffe and TensorRT execution path. It does not install packages, download models, or ship generated engines, shared libraries, Caffe weights, or installers.

The normal working directory for every command below is the repository root. Run python3 scripts/validate-legacy-model-assets.py --repo-root . before a build. Read references/workflows.md for the ordered build/inference procedure, api-reference.md for tensor and threshold contracts, and compatibility.md before selecting a TensorRT version.

Operating boundary

• Treat the checked-out repository and its documented source files as the evidence boundary. The checkout may contain large model weights, but this skill intentionally references them by path rather than copying them.
• Engine files are TensorRT-version-, platform-, and GPU-dependent outputs. Build them on the target and do not reuse an engine from another host unless deserialization has been explicitly proven for the exact runtime.
• The Caffe parser and the Cython pytrt extension are required for these two demos. A CPU-only Python check cannot prove inference readiness.
• Native --help is conditional: both demo scripts import pytrt at module import time, so help succeeds only after a compatible extension has been built and its TensorRT/CUDA libraries can load.
• Keep generated *.engine, *.so, pytrt.cpp, and build directories out of this runtime skill tree. The supplied validator is read-only.

Quick decision table

Need	Entry point	Required generated assets
Classify an image or stream	trt_googlenet.py	googlenet/deploy.engine, built pytrt
Detect faces and landmarks	trt_mtcnn.py	mtcnn/det1.engine, det2.engine, det3.engine, built pytrt
Build GoogLeNet engine	googlenet/create_engine.cpp	deploy.prototxt, deploy.caffemodel, TensorRT Caffe parser
Build MTCNN engines	mtcnn/create_engines.cpp	three det*_relu prototxt/weights, TensorRT Caffe parser
Build Python bridge	root Makefile → setup.py	Cython, NumPy, CUDA/TensorRT headers and libraries
Check source/assets without mutation	scripts/validate-legacy-model-assets.py	Python standard library only

If the request is to modernize these demos, export them, or replace Caffe, stop at the compatibility boundary and describe the migration separately; do not silently reinterpret a legacy workflow as modern TensorRT support.

Core workflow

1. Confirm assets and runtime. Validate source descriptors and expected filenames. Confirm that the target has a matching NVIDIA driver, CUDA, TensorRT Caffe parser/runtime, a C++ toolchain, Python 3, NumPy, OpenCV, and Cython. For Jetson, identify the JetPack/TensorRT pairing first.
2. Build engines on the target. From googlenet/, run make then ./create_engine; from mtcnn/, run make then ./create_engines. The programs serialize and immediately deserialize engines and print binding information. A successful compiler invocation alone is insufficient.
3. Build pytrt. From the repository root, run make PYTHON=python3. This invokes setup.py build_ext -if, compiles the Cython module against trtNet.cpp, and removes only the temporary build/ directory. The resulting extension is generated at the root and is intentionally not part of this skill.
4. Run a low-risk image trial. Use --image with a local image and keep the process in a display-capable session. Image input repeats indefinitely; press Esc to stop. Use --copy_frame for live sources if inference can outpace capture and the displayed frame is modified in place.
5. Check outputs. GoogLeNet must return a prob tensor with 1000 scores and display top three labels. MTCNN must return (N,5) boxes and (N,10) landmarks, with no NaNs and coordinates clipped to the image bounds. Record the exact TensorRT version, GPU, source commit, engine build command, and input mode with the result.

Safety and failure handling

Do not run install*.sh, download_yolo.sh, sudo pip, or network fetches as part of this workflow. Do not infer that an absent weight is safe to replace with a random or newer model: descriptor/weight compatibility is required. Use the read-only validator to distinguish missing assets from a broken build. For failures, classify them as missing source/weight, toolchain/header/linker, engine deserialization/version, extension import, input/GStreamer, or model post-processing. See troubleshooting.md.

Difficult synthetic usability cases

1. MTCNN scale-and-threshold case: Given a 1920×1080 BGR frame and the default minsize=40, predict that TrtMtcnn.detect rescales to at most 1280×720, raises the effective minimum size but never below 40, converts BGR→RGB, and applies PNet/RNet/ONet thresholds 0.7/0.6/0.7. Confirm that a caller asking for minsize=32 is rejected rather than silently accepted.
2. Mixed-runtime asset case: Given all six Caffe model files and descriptors but no generated engines or pytrt, use the validator to report source/model readiness with generated-runtime warnings; explain why both demo --help commands remain blocked until a compatible extension and engines exist.