---

name: paddle-workflows description: "Use face.evoLVe PaddlePaddle training, quantization, Paddle Inference, and Paddle Lite workflows safely." disable-model-invocation: true metadata: disco-role: operating license: MIT

paddle-workflows

Use this sub-skill when the task is about the PaddlePaddle side of face.evoLVe: Paddle training configuration, Paddle model component inspection, quantization-aware training, post-training quantization, Paddle Inference server demos, Paddle Lite edge demos, or diagnosing missing Paddle deployment artifacts.

Route here for

• Running or planning PaddlePaddle training from ImageFolder identity folders using IR/IR-SE/ResNet backbones, ArcFace/CosFace/SphereFace/Am_softmax/Softmax heads, and Focal/Softmax losses.
• Inspecting Paddle backbone/head/loss source components without accidentally importing the local paddle/ source package as the PaddlePaddle framework.
• Exporting a trained Paddle backbone to .pdmodel/.pdiparams, enabling QAT with SAVE_QUANT_MODEL, or preparing dynamic/static PaddleSlim quantization.
• Preparing the Paddle Inference FaceDatabase workflow, face_data.fdb, demo video input, GPU predictor settings, and recognition thresholds.
• Preparing Paddle Lite .nb model artifacts and runtime prerequisites for edge inference demos.

Route elsewhere

• Generic ImageFolder identity-folder validation, low-shot pruning, or dataset cleanup belongs in data-preparation.
• MTCNN alignment details and face-crop generation belong in face-alignment unless the Paddle demo-specific FaceDatabase flow is the focus.
• PyTorch training, PyTorch checkpoint extraction, and PyTorch verification metrics belong in pytorch-training or feature-extraction-verification.

Bundled references and scripts

• Read references/paddle-training-deployment.md when planning Paddle training, export, QAT, post-training quantization, Paddle Inference, or Paddle Lite deployment prerequisites.
• Read references/troubleshooting.md when diagnosing PaddlePaddle import shadowing, missing PaddleSlim/requests, absent .pdmodel/.pdiparams/.nb artifacts, hard-coded GPU predictor settings, FaceDatabase/video/font problems, or data layout errors.
• Run scripts/inspect_paddle_components.py for a safe CPU component smoke check that reports the installed PaddlePaddle version and a Paddle backbone output shape from a working face.evoLVe checkout.

Verification boundary

Construction evidence confirmed only safe import/component checks: PaddlePaddle and PaddleSlim imported, and a Paddle IR_50 CPU eval forward produced a [2, 512] embedding shape. Full Paddle training, QAT quality, post-training quantization, Paddle Inference video recognition, and Paddle Lite edge deployment remain artifact/hardware-dependent and are not verified unless the user supplies data, trained/exported models, videos, fonts, and compatible runtimes.