training-evaluation

# training-evaluation Use this sub-skill when you need to train 3DDFA, adapt a loss recipe, validate the training data layout, resume from a checkpoint, or interpret AFLW / AFLW2000 benchmark results. ## Use this route for - `train.py` command setup and the bundled `training/*.sh` recipes. - Choosing between WPDC, VDC, and PDC. - Checking filelists, param files, and the expected cropped-image layout. - Understanding checkpoint naming and resume behavior. - Running or interpreting `benchmark.py`, `benchmark_aflw.py`, and `benchmark_aflw2000.py`. ## Do not use this route for - Inference-only or demo flows; use the python-inference route instead. - Geometry rendering / visualization outputs; use the geometry-rendering route instead. - ONNX or C++ export; use the cpp-onnx-port route instead. ## Read first - `references/training-and-losses.md` - `references/data-layout.md` - `references/evaluation-benchmarks.md` - `references/troubleshooting.md` ## Skill-owned script - `scripts/validate_training_args.py` — safe checker for training command templates, filelists, param paths, and GPU device ids. ## Operating notes - Native training is CUDA-centric: `train.py` calls `torch.cuda.set_device(...)` and wraps the model with `nn.DataParallel(...).cuda()`. - The shipped loss modules import in a CPU-only Python environment, but their forward paths still expect CUDA tensors. - `--resume` restores the model state dict only; optimizer state is not bundled in the saved checkpoints. - The bundled shell recipes are templates. Adjust `--root`, filelists, param paths, and `--devices-id` to your local layout before running. - If you already have predicted params, use the benchmark helper flow described in the benchmark reference rather than re-running inference. ## Cross-links - For training command validation, run the skill-owned script. - For benchmark data requirements and NME meaning, read the evaluation reference.