---

name: training-evaluation description: "Use Torch Points3D Hydra training, evaluation, Trainer, checkpoints, forward inference, visualization, and experiment-output helpers." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Torch Points3D Training and Evaluation

Use this sub-skill when the user asks to compose Torch Points3D train.py or eval.py commands, debug Trainer, resume or evaluate checkpoints, inspect outputs/<date>/<time> runs, configure W&B/TensorBoard/visualization, or run forward inference on unlabeled data from a trained checkpoint.

Read First

• Read training/evaluation workflows for command patterns, safe CPU smoke options, and trainer behavior.
• Read configuration and checkpoints for Hydra selector rules, ModelCheckpoint, resume/eval fields, and output directory conventions.
• Read forward inference before adapting checkpoint-based inference for unlabeled point clouds.
• Read training/evaluation troubleshooting for Hydra errors, dataset downloads, logging/profiler side effects, checkpoint mismatch, and old config conversion.
• Run compose_config_smoke.py to validate selectors without training.
• Run summarize_runs.py to list checkpoint metrics without deleting runs.
• Run convert_checkpoint_omegaconf.py only when an old checkpoint stores OmegaConf containers that need conversion.
• Run forward_preflight.py before a real forward-inference job.

Main Workflows

Compose a training command

Training is a Hydra script, not an installed console entry point. In a Torch Points3D-style project containing train.py and conf/, specify the task, model config group, data config group, and model entry:

python train.py \
  task=segmentation \
  models=segmentation/pointnet2 \
  data=segmentation/shapenet-fixed \
  model_name=pointnet2_charlesssg \
  training.wandb.log=False \
  training.tensorboard.log=False \
  training.tensorboard.pytorch_profiler.log=False \
  debugging=early_break

Use the config smoke script first if the user is unsure whether selectors match. For real runs, remove debugging=early_break and configure data roots, logging, CUDA, batch size, and output directories intentionally.

Evaluate a checkpoint

eval.py uses conf/eval.yaml. The essential fields are checkpoint_dir, model_name, weight_name, batch_size, num_workers, cuda, precompute_multi_scale, and tracker_options.

python eval.py checkpoint_dir=/path/to/run model_name=pointnet2_charlesssg weight_name=latest cuda=-1

Set cuda=-1 for CPU-only evaluation. Keep tracker_options.make_submission and full_res explicit when evaluating datasets that support submissions or full-resolution voting.

Inspect output runs

python sub-skills/training-evaluation/scripts/summarize_runs.py --outputs-dir outputs --json

The bundled helper finds .pt checkpoints, reads their stats dict when possible, and reports malformed/empty folders without deleting anything.

Forward inference from a checkpoint

The repo forward workflow adapts a checkpoint's training data config to a forward dataset class and writes .npy predictions. It needs a real checkpoint, input data root, and output directory. Use the preflight helper before launching the real job:

python sub-skills/training-evaluation/scripts/forward_preflight.py \
  --checkpoint-dir /path/to/run --model-name pointnet2_charlesssg \
  --input-path /path/to/unlabeled-data --output-path /path/to/predictions

Then see forward inference for the command shape and caveats.

Boundary Rules

• For standalone PointNet2/KPConv/RSConv forward APIs without a Hydra run, use model-apis.
• For data layout, transform, or dataset-class lookup problems, use datasets-transforms.
• For registration descriptor evaluation or test-set protocols, use registration-workflows.

Safety Checklist

• Do not run full training, cross-validation, W&B checkpoint downloads, or dataset downloads without user approval for runtime, network, data size, and writes.
• For smoke runs, disable W&B/TensorBoard/profiler and set debugging=early_break.
• Prefer cuda=-1 for CPU checks; use GPU only after the dependency stack matches the target CUDA build.
• Treat skipped sparse/CUDA checks as optional limitations, not as passes.