---

name: bapps-training description: "Routes BAPPS training, fine-tuning, checkpointing, and smoke-test workflows for the PerceptualSimilarity package." disable-model-invocation: true metadata: disco-role: operating license: BSD 2-Clause

BAPPS Training

Use this sub-skill when the task is about training or fine-tuning the LPIPS metric on BAPPS-style 2AFC data.

Trigger phrases

• "train LPIPS"
• "fine-tune on BAPPS"
• "run the scratch variant"
• "run the tune variant"
• "smoke-test the training path"
• "save checkpoints"

What this route covers

• LPIPS/BAPPS ranking-loss training.
• Checkpoint directory creation and file naming.
• Smoke-friendly one-step or one-epoch runs.
• Scratch and trunk-tuning modes.
• Training-specific troubleshooting.

What this route excludes

• Direct image-pair comparison.
• Benchmark scoring only.
• Dataset-download automation.

If the task is only about scoring a split, route to bapps-evaluation. If the task is about direct LPIPS distances or LPIPS loss visualization, route to metric-usage.

Read these next

• references/workflows.md for the training command matrix.
• references/troubleshooting.md for checkpoint, dependency, and runtime issues.
• ../../references/bapps-dataset.md for the expected split layout.
• ../../references/api-reference.md for the verified Trainer API.

Run these helpers

• scripts/train_bapps.py for the bundled training loop.
• scripts/train_test_metric.sh for the standard train-then-score wrapper.
• scripts/train_test_metric_scratch.sh for the scratch variant.
• scripts/train_test_metric_tune.sh for the trunk-tune variant.
• ../../scripts/make_tiny_bapps_fixture.py to create a tiny BAPPS-style smoke fixture.

Working assumptions

• Training uses 2AFC splits with ref/, p0/, p1/, and judge/.
• The bundled helper creates the checkpoint directory automatically.
• The bundled helper avoids the old HTML/visdom stack used by the stock train.py.
• The smoke default is intentionally bounded so it is safe to run on small fixtures.