Skip to content

REPRODUCIBILITY.md

Latest commit

 

History

History
76 lines (56 loc) · 2.76 KB

REPRODUCIBILITY.md

File metadata and controls

76 lines (56 loc) · 2.76 KB
Raw

Reproducibility Guide

Environment used for the local evidence runs

Component Local evidence environment
Operating system Windows 11
Python 3.11.9
GPU NVIDIA GeForce RTX 4090 Laptop GPU
CUDA runtime reported by PyTorch 12.8
PyTorch 2.11.0+cu128
Transformers 4.51.3
PEFT 0.19.1
Model google/siglip2-base-patch16-224

The repository’s tracked CI intentionally does not require this GPU environment.

Local roots

Keep the following responsibilities separate:

Role Recommended location
Repository source, tests, configs, local venv repository root
Raw Fashionpedia data external dataset root, outside Git
Hugging Face cache external model-cache root, outside Git
Runs, logs, analysis, recovery records external output root, outside Git
Full local evidence release dist/ under the repository, ignored by Git

Do not commit raw dataset files, model caches, checkpoints, release archives, or runtime logs into Git history.

Reproduction flow

  1. Verify the raw Fashionpedia dataset layout and integrity.
  2. Run the train-only profile and task-definition audit.
  3. Build the image-group-disjoint train/development manifest.
  4. Fit the HOG baseline.
  5. Run the frozen SigLIP2 baseline.
  6. Audit SigLIP2 vision-attention LoRA targets.
  7. Run matched frozen-head versus LoRA development experiments.
  8. Freeze selected checkpoints from development results.
  9. Run untouched official-validation final confirmation.
  10. Freeze local evidence and execute local full-release contracts.
  11. Run hosted static-fixture contracts and documentation validation.

Hosted-CI verification

python -m pip install -r requirements-ci.txt
python -m pytest -q tests/test_ci_release_fixture_contracts.py
python -B scripts/validate_documentation.py

This path validates source-level and static-evidence contracts only.

Local full-evidence verification

With the local release bundle and project venv present:

.\.venv\Scripts\python.exe -B -m pytest -q `
  tests\test_release_evidence_contracts.py `
  -p no:cacheprovider

The local test reads the evidence release. It does not retrain the model or re-score Fashionpedia.

Dependency boundaries

  • requirements-ci.txt is intentionally minimal for hosted static-contract verification.
  • HOG, frozen SigLIP2, and LoRA dependencies are separated into dedicated requirements files.
  • The full local evidence environment is expected to include the project venv used for original evidence generation and verification.

Clean-environment caution

Do not delete .venv until a fresh environment rebuild has been qualified against the exact local evidence-contract tests. Several local drivers call .venv\Scripts\python.exe directly.