matching-pmh 1.5.1

Domain-robust PyTorch and sklearn: train on site A, deploy on site B with the same labels.

matching-pmh

Train on site A. Deploy on site B. Same labels.

Add a domain-robustness regularizer to your PyTorch model or sklearn pipeline — without replacing your architecture or reading a research paper first.

PyPI · GitHub · What is this? · First hour ·

---

Developer API (start here)

from pmh import check_applicability, robust_fit, evaluate_baseline_vs_pmh, evaluate_robust_fit

# Go / marginal / no-go before training
print(check_applicability(stack="pytorch", n_source=500, n_target=400).summary())

# PyTorch: one call
out = robust_fit(model, train_loader, source_batches=src, target_batches=tgt, hook="auto", epochs=20)

# sklearn: baseline vs PMH on target holdout
report = evaluate_baseline_vs_pmh(x_source, y_source, x_target, y_target, compare_to=("coral",))
print(report.summary())

# PyTorch: same report shape on val_loader
report = evaluate_robust_fit(model, train_loader, val_loader, source_batches=src, target_batches=tgt, hook="auto", epochs=10)
print(report.summary())

Three paths only: GOLDEN_PATHS.md · pmh-train wizard · starter template

---

Who this is for

You have	Start here
PyTorch model + source/target data	PMHTrainer + nuisance="domain_shift" · Colab
Frozen .npy / sklearn features	PMHMatcher in a Pipeline · Colab (sklearn)
LLM style/format drift	NLP walkthrough

Not for: new test-time classes, unrelated label definitions, or “make any model robust to everything.”

When PMH helps (honest expectations) → · What is PMH →

---

5-minute try

or locally:

pip install matching-pmh torch
git clone https://github.com/vishalstark512/matching-pmh.git
cd matching-pmh
python examples/00_first_run_domain_shift.py

You get baseline vs PMH target accuracy on synthetic data, then copy the pattern into your project.

Expected output: docs/DEMO_OUTPUT.md

pmh-train wizard

---

Default integration (PyTorch)

from pmh import PMHTrainer, PMHConfig

trainer = PMHTrainer(
    model,
    hook=backbone,
    head=classifier,
    nuisance="domain_shift",
    pmh_config=PMHConfig.balanced(),
    artifact_path="artifacts/my_run.pt",
)
trainer.fit(
    train_loader,
    source_batches=source_loader,
    target_batches=target_loader,
    epochs=20,
)

sklearn Pipeline (frozen features)

You already have embeddings x_source, x_target (same dimension). Adapt, then classify — like any other preprocessing step:

pip install "matching-pmh[sklearn]"

from pmh import PMHMatcher
from sklearn.pipeline import Pipeline
from sklearn.linear_model import LogisticRegression

pipe = Pipeline([
    ("adapt", PMHMatcher(nuisance="domain_shift").fit(x_source, x_target)),
    ("clf", LogisticRegression(max_iter=500)),
])
pipe.fit(x_source, y_source)
# Score on held-out TARGET rows — not source-only accuracy

Gallery: tabular · examples/06_office31_sklearn.py

---

How it works (30 seconds)

flowchart LR
  A[Site A data] --> E[Estimate shift geometry once]
  B[Site B data] --> E
  E --> T[Train your model]
  T --> H[Representation hook h]
  H --> L[Task loss + robustness penalty]

Same hook layer for estimate and train. First hour · Getting started · Troubleshooting glossary

---

How it relates to CORAL / domain adaptation

PMH keeps your task loss and penalizes sensitivity in representation space along directions that differ between train and deploy environments. See vs CORAL.

Under the hood: estimate deployment geometry once, train with an extra matched penalty on hook h. Details are optional: Theory.

---

Documentation map

I want to…	Read
Understand in plain language	WHAT_IS_PMH.md
Run + copy code in one hour	FIRST_HOUR.md
Integrate on my repo	GETTING_STARTED.md
ResNet / ViT / HF hooks	hooks.md · Gallery
Prove claims before production	Controls walkthrough
Replicate paper benchmarks	CORRECT_USAGE · Paper alignment

18 walkthroughs · Developer onboarding plan

---

Install

pip install matching-pmh
pip install "matching-pmh[sklearn]"   # classical ML path
pip install "matching-pmh[vision]"    # ResNet / timm examples
pip install "matching-pmh[hf]"        # LLM style shift

---

Advanced (estimators D1–D7, research)

The Grand Unification paper unifies many nuisance estimators (subspace, domain Gram, augmentations, style, …). The library exposes them when your shift story is not plain cross-domain:

pmh-train list-methods
pmh-train list-presets

Research / benchmark	Doc
Paper block presets	paper-presets-by-block.md
Office-31 reference table	BENCHMARKS.md
Lemma-level math	THEORY.md

---

Citation

@software{matching_pmh,
  title  = {matching-pmh: Matched PMH training from estimated deployment nuisance geometry},
  author = {Rajput, Vishal},
  year   = {2026},
  url    = {https://github.com/vishalstark512/matching-pmh}
}

License

MIT — see LICENSE.