matching-pmh 0.7.1

Matching Principle for PyTorch: estimate Sigma_task (D1-D7), matched PMH on any encoder (ResNet, ViT, GNN, LLM)

matching-pmh

Deployment geometry in. Matched robustness out.

Estimate Sigma_task (D1–D7) · train any encoder with matched PMH · falsify with controls

PyPI · GitHub · Walkthroughs · Theory · Quickstart

---

matching-pmh is a research-grade PyTorch library for the Matching Principle:

1. Name what changes at deployment without changing the label.
2. Estimate that nuisance geometry Sigma_task (covariance of label-preserving deployment variation).
3. Add a matched Jacobian penalty on your representations h = phi_theta(x).

Works with your stack: ResNet, ViT, GNN, Whisper-style encoders, causal LMs with LoRA, or frozen features + sklearn. Full math (LaTeX): docs/THEORY.md.

Design goal: two phases, one hook tensor h, no framework lock-in. The paper's thirteen task blocks are validation examples; integrate in an afternoon using the walkthroughs.

---

30-second start

pip install matching-pmh
python examples/01_domain_shift_d4.py
pmh-train list-methods

import pmh
from pmh import SigmaTaskConfig, PMHConfig, PMHLoss, collect_features, estimate_from_config

# Phase A — estimate (frozen encoder)
artifact = estimate_from_config(SigmaTaskConfig.for_domain(rank=32), h_source, h_target)
artifact.save("artifacts/sigma")

# Phase B — train (your loop)
pmh_loss = PMHLoss(artifact, PMHConfig(weight=0.3, cap_ratio=0.3, warmup_epochs=2))
total, _ = pmh_loss.capped_total(task_loss, h)

Quickstart guide · 16 walkthroughs

---

Problem, object, repair, unification

Problem	ERM uses every input direction that predicts labels—including nuisances harmful at deployment (lighting, site, sensor noise, formatting, renameable identifiers, …).
Object	Sigma_task = covariance of label-preserving deployment nuisance n (under law Q_n).
Repair	Matched PMH shrinks encoder sensitivity along Sigma_task, not uniformly (isotropic PMH / generic VAT).
Unification	CORAL, domain Grams, augmentation stacks, metric-learning directions, adversarial subspaces, and style Grams are different estimators of the same Sigma_task (Lemma D1–D7).

Matched loss (schematic): L = L_task + lambda * Tr(J_phi^T J_phi Sigma') with range(Sigma') covering range(Sigma_task). Details: THEORY.md.

---

How it fits your codebase

 Phase A (once)              Phase B (every step)
 ----------------              --------------------
 source/target data    ->      x, y from your loader
       |                            |
 encoder (eval)        ->      encoder (train) -> h
       |                            |
 estimate D1-D7        ->      L_task(h, y) + PMHLoss(h, Sigma_hat)
       |
 artifact.pt

You keep	Library adds
Model, optimizer, task loss	SigmaTaskConfig, estimate_from_config
Data loaders	collect_features (optional)
Training loop / Trainer	PMHLoss.capped_total or PMHTrainer

---

Walkthroughs (16 guides)

#	Guide	Run
1	PyTorch + D4	examples/01_domain_shift_d4.py
2	ResNet + D4	examples/12_resnet_hook_d4.py
3	Office-31 + sklearn	examples/06_office31_sklearn.py
4	Multi-layer CNN	examples/07_vision_multilayer.py
5	Compositional D5	examples/13_compositional_train_d5.py
6	LLM style D7	examples/08_hf_style_d7.py
7	HF Trainer + DPO	examples/11_dpo_lora_style_pmh.py
8	Falsification controls	examples/04_falsification_controls.py
9	CLI JSON jobs	pmh-train estimate --config ...
10	Lightning	examples/09_lightning_module.py
11	Temporal D6	API in guide
12	ViT / CLS + D4	examples/14_vit_cls_d4.py
13	Speech encoder + D4	examples/15_speech_encoder_d4.py
14	QM9 / molecules D5	examples/16_qm9_molecule_d5.py
15	Code / tokens D5	examples/17_code_tokens_d5.py
16	Augmentations D3	examples/18_augmentation_d3.py

---

Estimators at a glance (D1–D7)

Deployment story	Method	SigmaTaskConfig
Different site / camera / corpus; P(y given x) stable	D4	SigmaTaskConfig.for_domain(rank=32)
Low-rank shift; labels on both domains	D1	SigmaTaskConfig.for_subspace(rank=32)
Unstructured sensor / acquisition noise	D2	SigmaTaskConfig.for_isotropic(dim, noise_level)
Known augmentation modes (color, blur, crop, …)	D3	SigmaTaskConfig.for_augmentation() + aug_deltas
Nuisance on specific coordinates (atoms, tokens)	D5	SigmaTaskConfig.for_compositional(indices)
Drift along time within a sequence	D6	SigmaTaskConfig.for_temporal()
LLM style / format; semantics fixed	D7	SigmaTaskConfig.for_alignment(rank=32)

pmh-train list-methods

Hybrid nuisances: estimate separate Sigma matrices and add separate PMHLoss terms.

---

Install

pip install matching-pmh

Extra	Use case
pip install "matching-pmh[vision]"	ResNet / ViT examples
pip install "matching-pmh[hf]"	D7 style Gram (Transformers)
pip install "matching-pmh[hf-lora]"	LoRA + DPO example
pip install "matching-pmh[sklearn,vision]"	Office-31 pipeline
pip install "matching-pmh[lightning]"	Lightning callback
pip install "matching-pmh[all]"	Development + docs

From source:

git clone https://github.com/vishalstark512/matching-pmh.git
cd matching-pmh && pip install -e ".[dev]" && pytest -q

---

Documentation

Document	Purpose
QUICKSTART.md	First run in 10 minutes
THEORY.md	Full mathematics (LaTeX)
ARCHITECTURES.md	Hook points per stack
walkthroughs/	16 end-to-end guides

---

Citation

Cite the Grand Unification / Matching Principle manuscript. See CITATION.cff in the repository.

@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}
}

---

Contributing

See CONTRIBUTING.md.

License

MIT — see LICENSE.