eml-cost-torch 0.4.0

Per-layer Pfaffian profile and EML routing depth for any torch.nn.Module.

eml-cost-torch

Per-layer Pfaffian profile and EML routing depth for any torch.nn.Module.

Walks a PyTorch model, classifies each leaf module by mapping it to its symbolic SymPy equivalent, and runs eml-cost to produce a Pfaffian chain order, EML routing depth, and canonical axes-tuple per layer.

Install

pip install eml-cost-torch          # core (sympy + eml-cost only)
pip install "eml-cost-torch[torch]" # adds torch>=2.0 for live model walking

Quick start

import torch.nn as nn
from eml_cost_torch import summary, profile

model = nn.Sequential(
    nn.Linear(64, 32),
    nn.GELU(),
    nn.LayerNorm(32),
    nn.Linear(32, 16),
    nn.Sigmoid(),
)

print(summary(model))

Output:

==========================================================================================
  Per-layer Pfaffian profile  (5 leaf modules)
==========================================================================================
  name                              class                 axes                r     depth
  --------------------------------------------------------------------------------------
  0                                 Linear                p0-d2-w0-c0         r= 0  d=  2
  1                                 GELU                  p3-d5-w3-c0         r= 3  d=  6
  2                                 LayerNorm             p1-d4-w1-c0         r= 1  d=  4
  3                                 Linear                p0-d2-w0-c0         r= 0  d=  2
  4                                 Sigmoid               p1-d1-w1-c-1        r= 1  d=  2
  --------------------------------------------------------------------------------------
  total r (sum across leaves): 5
  max r in any leaf:           3
  max predicted_depth:         6
  distinct cost classes:       4
    p0-d2-w0-c0       x 2
    p3-d5-w3-c0       x 1
    p1-d4-w1-c0       x 1
    p1-d1-w1-c-1      x 1

Programmatic access

from eml_cost_torch import profile, profile_dict

# Returns list of LayerProfile dataclass instances
rows = profile(model)
for r in rows:
    print(r.class_name, r.axes, r.pfaffian_r)

# Or as JSON-friendly dicts
import json
json.dumps(profile_dict(model))

What's measured per layer

Field	Meaning
pfaffian_r	Total Pfaffian chain order (Khovanskii convention)
max_path_r	Chain order along the deepest path
eml_depth	EML routing tree depth
predicted_depth	Full predicted depth (chain + structural + corrections)
axes	Canonical fingerprint p<r>-d<n>-w<m>-c<k>
is_pfaffian_not_eml	True for Bessel, Airy, Lambert W, etc.

Supported torch.nn classes

50+ classes registered out of the box, including:

• Linear / Conv (Linear, Conv1d/2d/3d, ConvTranspose*)
• Activations (ReLU, GELU, Sigmoid, Tanh, SiLU, Mish, ELU, Softmax, ...)
• Normalisation (LayerNorm, BatchNorm*, GroupNorm, RMSNorm)
• Pooling (MaxPool*, AvgPool*, adaptive variants)
• Regularisation (Dropout, AlphaDropout)
• Embedding (Embedding, EmbeddingBag)
• Attention (MultiheadAttention — simplified score)

Unknown layer classes are reported as UNKNOWN rather than raising.

See also

• eml-cost — the underlying Pfaffian profile substrate
• monogate — EML arithmetic, witnesses, CLI
• monogate.org — research site

License

PROPRIETARY PRE-RELEASE — see LICENSE.