8000 GitHub - CodNoob100/canopy: Expressivity-Guided Neural Architecture Search via Tropical Geometry · GitHub
[go: up one dir, main page]

Skip to content
CodNoob100Public

About

Expressivity-Guided Neural Architecture Search via Tropical Geometry

Resources

Stars

1 star

Watchers

0 watching

Forks

Latest commit

 

History

42 Commits

Folders and files

Repository files navigation

CANOPY: Expressivity-Guided Neural Architecture Search via Tropical Geometry

License Python 3.10+

CANOPY is a research-oriented, training-free Neural Architecture Search (NAS) framework that uses tropical geometry to motivate zero-cost expressivity proxies.

Existing zero-cost proxies (like parameter count or kernel matrix rank) correlate with test accuracy empirically but lack theoretical justification for why they measure capacity. CANOPY is derived from the exact mathematical equivalence between ReLU networks and tropical rational functions.

For architecture-only scoring, CANOPY computes combinatorial bounds and structural proxies for linear-region expressivity. These are not exact counts of the regions formed by an untrained network. The weight-aware mode can count or estimate regions for an instantiated network, depending on the counting method.

The repository reports a Spearman rank correlation of $\rho = 0.51$ for its DARTS proxy on NAS-Bench-301, compared with $\rho = 0.33$ for its parameter-count baseline (a reported 56% relative improvement). Reproducing this result requires the relevant benchmark models and evaluation setup.

Key Results

CANOPY outperforms baseline proxies by correctly capturing the architectural expressivity, and scales to complex search spaces like DARTS where traditional metrics fail.

DARTS Rank Correlation Evolutionary Search Trajectory (DARTS)
DARTS Correlation DARTS Search Trajectory
NATS-Bench CIFAR-100 Correlation The Bound Tightness Paradox
CIFAR-100 Correlation Paradox Mechanism

Start here: two paths

Just trying CANOPY? Follow the five-minute, CPU-only quickstart to create a virtual environment, install the core package, and score your first architecture without any benchmark files or GPU.

Building on CANOPY? Use the granular developer route for isolated tests, Python API examples, real-data evaluation, provenance, and troubleshooting.

Installation

CANOPY's core scoring tools require Python 3.10+, NumPy, and SciPy. PyTorch and pycddlib are available as optional extras for downstream integrations; neither is needed to run the core scoring API.

git clone https://github.com/CodNoob100/canopy.git
cd canopy
pip install -e .

To run the benchmarking examples, you will also need the nats-bench dependency:

pip install -e .[bench]

Quickstart

Score a simple feedforward architecture's expressivity directly from its layer widths, with no weights instantiated:

from canopy.nas.architecture import ArchitectureSpec
from canopy.nas.tropical_score import tropical_expressivity_score

# Input width 32, 3 hidden layers [64, 128, 64], output width 10
arch = ArchitectureSpec(layer_widths=(32, 64, 128, 64, 10))

# 'generic' returns an integer upper bound, not a log10 score or exact count
score = tropical_expressivity_score(arch, mode="generic")
print(f"Architecture-level region upper bound: {score}")

Score a DARTS architecture with an approximate convolutional and skip-connection structural proxy (returned in log10 scale):

from canopy.nas.architecture import ArchitectureSpec
from canopy.nas.tropical_score import tropical_expressivity_score_darts

# Define a DARTS cell operation sequence
# e.g. Node 1 gets sep_conv_3x3 from Node 0, Node 2 gets skip_connect from Node 1, etc.
ops = (
    "sep_conv_3x3", "skip_connect", "max_pool_3x3", "none", 
    "sep_conv_5x5", "skip_connect", "none", "sep_conv_3x3", 
    "dil_conv_3x3", "skip_connect", "max_pool_3x3", "sep_conv_5x5", 
    "none", "none"
)

arch = ArchitectureSpec(layer_widths=(32, 32, 32, 32, 32), cell_ops=ops)
darts_score = tropical_expressivity_score_darts(arch)
print(f"DARTS Proxy Expressivity Score: {darts_score:.2f}")

Running the Examples

You can run the repository's NATS-Bench and DARTS evaluations using the scripts in examples/. Real benchmark data/models are required to make empirical accuracy claims. SyntheticBenchmark is useful for software testing, but it generates accuracies from the tropical score and cannot independently validate that score.

NATS-Bench Evaluation

# Set the NATS-Bench data path
export NATS_BENCH_TSS_PATH="/path/to/NATS-tss-v1_0-3ffb9-simple"

# Run the evolutionary search simulation
python examples/run_evaluation.py

DARTS Pipeline Evaluation

# Run a 10-seed evolutionary search over the DARTS search space
python examples/run_darts_search_trajectory.py --seeds 10

NAS-Bench-301 model download

The optional model downloader now requires a SHA-256 digest checked independently against a trusted model release before installing the archive. We have not independently verified an official checksum for the linked archive and intentionally do not bundle an unverified value. Do not use a checksum copied from the same untrusted download as proof of authenticity.

# Obtain VERIFIED_SHA256 from a trusted upstream release or independent source
python examples/download_nb301_models.py --sha256 "VERIFIED_SHA256"

For an archive you have already obtained:

python examples/download_nb301_models.py \\
  --archive "/path/to/nb301_models.zip" \\
  --sha256 "VERIFIED_SHA256" \\
  --dest "/path/to/nb301_models"

The downloader rejects unsafe archive paths and unsupported file types, applies size and compression-ratio limits, and will not overwrite an existing model directory. Models are not downloaded merely by importing the script.

License

This project is licensed under the Apache 2.0 License. See the LICENSE file for details.

About

Expressivity-Guided Neural Architecture Search via Tropical Geometry

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages

0