ctnas

CT-NAS

This module contains an API access to a Neural Architecture Search dataset on a search space called "computational themes". Such themes are small directed acyclic graphs (DAGs) with a vertex collapse condition. This condition makes the left and the right graph in the picture being equivalent / isomorphic.

The dataset contains computations of feed-forward neural networks with different hidden structural priors based on these themes. By this, you can search for connections between graph theoretic properties of the search space and computational properties of the neural network models.

Installation

Via poetry (recommended for projects) using PyPi: poetry add ctnas

Directly with pip from PyPi:

pip install ctnas

Via conda in your environment.yml (recommended for reproducible experiments):

name: exp01
channels:
- defaults
dependencies:
- pip>=20
- pip:
    - ctnas

From public GitHub:

pip install --upgrade git+ssh://[e-mail]:innvariant/ctnas.git

Usage examples

from ctnas.api import CTNASApi

api = CTNASApi()
print(api.get_datasets())
# Should give you:
# ['spheres-b8c16fd7', 'mnist', 'spheres-23aeba4d', 'spheres-bee36cd9',
#  'spheres-b758e9f4', 'spheres-0a19afe4', 'cifar10', 'spheres-6598864b']

from ctnas.api import CTNASApi

api = CTNASApi()
print(api.get_graph_properties().head())

Gives you s.th. like:

test_dev.py . graph_uuid num_nodes ... degree_var undir_ecc_var 0 6e302aa7-6208-42a9-b1e0-08ce6d9d83ba 6 ... 1.222222 0.222222 1 ecd9c934-90ae-460c-855f-90c0b24a4150 6 ... 0.666667 0.000000 2 d111e38f-3ed1-454f-9d0e-8ded0428c9d9 6 ... 1.000000 0.222222 3 d23cac47-047c-4ec6-aaa4-e393b2ebeccd 5 ... 0.640000 0.240000 4 c56bb6f8-a9ec-44db-8c17-37b166fb5b06 6 ... 0.888889 0.222222

[5 rows x 19 columns]

import networkx as nx
import matplotlib.pyplot as plt
from ctnas.api import CTNASApi

api = CTNASApi()
graph = api.get_graph("0a1ded7d-677a-41f7-9361-c7079c8a34a7")
nx.draw(graph)
plt.show()

Cite our work

@misc{stier2022ctnas,
    title={CT-NAS: Analysis of Hidden Structural Priors for Neural Architecture Search},
    author={Julian Stier and Michael Granitzer},
    year={2022}
}

MinIO Policy

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Action": [
        "s3:GetObject"
      ],
      "Effect": "Allow",
      "Resource": [
        "arn:aws:s3:::homes/stier/ctnas/*"
      ]
    }
  ]
}

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Action": [
        "s3:GetObject"
      ],
      "Effect": "Allow",
      "Principal": {
        "AWS": [
          "*"
        ]
      },
      "Resource": [
        "arn:aws:s3:::homes/stier/ctnas/*"
      ],
      "Sid": ""
    }
  ]
}