Skip to main content

API for NATS-Bench (a dataset for neural architecture topology and size).

Project description

NATS-Bench: Benchmarking NAS Algorithms for Architecture Topology and Size

Neural architecture search (NAS) has attracted a lot of attention and has been illustrated to bring tangible benefits in a large number of applications in the past few years. Network topology and network size have been regarded as two of the most important aspects for the performance of deep learning models and the community has spawned lots of searching algorithms for both of those aspects of the neural architectures. However, the performance gain from these searching algorithms is achieved under different search spaces and training setups. This makes the overall performance of the algorithms incomparable and the improvement from a sub-module of the searching model unclear. In this paper, we propose NATS-Bench, a unified benchmark on searching for both topology and size, for (almost) any up-to-date NAS algorithm. NATS-Bench includes the search space of 15,625 neural cell candidates for architecture topology and 32,768 for architecture size on three datasets. We analyze the validity of our benchmark in terms of various criteria and performance comparison of all candidates in the search space. We also show the versatility of NATS-Bench by benchmarking 13 recent state-of-the-art NAS algorithms on it. All logs and diagnostic information trained using the same setup for each candidate are provided. This facilitates a much larger community of researchers to focus on developing better NAS algorithms in a more comparable and computationally effective environment.

You can use pip install nats_bench to install the library of NATS-Bench.

If you are seeking how to re-create NATS-Bench from scratch or reproduce benchmarked results, please see AutoDL-Projects.

If you have questions, please ask at here or email me :)

Preparation and Download

In NATS-Bench, we (create and) use three image datasets -- CIFAR-10, CIFAR-100, and ImageNet16-120. For more details, please see Sec-3.2 in the NATS-Bench paper. To download these three datasets, please find them at Google Drive. To create the ImageNet16-120 PyTorch dataset, please call AutoDL-Projects/lib/datasets/ImageNet16, by using:

train_data = ImageNet16(root, True , train_transform, 120)
test_data  = ImageNet16(root, False, test_transform , 120)

The latest benchmark file of NATS-Bench can be downloaded from Google Drive. After download NATS-[tss/sss]-[version]-[md5sum]-simple.tar, please uncompress it by using tar xvf [file_name]. We highly recommend to put the downloaded benchmark file (NATS-sss-v1_0-50262.pickle.pbz2 / NATS-tss-v1_0-3ffb9.pickle.pbz2) or uncompressed archive (NATS-sss-v1_0-50262-simple / NATS-tss-v1_0-3ffb9-simple) into $TORCH_HOME. In this way, our api will automatically find the path for these benchmark files, which are convenient for the users. Otherwise, you need to indicate the file when creating the benchmark instance manually.

The history of benchmark files is as follows, tss indicates the topology search space and sss indicates the size search space. The benchmark file is used when creating the NATS-Bench instance with fast_mode=False. The archive is used when fast_mode=True, where archive is a directory containing 15,625 files for tss or contains 32,768 files for sss. Each file contains all the information for a specific architecture candidate. The full archive is similar to archive, while each file in full archive contains the trained weights. Since the full archive is too large, we use split -b 30G file_name file_name to split it into multiple 30G chunks. To merge the chunks into the original full archive, you can use cat file_name* > file_name.

Date benchmark file (tss) archive (tss) full archive (tss) benchmark file (sss) archive (sss) full archive (sss)
2020.08.31 NATS-tss-v1_0-3ffb9.pickle.pbz2 NATS-tss-v1_0-3ffb9-simple.tar NATS-tss-v1_0-3ffb9-full NATS-sss-v1_0-50262.pickle.pbz2 NATS-sss-v1_0-50262-simple.tar NATS-sss-v1_0-50262-full

These benchmark files (without pretrained weights) can also be downloaded from Dropbox, OneDrive or Baidu-Pan (extract code: h6pm).


1, create the benchmark instance:

from nats_bench import create
# Create the API instance for the size search space in NATS
api = create(None, 'sss', fast_mode=True, verbose=True)

# Create the API instance for the topology search space in NATS
api = create(None, 'tss', fast_mode=True, verbose=True)

2, query the performance:

# Query the loss / accuracy / time for 1234-th candidate architecture on CIFAR-10
# info is a dict, where you can easily figure out the meaning by key
info = api.get_more_info(1234, 'cifar10')

# Query the flops, params, latency. info is a dict.
info = api.get_cost_info(12, 'cifar10')

# Simulate the training of the 1224-th candidate:
validation_accuracy, latency, time_cost, current_total_time_cost = api.simulate_train_eval(1224, dataset='cifar10', hp='12')

3, others:

# Clear the parameters of the 12-th candidate.

# Reload all information of the 12-th candidate.

# Create the instance of th 12-th candidate for CIFAR-10.
# To keep NATS-Bench repo concise, we did not include any model-related codes here because they rely on PyTorch.
# The package of [models] is defined at
from models import get_cell_based_tiny_net
config = api.get_net_config(12, 'cifar10')
network = get_cell_based_tiny_net(config)

# Load the pre-trained weights: params is a dict, where the key is the seed and value is the weights.
params = api.get_net_param(12, 'cifar10', None)

Please see for more examples.

from nats_bench import api_test


If you find that NATS-Bench helps your research, please consider citing it:

  title={{NATS-Bench}: Benchmarking NAS Algorithms for Architecture Topology and Size},
  author={Dong, Xuanyi and Liu, Lu and Musial, Katarzyna and Gabrys, Bogdan},
  journal={arXiv preprint arXiv:2009.00437},

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for nats-bench, version 1.1
Filename, size File type Python version Upload date Hashes
Filename, size nats_bench-1.1-py3-none-any.whl (27.9 kB) File type Wheel Python version py3 Upload date Hashes View
Filename, size nats_bench-1.1.tar.gz (25.2 kB) File type Source Python version None Upload date Hashes View

Supported by

AWS AWS Cloud computing Datadog Datadog Monitoring DigiCert DigiCert EV certificate Facebook / Instagram Facebook / Instagram PSF Sponsor Fastly Fastly CDN Google Google Object Storage and Download Analytics Pingdom Pingdom Monitoring Salesforce Salesforce PSF Sponsor Sentry Sentry Error logging StatusPage StatusPage Status page