[ICLR 2021] HW-NAS-Bench: Hardware-Aware Neural Architecture Search Benchmark

Related tags

Deep Learning HW-NAS-Bench
Overview

HW-NAS-Bench: Hardware-Aware Neural Architecture Search Benchmark

Accepted as a spotlight paper at ICLR 2021.

Table of content

File structure

.
├── hw_nas_bench_api # HW-NAS-Bench API
│   ├── fbnet_models # FBNet's space
│   └── nas_201_models # NAS-Bench-201's space
│       ├── cell_infers
│       ├── cell_searchs
│       ├── config_utils
│       ├── shape_infers
│       └── shape_searchs
└── nas_201_api # NAS-Bench-201 API

Prerequisites

The code has the following dependencies:

  • python >= 3.6.10
  • pytorch >= 1.2.0
  • numpy >= 1.18.5

Preparation and download

No addtional file needs to be downloaded, our HW-NAS-Bench dataset has been included in this repository.

[Optional] If you want to use NAS-Bench-201 to access information about the architectures' accuracy and loss, please follow the NAS-Bench-201 guide, and download the NAS-Bench-201-v1_1-096897.pth.

How to use HW-NAS-Bench

More usage can be found in our jupyter notebook example

  1. Create an API instance from a file:
from hw_nas_bench_api import HWNASBenchAPI as HWAPI
hw_api = HWAPI("HW-NAS-Bench-v1_0.pickle", search_space="nasbench201")
  1. Show the real measured/estimated hardware-cost in different datasets:
# Example to get all the hardware metrics in the No.0,1,2 architectures under NAS-Bench-201's Space
for idx in range(3):
    for dataset in ["cifar10", "cifar100", "ImageNet16-120"]:
        HW_metrics = hw_api.query_by_index(idx, dataset)
        print("The HW_metrics (type: {}) for No.{} @ {} under NAS-Bench-201: {}".format(type(HW_metrics),

Corresponding printed information:

===> Example to get all the hardware metrics in the No.0,1,2 architectures under NAS-Bench-201's Space
The HW_metrics (type: <class 'dict'>) for No.0 @ cifar10 under NAS-Bench-201: {'edgegpu_latency': 5.807418537139893, 'edgegpu_energy': 24.226614330768584, 'raspi4_latency': 10.481976820010459, 'edgetpu_latency': 0.9571811309997429, 'pixel3_latency': 3.6058499999999998, 'eyeriss_latency': 3.645620000000001, 'eyeriss_energy': 0.6872827644999999, 'fpga_latency': 2.57296, 'fpga_energy': 18.01072}
...
  1. Show the real measured/estimated hardware-cost for a single architecture:
# Example to get use the hardware metrics in the No.0 architectures in CIFAR-10 under NAS-Bench-201's Space
print("===> Example to get use the hardware metrics in the No.0 architectures in CIFAR-10 under NAS-Bench-201's Space")
HW_metrics = hw_api.query_by_index(0, "cifar10")
for k in HW_metrics:
    if "latency" in k:
        unit = "ms"
    else:
        unit = "mJ"
    print("{}: {} ({})".format(k, HW_metrics[k], unit))

Corresponding printed information:

===> Example to get use the hardware metrics in the No.0 architectures in CIFAR-10 under NAS-Bench-201's Space
edgegpu_latency: 5.807418537139893 (ms)
edgegpu_energy: 24.226614330768584 (mJ)
raspi4_latency: 10.481976820010459 (ms)
edgetpu_latency: 0.9571811309997429 (ms)
pixel3_latency: 3.6058499999999998 (ms)
eyeriss_latency: 3.645620000000001 (ms)
eyeriss_energy: 0.6872827644999999 (mJ)
fpga_latency: 2.57296 (ms)
fpga_energy: 18.01072 (mJ)
  1. Create the network from api:
# Create the network
config = hw_api.get_net_config(0, "cifar10")
print(config)
from hw_nas_bench_api.nas_201_models import get_cell_based_tiny_net
network = get_cell_based_tiny_net(config) # create the network from configurration
print(network) # show the structure of this architecture

Corresponding printed information:

{'name': 'infer.tiny', 'C': 16, 'N': 5, 'arch_str': '|avg_pool_3x3~0|+|nor_conv_1x1~0|skip_connect~1|+|nor_conv_1x1~0|skip_connect~1|skip_connect~2|', 'num_classes': 10}
TinyNetwork(
  TinyNetwork(C=16, N=5, L=17)
  (stem): Sequential(
    (0): Conv2d(3, 16, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
    (1): BatchNorm2d(16, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
  )
  (cells): ModuleList(
    (0): InferCell(
      info :: nodes=4, inC=16, outC=16, [1<-(I0-L0) | 2<-(I0-L1,I1-L2) | 3<-(I0-L3,I1-L4,I2-L5)], |avg_pool_3x3~0|+|nor_conv_1x1~0|skip_connect~1|+|nor_conv_1x1~0|skip_connect~1|skip_connect~2|
      (layers): ModuleList(
        (0): POOLING(
          (op): AvgPool2d(kernel_size=3, stride=1, padding=1)
        )
        (1): ReLUConvBN(
...

Misc

Part of the devices used in HW-NAS-Bench:

Part of the devices used in HW-NAS-Bench

Acknowledgment

Comments
  • can i re-try your experiment again?

    can i re-try your experiment again?

    First, Congratulations to the ICLR spotlight.

    I now want to use a suitable backbone network on the Raspberry Pi.

    In your paper, it is said that the rpi 4 was measured with tensorflow lite, but it seems that only the pytorch model is provided on your github.

    Can you tell where is the tensorflow(tf lite) model for the experiment in your paper?

    thank you

    opened by cha-noong 4
  • without accurcy metrics?can`t I get the methrics throught the network config(string) instead of index?

    without accurcy metrics?can`t I get the methrics throught the network config(string) instead of index?

    HW-NAS-Bench looks like not provide the accurcy metircs? I can get the accurcy metrics in NAS-Bench-201 but I can not get the accurcy metrics in anywhere

    opened by dalin477 2
  • Is the implementation of ProxylessNAS available?

    Is the implementation of ProxylessNAS available?

    Thank you for sharing your great work!

    I'm trying to compare the results of the experiment of ProxylessNAS (Table 5) in your paper with other HW-NAS methods. So I would like to use the same hyperparameters and experimental setup as in the paper.

    On the OpenReview website, AnonReviewer3 said,

    It looks like the source code includes -- among other things -- (i) a reproduction of the FBNet search space, (ii) code for obtaining on-device benchmark numbers, (iii) a simple API for querying the inference times and energy usages of FBNet and NASBench-201 architectures, and (iv) implementations of the ProxylessNAS and FBNet search algorithms.

    but I couldn't find the implementation of ProxylessNAS mentioned here. Could you tell me where the implementation is, or what hyperparameters / experimental setup you used are?

    opened by mzsrkeen10 2
  • HLS estimates generation?

    HLS estimates generation?

    Hi, I wonder if you have plan to release the C code for Vivado HLS. Based on the paper, it seems like you have 9 blocks, and you ran C-synthesis in Vivado HLS for each block and get the latency/power estimates. Thanks in advance.

    opened by dj-park 1
  • About FBNet Search Space Measure/Estimate

    About FBNet Search Space Measure/Estimate

    I'm very interested in your work. And I'm confused about summing up the hardware-cost of all unique blocks of the architecture as its final hardware-cost. How do you deal with hardware-cost changes caused by layer to layer fusion?

    opened by yandan-sunny 1
  • About FBNet Search Space

    About FBNet Search Space

    Congratulations on the acceptance of ICLR!

    I'm confused about your dataset on FBNet search space.

    In your paper, you mentioned that "HW-NAS-Bench provides measured/estimated hardware-cost on six hardware devices for all the 10^21 architectures of FBNet search space".

    If measuring one architecture taking 0.1 seconds, then measuring all 10^21 architectures will take 10^21 * 0.1 / (606024*365)=3.1709792e+12 years. If you could estimate 10^9 architectures per second, then measuring all 10^21 architectures will take 10^12 seconds, roughly 30000 years.

    So, I'm really curious about how to perform such a big estimation task across such big search space.

    opened by Bowenwu1 1
Owner
Efficient and Intelligent Computing Lab
null
GitHub
NAS-HPO-Bench-II is the first benchmark dataset for joint optimization of CNN and training HPs.

NAS-HPO-Bench-II API Overview NAS-HPO-Bench-II is the first benchmark dataset for joint optimization of CNN and training HPs. It helps a fair and low-

yoichi hirose 8 Nov 21, 2022
NAS Benchmark in "Prioritized Architecture Sampling with Monto-Carlo Tree Search", CVPR2021

NAS-Bench-Macro This repository includes the benchmark and code for NAS-Bench-Macro in paper "Prioritized Architecture Sampling with Monto-Carlo Tree

null 35 Jan 3, 2023
CM-NAS: Cross-Modality Neural Architecture Search for Visible-Infrared Person Re-Identification (ICCV2021)

CM-NAS Official Pytorch code of paper CM-NAS: Cross-Modality Neural Architecture Search for Visible-Infrared Person Re-Identification in ICCV2021. Vis

JDAI-CV 40 Nov 25, 2022
Naszilla is a Python library for neural architecture search (NAS)

A repository to compare many popular NAS algorithms seamlessly across three popular benchmarks (NASBench 101, 201, and 301). You can implement your ow

null 270 Jan 3, 2023
[ICLR 2021] "Neural Architecture Search on ImageNet in Four GPU Hours: A Theoretically Inspired Perspective" by Wuyang Chen, Xinyu Gong, Zhangyang Wang

Neural Architecture Search on ImageNet in Four GPU Hours: A Theoretically Inspired Perspective [PDF] Wuyang Chen, Xinyu Gong, Zhangyang Wang In ICLR 2

VITA 156 Nov 28, 2022
Official PyTorch implementation of "Rapid Neural Architecture Search by Learning to Generate Graphs from Datasets" (ICLR 2021)

Rapid Neural Architecture Search by Learning to Generate Graphs from Datasets This is the official PyTorch implementation for the paper Rapid Neural A

null 48 Dec 26, 2022
Tgbox-bench - Simple TGBOX upload speed benchmark

TGBOX Benchmark This script will benchmark upload speed to TGBOX storage. Build

Non 1 Jan 9, 2022
code for paper "Does Unsupervised Architecture Representation Learning Help Neural Architecture Search?"

Does Unsupervised Architecture Representation Learning Help Neural Architecture Search? Code for paper: Does Unsupervised Architecture Representation

null 39 Dec 17, 2022
[CVPR 2021] 'Searching by Generating: Flexible and Efficient One-Shot NAS with Architecture Generator'

[CVPR2021] Searching by Generating: Flexible and Efficient One-Shot NAS with Architecture Generator Overview This is the entire codebase for the paper

null 35 Dec 1, 2022
Code release to accompany paper "Geometry-Aware Gradient Algorithms for Neural Architecture Search."

Geometry-Aware Gradient Algorithms for Neural Architecture Search This repository contains the code required to run the experiments for the DARTS sear

null 18 May 27, 2022
Densely Connected Search Space for More Flexible Neural Architecture Search (CVPR2020)

DenseNAS The code of the CVPR2020 paper Densely Connected Search Space for More Flexible Neural Architecture Search. Neural architecture search (NAS)

Jamin Fong 291 Nov 18, 2022
A Research-oriented Federated Learning Library and Benchmark Platform for Graph Neural Networks. Accepted to ICLR'2021 - DPML and MLSys'21 - GNNSys workshops.

FedGraphNN: A Federated Learning System and Benchmark for Graph Neural Networks A Research-oriented Federated Learning Library and Benchmark Platform

FedML-AI 175 Dec 1, 2022
[ICLR2021oral] Rethinking Architecture Selection in Differentiable NAS

DARTS-PT Code accompanying the paper ICLR'2021: Rethinking Architecture Selection in Differentiable NAS Ruochen Wang, Minhao Cheng, Xiangning Chen, Xi

Ruochen Wang 86 Dec 27, 2022
DeepHyper: Scalable Asynchronous Neural Architecture and Hyperparameter Search for Deep Neural Networks

What is DeepHyper? DeepHyper is a software package that uses learning, optimization, and parallel computing to automate the design and development of

DeepHyper Team 214 Jan 8, 2023
Model search is a framework that implements AutoML algorithms for model architecture search at scale

Model search (MS) is a framework that implements AutoML algorithms for model architecture search at scale. It aims to help researchers speed up their exploration process for finding the right model architecture for their classification problems (i.e., DNNs with different types of layers).

Google 3.2k Dec 31, 2022
An implementation for Neural Architecture Search with Random Labels (CVPR 2021 poster) on Pytorch.

Neural Architecture Search with Random Labels(RLNAS) Introduction This project provides an implementation for Neural Architecture Search with Random L

null 18 Nov 8, 2022
Based on the paper "Geometry-aware Instance-reweighted Adversarial Training" ICLR 2021 oral

Geometry-aware Instance-reweighted Adversarial Training This repository provides codes for Geometry-aware Instance-reweighted Adversarial Training (ht

Jingfeng 47 Dec 22, 2022
An implementation demo of the ICLR 2021 paper Neural Attention Distillation: Erasing Backdoor Triggers from Deep Neural Networks in PyTorch.

Neural Attention Distillation This is an implementation demo of the ICLR 2021 paper Neural Attention Distillation: Erasing Backdoor Triggers from Deep

Yige-Li 84 Jan 4, 2023
Seach Losses of our paper 'Loss Function Discovery for Object Detection via Convergence-Simulation Driven Search', accepted by ICLR 2021.

CSE-Autoloss Designing proper loss functions for vision tasks has been a long-standing research direction to advance the capability of existing models

Peidong Liu(刘沛东) 54 Dec 17, 2022