Learning Versatile Neural Architectures by Propagating Network Codes

Related tags

Deep LearningNCP
Overview

Learning Versatile Neural Architectures by Propagating Network Codes

Mingyu Ding, Yuqi Huo, Haoyu Lu, Linjie Yang, Zhe Wang, Zhiwu Lu, Jingdong Wang, Ping Luo

diagram

Introduction

This work includes:
(1) NAS-Bench-MR, a NAS benchmark built on four challenging datasets under practical training settings for learning task-transferable architectures.
(2) An efficient predictor-based algorithm Network Coding Propagation (NCP), which back-propagates the gradients of neural predictors to directly update architecture codes along desired gradient directions for various objectives.

This framework is implemented and tested with Ubuntu/Mac OS, CUDA 9.0/10.0, Python 3, Pytorch 1.3-1.6, NVIDIA Tesla V100/CPU.

Dataset

We build our benchmark on four computer vision tasks, i.e., image classification (ImageNet), semantic segmentation (CityScapes), 3D detection (KITTI), and video recognition (HMDB51). Totally 9 different settings are included, as shown in the data/*/trainval.pkl folders.

Note that each .pkl file contains more than 2500 architectures, and their corresponding evaluation results under multiple metrics. The original training logs and checkpoints (including model weights and optimizer data) will be uploaded to Google drive (more than 4T). We will share the download link once the upload is complete.

Quick start

First, train the predictor

python3 tools/train_predictor.py  # --cfg configs/seg.yaml

Then, edit architecture based on desired gradients

python3 tools/ncp.py  # --cfg configs/seg.yaml

Examples

  • An example in NAS-Bench-MR (Seg):
{'mIoU': 70.57,
 'mAcc': 80.07,
 'aAcc': 95.29,
 'input_channel': [16, 64],
 # [num_branches, [num_convs], [num_channels]]
 'network_setting': [[1, [3], [128]],
  [2, [3, 3], [32, 48]],
  [2, [3, 3], [32, 48]],
  [2, [3, 3], [32, 48]],
  [3, [2, 3, 2], [16, 32, 16]],
  [3, [2, 3, 2], [16, 32, 16]],
  [4, [2, 4, 1, 1], [96, 112, 48, 80]]],
 'last_channel': 112,
 # [num_branches, num_block1, num_convs1, num_channels1, ..., num_block4, num_convs4, num_channels4, last_channel]
 'embedding': [16, 64, 1, 3, 128, 3, 3, 3, 32, 48, 2, 2, 3, 2, 16, 32, 16, 1, 2, 4, 1, 1, 96, 112, 48, 80]
}
  • Load Datasets:
import pickle
exps = pickle.load(open('data/seg/trainval.pkl', 'rb'))
# Then process each item in exps
  • Load Model / Get Params and Flops (based on the thop library):
import torch
from thop import profile
from models.supernet import MultiResolutionNet

# Get model using input_channel & network_setting & last_channel
model = MultiResolutionNet(input_channel=[16, 64],
                           network_setting=[[1, [3], [128]],
                            [2, [3, 3], [32, 48]],
                            [2, [3, 3], [32, 48]],
                            [2, [3, 3], [32, 48]],
                            [3, [2, 3, 2], [16, 32, 16]],
                            [3, [2, 3, 2], [16, 32, 16]],
                            [4, [2, 4, 1, 1], [96, 112, 48, 80]]],
                          last_channel=112)

# Get Flops and Parameters
input = torch.randn(1, 3, 224, 224)
macs, params = profile(model, inputs=(input, ))  

structure

Data Format

Each code in data/search_list.txt denotes an architecture. It can be load in our supernet as follows:

  • Code2Setting
params = '96_128-1_1_1_48-1_2_1_1_128_8-1_3_1_1_1_128_128_120-4_4_4_4_4_4_128_128_128_128-64'
embedding = [int(item) for item in params.replace('-', '_').split('_')]

embedding = [ 96, 128,   1,   1,  48,   1,   1,   1, 128,   8,   1,   1,
           1,   1, 128, 128, 120,   4,   4,   4,   4,   4, 128, 128,
         128, 128, 64]
input_channels = embedding[0:2]
block_1 = embedding[2:3] + [1] + embedding[3:5]
block_2 = embedding[5:6] + [2] + embedding[6:10]
block_3 = embedding[10:11] + [3] + embedding[11:17]
block_4 = embedding[17:18] + [4] + embedding[18:26]
last_channels = embedding[26:27]
network_setting = []
for item in [block_1, block_2, block_3, block_4]:
    for _ in range(item[0]):
        network_setting.append([item[1], item[2:-int(len(item) / 2 - 1)], item[-int(len(item) / 2 - 1):]])

# network_setting = [[1, [1], [48]], 
#  [2, [1, 1], [128, 8]],
#  [3, [1, 1, 1], [128, 128, 120]], 
#  [4, [4, 4, 4, 4], [128, 128, 128, 128]], 
#  [4, [4, 4, 4, 4], [128, 128, 128, 128]], 
#  [4, [4, 4, 4, 4], [128, 128, 128, 128]], 
#  [4, [4, 4, 4, 4], [128, 128, 128, 128]]]
# input_channels = [96, 128]
# last_channels = [64]
  • Setting2Code
input_channels = [str(item) for item in input_channels]
block_1 = [str(item) for item in block_1]
block_2 = [str(item) for item in block_2]
block_3 = [str(item) for item in block_3]
block_4 = [str(item) for item in block_4]
last_channels = [str(item) for item in last_channels]

params = [input_channels, block_1, block_2, block_3, block_4, last_channels]
params = ['_'.join(item) for item in params]
params = '-'.join(params)
# params
# 96_128-1_1_1_48-1_2_1_1_128_8-1_3_1_1_1_128_128_120-4_4_4_4_4_4_128_128_128_128-64'

License

For academic use, this project is licensed under the 2-clause BSD License. For commercial use, please contact the author.

Owner
Mingyu Ding
Mingyu Ding
Implementation of Nyström Self-attention, from the paper Nyströmformer

Nyström Attention Implementation of Nyström Self-attention, from the paper Nyströmformer. Yannic Kilcher video Install $ pip install nystrom-attention

Phil Wang 95 Jan 02, 2023
Simulation of the solar system using various nummerical methods

solar-system Simulation of the solar system using various nummerical methods Download the repo Make shure matplotlib, scipy etc. are installed execute

Caspar 7 Jul 15, 2022
Download files from DSpace systems (because for some reason DSpace won't let you)

DSpaceDL A tool for downloading files from DSpace items. For some reason, DSpace systems have a dogshit UI, and Universities absolutely LOOOVE to use

Soumitra Shewale 5 Dec 01, 2022
Learning Efficient Online 3D Bin Packing on Packing Configuration Trees

Learning Efficient Online 3D Bin Packing on Packing Configuration Trees This repository is being continuously updated, please stay tuned! Any code con

86 Dec 28, 2022
Official Repository for the ICCV 2021 paper "PixelSynth: Generating a 3D-Consistent Experience from a Single Image"

PixelSynth: Generating a 3D-Consistent Experience from a Single Image (ICCV 2021) Chris Rockwell, David F. Fouhey, and Justin Johnson [Project Website

Chris Rockwell 95 Nov 22, 2022
Simple image captioning model - CLIP prefix captioning.

CLIP prefix captioning. Inference Notebook: 🥳 New: 🥳 Our technical papar is finally out! Official implementation for the paper "ClipCap: CLIP Prefix

688 Jan 04, 2023
Example of semantic segmentation in Keras

keras-semantic-segmentation-example Example of semantic segmentation in Keras Single class example: Generated data: random ellipse with random color o

53 Mar 23, 2022
This code is part of the reproducibility package for the SANER 2022 paper "Generating Clarifying Questions for Query Refinement in Source Code Search".

Clarifying Questions for Query Refinement in Source Code Search This code is part of the reproducibility package for the SANER 2022 paper "Generating

Zachary Eberhart 0 Dec 04, 2021
project page for VinVL

VinVL: Revisiting Visual Representations in Vision-Language Models Updates 02/28/2021: Project page built. Introduction This repository is the project

308 Jan 09, 2023
An official source code for "Augmentation-Free Self-Supervised Learning on Graphs"

Augmentation-Free Self-Supervised Learning on Graphs An official source code for Augmentation-Free Self-Supervised Learning on Graphs paper, accepted

Namkyeong Lee 59 Dec 01, 2022
List of content farm sites like g.penzai.com.

内容农场网站清单 Google 中文搜索结果包含了相当一部分的内容农场式条目,比如「小 X 知识网」「小 X 百科网」。此种链接常会 302 重定向其主站,页面内容为自动生成,大量堆叠关键字,揉杂一些爬取到的内容,完全不具可读性和参考价值。 尤为过分的是,该类网站可能有成千上万个分身域名被 Goog

WDMPA 541 Jan 03, 2023
Motion and Shape Capture from Sparse Markers

MoSh++ This repository contains the official chumpy implementation of mocap body solver used for AMASS: AMASS: Archive of Motion Capture as Surface Sh

Nima Ghorbani 135 Dec 23, 2022
The source code for the Cutoff data augmentation approach proposed in this paper: "A Simple but Tough-to-Beat Data Augmentation Approach for Natural Language Understanding and Generation".

Cutoff: A Simple Data Augmentation Approach for Natural Language This repository contains source code necessary to reproduce the results presented in

Dinghan Shen 49 Dec 22, 2022
DSEE: Dually Sparsity-embedded Efficient Tuning of Pre-trained Language Models

DSEE Codes for [Preprint] DSEE: Dually Sparsity-embedded Efficient Tuning of Pre-trained Language Models Xuxi Chen, Tianlong Chen, Yu Cheng, Weizhu Ch

VITA 4 Dec 27, 2021
[Preprint] "Chasing Sparsity in Vision Transformers: An End-to-End Exploration" by Tianlong Chen, Yu Cheng, Zhe Gan, Lu Yuan, Lei Zhang, Zhangyang Wang

Chasing Sparsity in Vision Transformers: An End-to-End Exploration Codes for [Preprint] Chasing Sparsity in Vision Transformers: An End-to-End Explora

VITA 64 Dec 08, 2022
Infrastructure as Code (IaC) for a self-hosted version of Gnosis Safe on AWS

Welcome to Yearn Gnosis Safe! Setting up your local environment Infrastructure Deploying Gnosis Safe Prerequisites 1. Create infrastructure for secret

Numan 16 Jul 18, 2022
PyVideoAI: Action Recognition Framework

This reposity contains official implementation of: Capturing Temporal Information in a Single Frame: Channel Sampling Strategies for Action Recognitio

Kiyoon Kim 22 Dec 29, 2022
This is the replication package for paper submission: Towards Training Reproducible Deep Learning Models.

This is the replication package for paper submission: Towards Training Reproducible Deep Learning Models.

0 Feb 02, 2022
PyTorch code for MART: Memory-Augmented Recurrent Transformer for Coherent Video Paragraph Captioning

MART: Memory-Augmented Recurrent Transformer for Coherent Video Paragraph Captioning PyTorch code for our ACL 2020 paper "MART: Memory-Augmented Recur

Jie Lei 雷杰 151 Jan 06, 2023