tonylibing / inceptionv4_and_inception-resnetv2.pytorch

A PyTorch implementation of Inception-v4 and Inception-ResNet-v2.

Introduction

An inofficial PyTorch implementation of Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning

Models

• Inception-v4
• Inception-ResNet-v2

Analysis

All the results reported here are based on this repo, and 50000 ImageNet validation sets。

• top-1 accuracy
• top-5 accuracy
• # model parameters / FLOPs
• inference time (average)
• bottom10 accuracy
• Hyper parameters
• blacklists

• Top-1 and top-5 accuracy with blacklisted entities

  Model	top-1(TF)	top-1(this repo)	top-5(TF)	top-5(this repo)
  Inception-v4	0.801	0.801	0.952	0.950
  Inception-ResNet-v2	0.804	0.803	0.953	0.951

• Other hyper-parameters in Inception-v4

  eps in BatchNorm2d and count_include_pad in AvgPool2d

  Config	#top-1	top-1	#top-5	top-5
  eps=0.001, count_include_pad=False	40041	0.801	47445	0.949
  eps=0.001, count_include_pad=True	39970	0.799	47395	0.948
  eps=1e-5, count_include_pad=False	40036	0.801	47438	0.949

• Model parameters and FLOPs

  Model	Params(M)	FLOPs(G)
  Inception-v4	42.68	6.31
  Inception-ResNet-v2	55.84	6.65

• Average inference time(RTX 2080Ti)

  Model	Single inference time(ms)
  Inception-v4	40.54
  Inception-ResNet-v2	61.62

• Top-1 and top-5 accuracy(bottom-10 classes)

  • Inception-v4

    Top-1 accuracy	Classes	Top-5 accuracy	Classes
    0.16	n04152593 : screen, CRT screen	0.62	n03692522 : loupe, jeweler's loupe
    0.22	n04286575 : spotlight, spot	0.64	n04286575 : spotlight, spot
    0.22	n02123159 : tiger cat	0.64	n04525038 : velvet
    0.22	n03642806 : laptop, laptop computer	0.68	n04081281 : restaurant, eating house, eating place, eatery
    0.22	n04355933 : sunglass	0.72	n03532672 : hook, claw
    0.24	n04560804 : water jug	0.72	n03658185 : letter opener, paper knife, paperknife
    0.26	n04525038 : velvet	0.74	n03476684 : hair slide
    0.26	n02979186 : cassette player	0.74	n02910353 : buckle
    0.28	n02107908 : Appenzeller	0.76	n02776631 : bakery, bakeshop, bakehouse
    0.34	n03710637 : maillot	0.76	n03347037 : fire screen, fireguard

  • Inception-ResNet-v2

    Top-1 accuracy	Classes	Top-5 accuracy	Classes
    0.18	n04152593 : screen, CRT screen	0.6	n04286575 : spotlight, spot
    0.22	n03710637 : maillot	0.64	n04525038 : velvet
    0.22	n02123159 : tiger cat	0.64	n03692522 : loupe, jeweler's loupe
    0.28	n02979186 : cassette player	0.66	n03658185 : letter opener, paper knife, paperknife
    0.28	n04008634 : projectile, missile	0.7	n04081281 : restaurant, eating house, eating place, eatery
    0.28	n04355933 : sunglass	0.72	n03532672 : hook, claw
    0.3	n03658185 : letter opener, paper knife, paperknife	0.74	n04591157 : Windsor tie
    0.3	n03642806 : laptop, laptop computer	0.74	n03016953 : chiffonier, commode
    0.3	n04286575 : spotlight, spot	0.74	n04239074 : sliding door
    0.32	n02089973 : English foxhound	0.74	n03476684 : hair slide

Inception-Resnet-v2 Architecture

• Stem

  

• Overall schema

  

  The output of the last Inception-ResNet-C layer has no ReLU activation.

Reference

• https://github.com/tensorflow/models/tree/master/research/slim/nets
• https://github.com/tensorflow/models/tree/master/research/inception/inception/data
• https://github.com/Cadene/pretrained-models.pytorch
• https://github.com/kentsommer/keras-inceptionV4
• https://github.com/Lyken17/pytorch-OpCounter