imagenet-validation.torch

A modified subset of https://github.com/soumith/imagenet-multiGPU.torch to allow easy testing of imagenet models.

Also used to make sure that models converted from other frameworks are correct.

This code was used in "An Analysis of Deep Neural Network Models for Practical Applications" report http://arxiv.org/abs/1605.07678

Evaluation

Model name	top-1 (central crop) accuracy	test time (s)	framework	trained by
NIN	0(?)	155.66	caffe	authors
bn-NIN	62.62	170.71	torch	me
inception-v2	51.00(?)	156.21	?	google
inception-v3	78.53	536.67	?	google
bvlc_alexnet	54.61	172.29	caffe	bvlc
bn-alexnet	56.60	170.56	torch	me
VGG_CNN_S	63.06	175.31	caffe	VGG
VGG-16	70.62	738.82	caffe	VGG
VGG-19	70.74	898.13	caffe	VGG
ResNet-18	69.31	208.53	torch	facebook
ResNet-34	72.94	363.11	torch	facebook
ResNet-50	75.76	439.36	torch	facebook
ResNet-101	77.34	695.78	torch	facebook

Time is given for Titan Black.

Networks

1. inception-v2: https://github.com/soumith/inception.torch
2. inception-v3: https://github.com/szagoruyko/inception-v3.torch
3. bvlc_alexnet: https://github.com/BVLC/caffe/tree/master/models/bvlc_alexnet
4. bn_alexnet: https://gist.github.com/szagoruyko/dd032c529048492630fc
5. bn-nin: https://gist.github.com/szagoruyko/0f5b4c5e2d2b18472854
6. VGG_CNN_S: https://gist.github.com/ksimonyan/fd8800eeb36e276cd6f9#file-readme-md
7. VGG-16: https://gist.github.com/ksimonyan/211839e770f7b538e2d8#file-readme-md
8. VGG-19: https://gist.github.com/ksimonyan/3785162f95cd2d5fee77#file-readme-md