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
| 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 | ? | |
| inception-v3 | 78.53 | 536.67 | ? | |
| 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 | |
| ResNet-34 | 72.94 | 363.11 | torch | |
| ResNet-50 | 75.76 | 439.36 | torch | |
| ResNet-101 | 77.34 | 695.78 | torch |
Time is given for Titan Black.