kisstea / ResNeXt

Reproduce ResNet-v3(Aggregated Residual Transformations for Deep Neural Network) with Caffe

ResNeXt

Reproduce ResNet-v3(Aggregated Residual Transformations for Deep Neural Network)

Results

0. 1-crop validation error on ImageNet (center 224x224 crop from resized image with shorter side=256):

   model	top-1	top-5
   VGG-16	28.5%	9.9%
   ResNet-50	24.7%	7.8%
   ResNeXt-50	23.1%	6.7%
   ResNet-101	23.6%	7.1%
   ResNeXt-101(32x4d)	22.1%	5.8%
   ResNet-152	23.0%	6.7%
   ResNeXt-152	21.3%	5.6%
   ResNeXt-101(64x4d)	20.8%	5.5%

1. 1-crop validation error on ImageNet (center 320x320 crop from resized image with shorter side=320):

Network	crop-size	top-1	top-5
ResNet-152-v1	320x320	21.3%	5.5%
ResNet-152-v2	320x320	21.1%	5.5%
ResNet-200-v1	320x320	21.8%	6.0%
ResNet-200-v2	320x320	20.7%	5.3%
ResNeXt-50	320x320	21.9%	5.9%
ResNeXt-101(32x4d)	320x320	20.2%	4.9%
ResNeXt-152	320x320	19.9%	4.8%
ResNeXt-101(64x4d)	320x320	19.5%	4.8%

Training Curve:

Model files:

ResNeXt-50 OneDrive download: link

ResNeXt-101 OneDrive download: link

ResNeXt-152 OneDrive download: link