Skip to content

SE_ResNet && SE_ResNeXt with pretrained weights on ImageNet (SENet In TensorFlow)

License

Notifications You must be signed in to change notification settings

HiKapok/TF-SENet

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

7 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

SE_ResNet and SE_ResNeXt model for Tensorflow with pre-trained weights on ImageNet

This repository contains code of the un-official re-implement of SE_ResNe?t50 and SE_ResNe?t101 model. Here is the authors' implementation in Caffe.

SENet is one state-of-the-art convolutional neural network architecture, where dynamic channelwise feature recalibration have been introduced to improve the representational capacity of CNN. More details can be found in the original paper: Squeeze-and-Excitation Networks.

In order to accelerate further research based on SE_ResNet and SE_ResNeXt, I would like to share the pre-trained weights on ImageNet for them, you can download from Google Drive. The pre-trained weights are converted from official weights in Caffe using MMdnn with other post-processing. And the outputs of all the network using the converted weights has almost the same outputs as original Caffe network (errors<1e-5). All rights related to the pre-trained weights belongs to the original author of SENet.

This code and the pre-trained weights only can be used for research purposes.

The canonical input image size for this SE_ResNe?t is 224x224, each pixel value should in range [-128,128](BGR order), and the input preprocessing routine is quite simple, only normalization through mean channel subtraction was used. According to the official open-source version in Caffe, SE-ResNe?t models got to 22.37%(SE-ResNet-50) and 20.97%(SE-ResNeXt-50) on ImageNet-1k for the single crop top-1 validation error.

The codes was tested under Tensorflow 1.6, Python 3.5, Ubuntu 16.04.

BTW, other scaffold need to be build for training from scratch. You can refer to resnet/imagenet_main for adding weight decay to the loss manually.

Apache License 2.0

About

SE_ResNet && SE_ResNeXt with pretrained weights on ImageNet (SENet In TensorFlow)

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages