Awesome

TF-resnet

This is a Tensorflow implementation of ResNet, compatible with Tensorflow 1.2.1.

Currently it only supports testing the 101 layer model by converting the caffemodel provided by Kaiming. Although supporting other ResNet variants and training should be quick and easy.

The caffemodel2npy.py is modified from here, and the resnet_model.py is modified from here.

Example Usage

• Download the prototxt and caffemodel provided by Kaiming
• Convert caffemodel to npy file

python caffemodel2npy.py ../ResNet/ResNet-101-deploy.prototxt ../ResNet/ResNet-101-model.caffemodel ./model/ResNet101.npy

• Convert npy file to tfmodel

python npy2tfmodel.py 0 ./model/ResNet101.npy ./model/ResNet101_init.tfmodel

• Test on a single image

python resnet_main.py 0 single

• To test on the ILSVRC12 validation set, run

python resnet_main.py 0 imagenet

Performance

The converted ResNet 101 model achieves top 5 error of 7.48% and top 1 error of 24.58% on ILSVRC12 validation set. This is without any cropping/flipping/multi-scale, using only the original image.

TODO

• Support ResNet 50 and 152
• Training code