VGG-16 is my favorite image classification model to run because of its simplicity and accuracy. The creators of this model published a pre-trained binary that can be used in Caffe.
https://gist.github.com/ksimonyan/211839e770f7b538e2d8#file-readme-md
MD5 (VGG_ILSVRC_16_layers.caffemodel) = 441315b0ff6932dbfde97731be7ca852
This is to convert that specific file to a TensorFlow model and check its correctness.
Run make
to download the original caffe model and convert it.
tf_forward.py
has an example of how to use the generated vgg16.tfmodel
file.
If you don't feel like installing caffe, you can download the output here https://github.com/ry/tensorflow-vgg16/raw/master/vgg16-20160129.tfmodel.torrent
The input ("images") to the TF model is expected to be [batch, height, width, channel] where height = width = 224 and channel = 3. Values should be between 0 and 1.
The output ("prob") is a 1000 dimensional class probabiity vector whose indexes correspond to line numbers in syset.txt.