Image invert : 'Understanding Deep Image Representations by Inverting Them'
A simple implementation of 'Understanding Deep Image Representations by Inverting Them' , to see what it does!
- Omitted 'natural image prior of training dataset' ( means that did not used 'sigma' in the paper )
- Only L2 and Total Variation losses are used.
- Roughly set TV loss ratio to 5e-7
Download vgg19 pre-trained : http://www.vlfeat.org/matconvnet/models/beta16/imagenet-vgg-verydeep-19.mat
vgg.py is borrowed from 'https://github.com/anishathalye/neural-style'
Test results from 'conv5_1' :
< Input Image >
< L2 Loss only >
< L2 and TV loss >
