Shaofan Lai

Same issue here. Even cannot compile the graph in 1 hour

You might mix up the image_channel_axis. Theano and TF use different data format.

I implemented the code based on the [paper](https://arxiv.org/pdf/1703.10593.pdf). The only hyper-parameter in the Equ.3 is lambda. Feel free to change the code to include more tricks from the official repository.

Indeed. There is a difference between TF and Theano in dimension. I think I used `"image_data_format": "channel_last"` in `$HOME/.keras/keras.json` when I wrote this repository. Since they stop to update Theano...