search

imatge-upc / salgan

SalGAN: Visual Saliency Prediction with Generative Adversarial Networks
https://imatge-upc.github.io/salgan
MIT License
371 stars 105 forks source link

Unable to reproduce the standard saliency map. #30

Open YenchungChen opened 6 years ago

YenchungChen commented 6 years ago

Hi @junting,

I'm super interesting in SalGAN, so here I am.

I've tried to run the model with provided VGG-16 weights, and pertained generator and discriminator which are provided in your Github repo, really appreciate your work.

While running (03-predict.py) with the setting above, following is the result I've received:

Use this picture as the input of SalGAN i112_original

I get the result i112

which is slightly different from the result I expected according to the website i1123

​I'm wondering if there's some mistake with my operation, or shall I do some training on the provided weights in order to get the expected results ? Hope you could give me some assistance. Nice work for the SalGAN again.

Cheers,

pascalxia commented 6 years ago

I obtained the same result as @YenchungChen . I made a Docker image for people to replicate it easily. The Docker image is blindgrandpa/theano_cudnn5.1. It uses CUDA 8.0 + cuDNN 5.1.10 + Theano 0.9.0 + Lasagne 0.2.dev1.

iriszxcy commented 5 years ago

I got the same result as @YenchungChen and @pascalxia, any ideas? i112sa