weird results with pretrained models

[ahmedsalaheldin]

I can't reproduce the results even with the pretrained models. I get similar results for all data sets on gpu and cpu. A screen shot of the results below.

[junyanz]

See this two posts 1 and 2 for solutions. Also check out our PyTorch version, which is more stable and still under active development.

[ahmedsalaheldin]

reinstalling torch didn't make a difference for me. I'll try the Pytorch version, thanks.

[happsky]

I also meet the same problem, @ahmedsalaheldin did you solve it?

[junyanz]

@happsky and @ahmedsalaheldin Could you report your CUDA and Torch version?

[ahmedsalaheldin]

@happsky no I moved to the pytorch implementation as suggested and that works fine for me. @junyanz Cuda 9.0 . I'm not aware that torch keeps a version system. edit: torch commit: commit 0219027e6c4644a0ba5c5bf137c989a0a8c9e01b

[happsky]

@junyanz Cuda 8.0 and torch version is commit 0219027e6c4644a0ba5c5bf137c989a0a8c9e01b. @ahmedsalaheldin did you find any pretrained pytorch model for style tranfer?

[taesungp]

Hello, I did a fresh install of the newest torch version 0219027e6c4644a0ba5c5bf137c989a0a8c9e01b, and the running the pretrained model produced fine results.

I have CUDA 9.0 and libcudnn.so.7.1.4. There were a few glitches when I installed Torch. I had to install with export TORCH_NVCC_FLAGS="-D__CUDA_NO_HALF_OPERATORS__" that disabled half-precision operations. I also followed this link to run Torch with cudnn v7.

Unfortunately, I could not replicate the problem... I would appreciate a lot if someone who had the same problem could share a fix.

[happsky]

@taesung89 what's you Lua version, 5.1 or 5.2?

[ahmedsalaheldin]

@happsky I trained the pytorch model myself. training the torch model gave me weird results too which is why I tried the pretrained models P.S I tried the torch version with all lua versions

[happsky]

@ahmedsalaheldin may you share the pytorch pre-trained models?