Result of cezanne_pretrained looking like heatmaps

[totopia]

Ran commands:

bash ./datasets/download_dataset.sh ae_photos
bash ./pretrained_models/download_model.sh style_cezanne
DATA_ROOT=./datasets/ae_photos name=style_cezanne_pretrained model=one_direction_test phase=test loadSize=256 fineSize=256 resize_or_crop="scale_width" th test.lua

Result:

Is it expected?

[junyanz]

It is not expected. It is caused by Torch version. I resolved the issue by removing the old torch and installing a new torch from the scratch following this official instruction. See #57 for more discussion.

[totopia]

Thanks for the reply. seems like I may have more infra problems to solve before I can get Torch re-installed. Could you let me know your output for cd ~torch; git log -n 1? This should display the "version" of torch.

Somehow the very edge commit has the openblas_utest taking forever on installl on Ubuntu 16.04, but it may be my own env issues.

[junyanz]

Here is my torch version: commit 20e523771377568841e9d11bc3275a4404edcd1e

[totopia]

@junyanz Thanks for the info. I tried the commit and it doesn't seem to be specific to torch or app code. It might be a deeper issure in gcc or nvcc reference.

Do you know which Cuda version you were using at the time of training the model?

I actually managed to fix the problem. Current install-deps sometimes stuck at a openblas install step, so I had to manually install openblas. I don't know if the current project deletes previous results dir, but a deletion of corrupt result directory is also required to fix the problem.

[junyanz]

Cool. I am using cuda 8.0. I always recommend a complete re-installation. I am never able to fix a broken Torch.