I tried AnimeGAN, which converts live-action film into anime style.
Clone AnimeGAN.
Create an environment for animegan.
$ conda create -n animegan python=3.6
$ conda activate animegan
$ pip install tensorflow-gpu==1.8.0
$ pip install tqdm
$ pip install scipy
$ pip install opencv-python
$ cd AnimeGAN-master
Place dataset on AnimeGAN-master.
Copy the contents of Haoyao-style to checkpoint \ AnimeGAN_Hayao_lsgan_300_300_1_3_10
Install CUDA 9.0.
Copy and paste C: \ Program Files \ NVIDIA GPU Computing Toolkit \ CUDA \ v9.0 \ bin \ cudart64_90.dll into AnimeGAN-master.
If you do the following
$ python test.py --checkpoint_dir checkpoint/AnimeGAN_Hayao_lsgan_300_300_1_3_10 --test_dir dataset/test/real --style_name H
I got an error.
return _bootstrap._gcd_import(name[level:], package, level)
ImportError: DLL load failed: The specified module cannot be found.
vc_redist.x64.exe by referring to Installing tensorflow on windows10 (for specific errors) -jp / help / 2977003 / the-latest-supported-visual-c-downloads) will be installed.
~~ Put vgg19.npy in vgg19_weight ~~
When I ran it, I got an error again.
tensorflow.python.framework.errors_impl.InternalError: Blas SGEMM launch failed : m=36864, n=128, k=64
Specify GPU1 by referring to What to do if "Blas GEMM launch failed" appears in TensorFlow. I needed to.
Change os.environ ["CUDA_VISIBLE_DEVICES"] = "0" to "1" on the 9th line of test.py and execute.
======================End of Report==========================
FLOPs: 7937325
100%|██████████████████████████████████████████████████████████████████████████████████| 22/22 [01:24<00:00, 3.84s/it]
test-time: 84.46189665794373 s
one image test time : 3.839177120815624 s
It took 3.84 seconds per sheet!
before There is an input image in the dataset / test / real folder.
after The output is in results \ H.
Thank you for your hard work.
Recommended Posts