search

lltcggie / waifu2x-caffe

waifu2xのCaffe版
MIT License
8.01k stars 842 forks source link

"possibility of split size is too large" error occuring in version 1.2.0.3, but not in version 1.2.0.2 #202

Open violisto opened 3 years ago

violisto commented 3 years ago

Both versions of the program are using the same settings, but ver 1.2.0.3 always comes up with this error. Images below are what happens after attempting to start the conversion.

waifu2x-caffe 1 2 0 2 successful run waifu2x-caffe 1 2 0 3 error error_log_20200722-004546.4896.txt

seiya-git commented 3 years ago

its bug with cuda 10.2, use cudann for cuda 10.1

download dll for cuda 10.1 and replace dll https://developer.nvidia.com/compute/machine-learning/cudnn/secure/7.6.5.32/Production/10.1_20191031/cudnn-10.1-windows10-x64-v7.6.5.32.zip

https://developer.nvidia.com/rdp/cudnn-download

i0ntempest commented 3 years ago

+1 this error. Changed the cudnn DLL and all fine.

lltcggie commented 3 years ago

ver 1.2.0.4でCUDA ToolkitとcuDNNを更新したので直ったと思われます。

i0ntempest commented 3 years ago

Doesn't seem to work for me here, using split size 64 and batch size 1, with a GTX 1650 eGPU. Changing DLL now does nothing. However it does work on a notebook 1060 with the same settings. Is 4GB of VRAM simply too small? Also: it errors out only when I scale to 3x or above, not if I do 2x for twice.

i0ntempest commented 3 years ago

Figured it out: I'm running the 1650 as a secondary GPU, and my main GPU, a 5700 XT, is conflicting with it. If I disable my 5700 XT then all is fine. Don't know if it's a problem in cuDNN or the software itself.

tomchen commented 3 years ago

Just change "App Setting" -> "Use Processor" to "CPU" instead of "CUDA (uses cuDNN)"

(The older version of cuda dll that seiya-git suggested does not work for me)

AmoArt commented 3 years ago

Oh, I've run into this problem myself and it's pretty disappointing to see it still not solved in any shape or form. While running the program on CPU works, it's pretty slow.

ccrowles commented 3 years ago

Same issue, and replacing with the dll for cuda 10.1 does not work, it still tries to search for the 64_8 dll (which is 10.2) instead of the 64_7 (10.1) that I put in the folder.

image

ccrowles commented 2 years ago

Different pc, same issue as before.

erickyun commented 2 years ago

This problem occurs with mine laptop too but only with 2-D illust (CUnet Model) with 2-D illust (UpRGB Model) there is no problem?

EDİT: With mine CPU 2-D illust (CUnet Model) works but it does not work with mine GPU?

CPU:Intel(R) Core(TM) i5-10300H CPU @ 2.50GHz 2.50 GHz GPU:Intel (R) UHD Graphics GPU: NVIDIA GeForce GTX 1650 Windows10

i0ntempest commented 2 years ago

Yes, CUnet is known to use more memory. Same behavior here.

Horizuki commented 1 year ago

Problem solved, if anyone still interesting and use waifu2x-caffe (ver 1.2.0.4) still get this problem. Just download newest 10.2 dll (for now it is cuDNN v8.4.1 (May 27th, 2022), for CUDA 10.2) and replace it with the older one, which should fix the error. After that, it work pretty well image

i0ntempest commented 1 year ago

The solution above breaks the "Check cuDNN" button/function but upscaling works nonetheless. Tried with v8.5.0.96 dlls. image