search

lltcggie / waifu2x-caffe

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

Waifu2x-caffe uses too much ram #224

Open mikebilly opened 2 years ago

mikebilly commented 2 years ago

I tried to upscale a 2720x1920 picture with these output settings -p cudnn -m noise_scale -n 3 -s 8 -t 1 -i using waifu2x-caffe on Google Colab but I got this output:

tcmalloc: large alloc 4010811392 bytes == 0x55db15804000 @  0x7f28ceaf11e7 0x7f28cc497382 0x7f28cc59621b 0x55da05ce6afb 0x55da05ce688c 0x55da05ce51cf 0x55da05d1ef73 0x55da05d1eb47 0x55da05d1e016 0x55da05d1dbf5 0x55da05d1caf7 0x55da05cc5e2e 0x7f28cb75abf7 0x55da05cc13ca
tcmalloc: large alloc 4010811392 bytes == 0x55dc097ba000 @  0x7f28ceaf11e7 0x7f28cc497382 0x7f28cc59621b 0x55da05ce6afb 0x55da05ce688c 0x55da05ce51cf 0x55da05d1ef73 0x55da05d1eb47 0x55da05d1e016 0x55da05d1dbf5 0x55da05d1caf7 0x55da05cc5e2e 0x7f28cb75abf7 0x55da05cc13ca
tcmalloc: large alloc 4010811392 bytes == 0x55dcf88bc000 @  0x7f28ceaf11e7 0x7f28cc497382 0x7f28cc59621b 0x7f28cc5b46e5 0x7f28cc5b6113 0x55da05d1e799 0x55da05d1e8b6 0x55da05d1ec07 0x55da05d1e016 0x55da05d1dbf5 0x55da05d1caf7 0x55da05cc5e2e 0x7f28cb75abf7 0x55da05cc13ca

and it didn't create output picture.

However, waifu2x-ncnn-vulkan successfully upscaled that same input picture, with those same output settings (except for using GPU instead of cuDNN). While waifu2x-ncnn-vulkan only used 2.95gb ram and took 1 hour and 10 minutes to upscale, waifu2x-caffe used more than 12.69gb ram (exceeding the allowed ram usage) and after 40 minutes, it failed.

image

(Right before waifu2x-caffe stopped processing and failed)