Skip to content

Latest commit

 

History

History
21 lines (16 loc) · 840 Bytes

readme.md

File metadata and controls

21 lines (16 loc) · 840 Bytes
Raw

cnn: convolutional neural network denoise

this reads the weights of gmic's resnet from a texture and evaluates the network.

the current implementation runs in small tiles (24x24) and has only been tested on a slow nvidia 1650 GTX. a future implementation should probably focus on the turing architecture or more recent ones.

to generate the required weights data/cnn.lut, you need gmic-3.0.0 and dump the network weights like so (all in this directory):

gmic ./gmic_denoise_cnn.gmz k[3] unserialize repeat '$!' 'l[$>]' o '{b}'.pfm endl done
clang -Wall -std=c11 -I ../../../ ingest.c -g -o ingest
./ingest
mv cnn.lut ${VKDT_BIN_DIR}/data/

assuming that your git repo bin/ or system wide installed /usr/lib/vkdt directory names are stored in the envvar ${VKDT_BIN_DIR}.