search

run-youngjoo / SC-FEGAN

SC-FEGAN : Face Editing Generative Adversarial Network with User's Sketch and Color (ICCV2019)
Other
3.35k stars 530 forks source link

question in GatedConv2d #45

Open tangeroo opened 5 years ago

tangeroo commented 5 years ago

hi, Thanks very much for your work.Here I have a question: in GatedConv2d, why does sigmoid function add to the conved inp instead of conved mask. Thank you image

abtExp commented 4 years ago

the concept of gated convolutions is based on learning weights based on the features, by activating only certain nodes in the kernel depending on the input. The gated convolution filter learns the relations between the input feature maps and the most prominent weights for that input feature map, thus it applies the sigmoid on the input allowing certain parts of input to contribute more to the output. (Giving higher weights to some features and lower weights to other features.)