search

1adrianb / face-alignment-training

Training code for the networks described in "How far are we from solving the 2D & 3D Face Alignment problem? (and a dataset of 230,000 3D facial landmarks)" paper.
https://www.adrianbulat.com/face-alignment
Other
145 stars 48 forks source link

Batchnorm parameter #15

Open zhengge opened 6 years ago

zhengge commented 6 years ago 
local function convBlock(numIn, numOut, order)
    local cnet = nn.Sequential()
        :add(batchnorm(numIn,1e-5,false))
        :add(relu(true))
        :add(conv(numIn,numOut/2,3,3,1,1,1,1):noBias())
        :add(nn.ConcatTable()
            :add(nn.Identity())
            :add(nn.Sequential()
                :add(nn.Sequential()
                    :add(batchnorm(numOut/2,1e-5,false))
                    :add(relu(true))
                    :add(conv(numOut/2,numOut/4,3,3,1,1,1,1):noBias())
                )

batchnorm(numIn,1e-5,false) Dose this statement want to set affine = false？ But in fact, four parameters (nOutput, eps, momentum, affine) are needed. The batchnorm will have wight and bias.

lippman1125 commented 5 years ago

Same question as above