yuuho / stylegans-pytorch Goto Github PK

as paper said

The activation funtion(Lrelu) is always applied right after adding the bias

you only have bias but not act

reference original repo:
https://github.com/NVlabs/stylegan2/blob/cec605e0834de5404d5c7e5cead7053bdd0e4dde/dnnlib/tflib/ops/fused_bias_act.py#L72-L96

the size of weight should be (N, iC, oC, kH, kW)

the size of modulated_weight is [N, iC, oC, kH, kW]
the size of demod_norm is [N, oC]

This line is view demod_norm to [N, 1, oC, 1, 1]

but

this line is view demod_norm to [N, oC, 1, 1, 1]

this two function has a difficult result I guess

GTX1080ti (11GB) 以上の環境で動かすことしか想定していなかったため
バッチサイズ16固定としていた

• 基本的にはバッチサイズ 1
• 選択して16とか選べるようにする

the NVlabs:
https://github.com/NVlabs/stylegan2/blob/cec605e0834de5404d5c7e5cead7053bdd0e4dde/dnnlib/tflib/ops/upfirdn_2d.py#L234-L292

This part code make a filter_kernel：k
parameter of input k is [1, 3, 3, 1]
and after this line the k is:
[4/64, 12/64, 12/64, 4/64]
[12/64, 36/64, 36/64, 12/64]
[12/64, 36/64, 36/64, 12/64]
[4/64, 12/64, 12/64, 4/64]

but your code the FusedBlur3x3
it use filter kernel is :
[1/16, 2/16, 1/16]
[2/16, 4/16, 2/16]
[1/16, 2/16, 1/16]
why?