Model Format

structure

現時点ではリストの後ろが入力側な点に注意

• module(compound list)
  • activation
    • id(string) どの活性化関数を使うか
  • layer(compound list)
    • node(compound list)
      • affine
        • weight(fixed list)
        • bias(fixed)
    • functions(string list) 層ごとに個別設定する処理(softmaxなど)

例

• NBT Format {module:[{activation:"relu",layer:[{node:[{affine:{weight:[int list],bias:int}},{...}]},{...}]},{...}]}

• Path Format module[0].layer[0].node[0].affine.weight[0]