ML AutoML NAS ENAS

[junxnone]

ENAS

• 子模型共享参数, 提高效率，减少所需计算资源
• macro search space over entire convolutional models
• micro search space over sonvolutional cells
• Decision Block
• Conv Cell
• Reduction Cell

Search Space

• Model = Blocks x N
• Blocks = Conv Cell x N + Reduction x M
• Cell = Node x B (B=4)

Entire Search Space -- |

macro Search Space	Micro Search Space

Macro Search Space - Node Avaliable Operations	Micro Search Space - Node Available Operations
	identity
convolutions with filter sizes 3 × 3
convolutions with filter sizes 5 × 5
depthwise-separable convolutions with filter sizes 3×3	separable convolution with kernel size 3 × 3
depthwise-separable convolutions with filter sizes 5×5	separable convolution with kernel size 5 × 5
max pooling of kernel size 3 × 3	max pooling with kernel size 3×3
average pooling of kernel size 3 × 3	average pooling with kernel size 3×3

Recurrent Cell

Discovered

Name	Discovered
RNN Cell
CNN Network
Convolution Cell
Reduction Cell

Reference

• paper - 2018 - Efficient Neural Architecture Search via Parameter Sharing
• 论文笔记系列-Efficient Neural Architecture Search via Parameter Sharing
• Github - Authors - Tensorflow
• ENAS-pytorch
• AutoDL论文解读（四）：权值共享的搜索
• 【ENAS】2018-ICML-Efficient Neural Architecture Search via Parameter Sharing-论文阅读

[junxnone]

junxnone/tech-io#805 junxnone/aiwiki#393