Open junxnone opened 3 years ago
Normal Cell
Reduction Cell
Addition
3x3 depthwise-separable convolution 5x5 depthwise-separable convolution 7x7 depthwise-separable convolution 1x7 followed by 7x1 convolution identity 3x3 average pooling 3x3 max pooling 3x3 dilated convolution
B1
B1:5
4x64
9x64
16x64
25x64
36x64
B=1
B=2
B=3
b
b+1
mean squared error
LSTM & MLP -- |
junxnone/tech-io#805
PNAS
Normal Cell
&Reduction Cell
Normal Cell
功能实现 - stride=1Reduction Cell
功能实现 - stride=2Arch
Addition
3x3 depthwise-separable convolution 5x5 depthwise-separable convolution 7x7 depthwise-separable convolution 1x7 followed by 7x1 convolution identity 3x3 average pooling 3x3 max pooling 3x3 dilated convolution
Search Space
B1
= 2x2x8x8x1 = 256Cell = Block x 5
B1:5
=4x64
x9x64
x16x64
x25x64
x36x64
= 5.6 x10e14去掉对称 Cells ~10e12
Algos
B=1
->B=2
->B=3
...Surrogate Model
b
Block , predict withb+1
Blockmean squared error
, 但是要给出真实的 performance 排序LSTM & MLP -- |![image](https://user-images.githubusercontent.com/2216970/88249899-6653ed80-ccd8-11ea-9960-d5117203bf25.png)
Search Result
Reference