Architecture & Scoring Question

[c31601]

Hi Assistants,

Could we change the layer, such as ReLU -> PReLU, short-cut from 1x1 to 3x3...etc.
Or we could only modify the hyper-parameters, such as iteration, lr...etc.?

The score of the exercise only depends on performance of the Res110?
Res20 & Res56 not considered?

Thank you so much!

[JiaRenChang]

Hi,
The paper of "wide ResNet" had experiments on shortcut.
https://arxiv.org/pdf/1605.07146.pdf

You can change ReLU layer to another activation function, but it may not improve your net.
You also can modify hyper-parameters, and it may not improve your net.

Yes, the network performance score of your homework 1 depend on the result of ResNet-110.
However, you still need to report your result of ResNet-20 and ResNet-56.

Jia-Ren