Questions about the network structure

[catherineytw]

hi, I was carefully reading the code and paper in the past few days, and had to say that it was really amazing!

but when I dug into the details and was confused about the implementation of the Encoder structure.

In the paper, the network has two layers, each of which has 3 operations and comes in the sequence of { skeletonConv, ReLu, and pooling}, but in the code, the order of the skeleton pooling and relu are inverted. Below is the module list I printed out to show the inconsistency, and was wondering which one was correct?

Many thanks!

The module list of encoder I printed: Encoder: ModuleList( (0): Sequential( (0): SkeletonConv( (offset_enc): SkeletonLinear() ) (1): SkeletonPool() (2): LeakyReLU(negative_slope=0.2) ) (1): Sequential( (0): SkeletonConv( (offset_enc): SkeletonLinear() ) (1): SkeletonPool() (2): LeakyReLU(negative_slope=0.2) ) )

[PeizhuoLi]

Hi, thanks for your interest in this project. You are right about the code. We made a mistake in the paper. The correct order is Conv, Pool, LeakyReLU.

[catherineytw]

Many thanks. 

Sent from Yahoo Mail for iPhone

On Tuesday, August 11, 2020, 22:47, Peizhuo notifications@github.com wrote:

Hi, thanks for your interest in this project. You are right about the code. We made a mistake in the paper. The correct order is Conv, Pool, LeakyReLU.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.