How do you learn reduction cell?

quark0 / darts

Differentiable architecture search for convolutional and recurrent networks

https://arxiv.org/abs/1806.09055

Apache License 2.0

3.92k stars 843 forks source link

How do you learn reduction cell? #79

Open imirzadeh opened 5 years ago

imirzadeh commented 5 years ago

Hi,

In the paper, the best-found reduction cell consists only pooling layers and skip connections. But, in the code, you don't select pooling layers explicitly and you just set stride to 2.

Does that mean after learning and pruning, the best-found weights belong to pooling/skip connection operations? Just want to make sure :)

Catosine commented 5 years ago

As far as I know, you are right! The cells are selected based on alpha; operations with the highest alpha value after search will be used to build the cells. So, you may conclude that the algorithm picks the poolings and skip-connections for reduction cell.

Best,