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

The alpha variable's value of "avg_pool" and "skip_connect" is the mostly biggest over the others #124

Open shuangjian24 opened 5 years ago

shuangjian24 commented 5 years ago

The alpha variable's value of "avg_pool" and "skip_connect" is the mostly biggest over the others, and "none" alpha is very big to 0.6, when search the normal_cell. And it is the same with reduce_cell.

shuangjian24 commented 5 years ago

This thing happens in my own time series dataset.

rangerToby commented 5 years ago

@shuangjian24 this happens on my own dataset cnn classification too, and i find something in paper PC-darts, which can explain this, pooling op(or skip connect) has no parameters so it will perform better than parameter-op