Seeming inconsistency between cifar10 and ptb in terms of reward set for arc selection

melodyguan / enas

TensorFlow Code for paper "Efficient Neural Architecture Search via Parameter Sharing"

https://arxiv.org/abs/1802.03268

Apache License 2.0

1.58k stars 390 forks source link

Seeming inconsistency between cifar10 and ptb in terms of reward set for arc selection #40

Open AranKomat opened 6 years ago

AranKomat commented 6 years ago

In the final phase where you choose the best architecture based on their reward, the reward of ptb and cifar10 is set to be c/ppl + (entropy term) and accuracy, respectively. Why did you use entropy term for arc selection of ptb and not for that of cifar10?