Fix Bugs of Model Structure and Learning Rate

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

Open a00012025 opened 5 years ago

a00012025 commented 5 years ago

Hi, I've fixed the following bugs when I was reproducing your results and modifying the code:

Change average pooling kernel size from 1 to 2
- It's more reasonable to have 2 as kernel size. Please correct me if 1 is intended to be used.
Fix bug to support exponential learning rate decay
Fix shape bug when creating fc layer
- get_shape() will only return 2 values because global_average_pooling has been applied above