search

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

Possible Bug in in micro_child.py #83

Open yashkant opened 5 years ago

yashkant commented 5 years ago

Hi,

It appears to me that while building the input to the next pooling layer, we are throwing away layers[0] element with this line. https://github.com/melodyguan/enas/blob/d1a90ac915301198f2a30ce136e9040e6c4235ff/src/cifar10/micro_child.py#L277

I think it should be instead layers = layers[layers[0], x]

Before the factorized_reduction operation x would already contain the layers[-1] element from previous loop and I think we should only replace the last element of layers with the factorized_reduction(x).

Instead what is being done is: layers = [x, factorized_reduction(x)]

yashkant commented 5 years ago

@melodyguan Hi, could you please have a look into this.

AlloNighthawk commented 5 years ago

I am also curious about this problem. Hope to get response from author.