[Question] In micro_controller, about the shape of params

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

[Question] In micro_controller, about the shape of params #20

Closed yukang2017 closed 6 years ago

yukang2017 commented 6 years ago

Thank you for your code.

https://github.com/melodyguan/enas/blob/2734eb2657847f090e1bc5c51c2b9cbf0be51887/src/cifar10/micro_controller.py#L178

This line of code is supposed to decide the type of op (e.g. conv 3x3 or 5x5). But the shape of self.w_soft, self.b_soft is [self.lstm_size, self.num_branches] and [1, self.num_branches].

I think they should be [self.lstm_size, self.num_type_ops] and [1, self.num_type_ops] ?

hyhieu commented 6 years ago

In our code, num_branches is the class variable that means number of possible operations, which is num_type_ops in your understanding.