how to train?

[jianlong-yuan]

Could you give us your train script?

[chenxi116]

This repo is a simplified version of the official implementation in TF Slim. If you want to train, I suggest using the official implementation. Here may be a good place to start.

[jianlong-yuan]

Hi Thank you for your answer. But TF Slim also can not reproduce the results of your article by using RL. I don't find it uses RL.

[chenxi116]

I assume by training, you mean training our found architecture, PNASNet-5. I assume by RL, you mean reinforcement learning. First, training a network does not involve reinforcement learning: it uses stochastic gradient descent. Second, our search algorithm does not use reinforcement learning either.

[jianlong-yuan]

I apologize for my statement. I want to search for network structures, not structures you've already searched for.

[chenxi116]

I see. Sorry we did not release the code for architecture search, as that part is too heavily dependent on infrastructure.