What's the idea of `betas2path ` ?

[knsong]

Hi, thanks for the work first, but can you please explain the idea of betas2path function? For example, shouldn't down_idx = np.argmax([ beta[0] for beta in betas[1][1:-1].cpu().numpy() ]) + 1 be down_idx = np.argmax([ beta[0] for beta in betas[1][0:-1].cpu().numpy() ]) + 1 ?

[chenwydj]

Hi @knsong,

Thank you for your interest in our work!

betas[1][0] and betas[2][:2] are actually dummy variables. From Figure 1 in our paper we can see that paths of 1/16 and 1/32 scale have one and two fewer layers than the path of 1/8 scale. We insert these dummy beta variable just for the convenience of forming a matrix.

[knsong]

Hi @knsong,

Thank you for your interest in our work!

betas[1][0] and betas[2][:2] are actually dummy variables. From Figure 1 in our paper we can see that paths of 1/16 and 1/32 scale have one and two fewer layers than the path of 1/8 scale. We insert these dummy beta variable just for the convenience of forming a matrix.

thanks, but from here, the setting of size of betas has already considered removing the dummy data, right? By the way, why there is no comparing with AutoDeeplab in the paper?

[chenwydj]

taking last = 1 (1/16 scale) as the example, the line you mentioned is to consider the one less node, and betas[1][1:-1] is to further consider that the first node only has downsampled in-degree path, which makes the first node's beta[1] no sense.

We did not compare with AutoDeeplab since it is only performance-driven whereas we are both performance- and latency-driven.