Question about stats pooling in 2D convnet paper

[Sreyan88]

Hi there!

Great paper and great repo. My question is rather related to your paper. In the paper you mention:

The statistics pooling layer in speaker embeddings networks with 2D CNN architectures is a concatenation of the mean and std of each of the F × C frequency-channel pairs

I am a bit confused on this end. In pytorch terms if my resent output is B x C x T x F, how exactly do I implement stats pooling?

would it be:

x = x.permute(0,2,3,1) #B x T x F x C
x = x.reshape(B,T,F x C) # B x T x (FxC)

followed by a stats pooling layer?

Thank You for the help!

[tstafylakis]

Hi Sreyan,
Thanks for your interest. Stats pooling is very simple.
Your tensor is x = x.reshape(B,T,F x C),
Do something like:
x = torch.cat((torch.mean(x,dim=1),torch.std(x,dim=1))).reshape(B,FxCx2)
so that each example in your minibatch is a vector of dim = 2xFxC.
Is that clear?

[Sreyan88]

Hi @tstafylakis ,

Thank You so much for your reply. Yes, that is clear. I am currently trying to implement an attentive stats pooling layer. Will update you here once I am able to find a solution. Thank You!