iclr18_estimate_bits_z entropy model does not match paper

[zzzzz314314]

Hi, thanks to your hardwork and the opensourcing.
As said in title, the BitEstimator module (

iclr_17_compression/models/bitEstimator.py

Line 22 in bce72bc

return torch.sigmoid(x * F.softplus(self.h) + self.b)

) used in iclr18_estimate_bits_z does not match the paper, where h should be matrices multiplication to each channel instead of scaling factors applied to each channel. I was wondering if you have experimented this and found there is little difference. Thanks!