huangleibuaa / iternorm-pytorch Goto Github PK

If I use ./extension/normalization/iterative_normalization.py directly, I get the error - "more than one element of the written-to tensor refers to a single memory location". This error occurs due to the PyTorch version update. I could find the solution here.

Basically,
change self.register_buffer('running_wm', torch.eye(num_channels).expand(num_groups, num_channels, num_channels) to self.register_buffer('running_wm', torch.eye(num_channels).expand(num_groups, num_channels, num_channels).clone()).

I tried to run DBN, but the implementation cause gpu memory leaks in my computer whose torch version is 0.4.1. I fix this by changing the following lines of codes
self.running_mean = (1. - self.momentum) * self.running_mean + self.momentum * mean
self.running_projection = (1. - self.momentum) * self.running_projection + self.momentum * wm
to
self.running_mean = (1. - self.momentum) * self.running_mean + self.momentum * mean.data
self.running_projection = (1. - self.momentum) * self.running_projection + self.momentum * wm.data

Thanks for the implementation! I get the following error when using a 2D vector as input Bxd. Am I using this layer wrong? I also tried reshaping to Bxdx1x1 (d=256)

xc = saved[0]  # centered input
IndexError: tuple index out of range

with the following construction

self.norm = IterNorm(256, num_groups=8, affine=False, momentum=1)