Output None when using batchnorm1d in eval mode?

xiaoyulun · August 9, 2019, 6:40am

Below is test code

import torch
import torch.nn as nn
data = torch.rand(3, 10, 145)
bn = nn.BatchNorm1d(10)
bn.eval()
print bn(data) is None

Mazhar_Shaikh · August 9, 2019, 7:14am

Hi xiaoyulun,
I have run your test code and do not get output as None.

xiaoyulun · August 10, 2019, 5:55am

Really , which pytorch version do you use?

Mazhar_Shaikh · August 10, 2019, 6:34am

I tried in Pytorch 0.4 and 1.1.

Mazhar_Shaikh · August 10, 2019, 6:37am

xiaoyulun · August 12, 2019, 6:55am

Thanks for your help. Below is my result

But when I update pytorch to 1.2.0, the problem is solved

ptrblck · August 12, 2019, 1:54pm

I’ve installed PyTorch 1.1.0 for Python2.7 and used your code snippet in python and ipython.
Both yield valid results and I cannot reproduce this issue.
Do you also get a None output for nn.BatchNorm2d?

Could you update PyTorch to the latest stable release (1.2.0) and check it again?

xiaoyulun · August 13, 2019, 5:28am

Yes, When I update pytorch to 1.2.0. It can yield valid results. But I didn’t figure out why it is occured in pytorch 1.1.0

ptrblck · August 13, 2019, 1:54pm

I’m still unsure, how this might happen on your system, as I’ve tried to reproduce it using matching versions of Python, Pytorch etc. and wasn’t able to.