AttributeError: module 'torch.nn.functional' has no attribute 'instance_norm'

vision
1

Hi all,

While I’m trying to use the instance_norm function in the functional class with my own weight and bias, it raises some errors like below,

Traceback (most recent call last):
File “learn_train.py”, line 320, in
pred_v = model([inputParaTensor_v, inputTensor_v])
File “/root/anaconda3/lib/python3.6/site-packages/torch/nn/modules/module.py”, line 357, in call
result = self.forward(*input, **kwargs)
File “/root/anaconda3/lib/python3.6/site-packages/torch/nn/parallel/data_parallel.py”, line 71, in forward
return self.module(*inputs[0], **kwargs[0])
File “/root/anaconda3/lib/python3.6/site-packages/torch/nn/modules/module.py”, line 357, in call
result = self.forward(*input, **kwargs)
File “learn_train.py”, line 90, in forward
conv1_feat = self.relu(F.layer_norm(conv1_feat_norm))
AttributeError: module ‘torch.nn.functional’ has no attribute ‘layer_norm’

The code snippet is very simple,

conv1_feat_norm = F.conv2d(x_img, self.conv1_weight, bias=None, padding=1)
conv1_feat = self.relu(F.layer_norm(conv1_feat_norm))

Does anyone have any idea why this happens?

Best

2

Hi,

What is the version of pytorch you’re using?
If you use the latest binary release which is 0.3.1, this function is not in the doc.
This function has been added in master and is only available if you compile from source at the moment.

3

Hi,

Thank you very much. This works for me very well!

Best wishes

4

Hi, I am also getting a similar error after the announcement that nn.Upsample is DEPRECATED in favor of interpolate:

My ERROR:

AttributeError: module 'torch.nn.functional' has no attribute 'interpolate'

I have added this interpolate layer inside a nn.Sequential.

Any insight into this.

1 Like
5

I think torch removed the interpolate layer inside the nn.function and created the equivalent methods for Upsampling and other in different modes.

https://pytorch.org/docs/stable/nn.html?highlight=interpolate

11%20PM
11%20PM1390×1106 88.4 KB

6

I think torch removed the interpolate layer inside the nn.function and created the equivalent methods for Upsampling and other in different modes.

It’s opposite of what you said.

I checked my PyTorch version. It is 4.0.0 py36_cuda8.0.61_cudnn7.1.2_1 . So, it was my mistake, because as per the doc, UpSample() is deprecated in version > 4.0.0