Demo error - 'BatchNorm2d' object has no attribute 'track_running_stats'

[tobiascz]

Hi,

I'am using torch version 0.5 and i get that error in that line

output = model(input_var)

The error message looks like that:

  File "src/demo.py", line 26, in main
    output = model(input_var)
  File "/home/narvis/miniconda3/envs/hpe/lib/python3.6/site-packages/torch/nn/modules/module.py", line 477, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/narvis/Dev/pytorch-pose-hg-3d/src/models/hg_3d.py", line 101, in forward
    x = self.bn1(x)
  File "/home/narvis/miniconda3/envs/hpe/lib/python3.6/site-packages/torch/nn/modules/module.py", line 477, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/narvis/miniconda3/envs/hpe/lib/python3.6/site-packages/torch/nn/modules/batchnorm.py", line 66, in forward
    self.training or not self.track_running_stats,
  File "/home/narvis/miniconda3/envs/hpe/lib/python3.6/site-packages/torch/nn/modules/module.py", line 518, in __getattr__
    type(self).__name__, name))
AttributeError: 'BatchNorm2d' object has no attribute 'track_running_stats'

I found that other repositories have similar issues when upgrading from torch 0.3 to 0.4:
kunglab/ddnn#2

And I checked the specific commit to find out how to solve this errorr: kunglab/ddnn@071c82f

But I couldn't find the right solution yet.

I would like to use the most recent torch version so downgrading is not really an option for me.

[tobiascz]

In the mean time i started training the model myself on the dataset with python3 and pytorch 0.5. I can load the model and do the forward pass without getting the track_running_stats missing error.

[llcshappy]

"train_params = [x for x in model.parameters() if x.requires_grad] "
It works.

[jixinhe111]

In the mean time i started training the model myself on the dataset with python3 and pytorch 0.5. I can load the model and do the forward pass without getting the track_running_stats missing error.

Hi,
I'am using torch version 0.4 and i get that error, can you tell how to solve this problem,thanks.

[tobiascz]

The Problem is that higher torch versions use the attribute track_running_stats. But track_runninh_stats is a new attribute and by the time this model was trained it was not available. So you try loading a model without this Attribute with a pytorch version that needs that attribute. I couldn’t find a way to add this attribute to the already trained model that is why I trained my own model with a higher pytorch version so that the model includes the attribute. Hope I could help

[zkyf]

pytorch 0.3.1 works well on my computer. My configuration is Ubuntu 16.04, cuda 9.0, pytroch 0.3.1, python 2.7.15.

[aseerkhan]

I am facing the same problem ...is any can help me

[zkyf]

I believe you need to manually download an older version of pytorch like 0.3.0 or 0.3.1

[xingyizhou]

Hi,
The current version should work on python3 and pytorch 0.4+

[heroinlin]

Your model train by pytorch0.3.x, but run in pytorch > 0.4.0.
Change the parameter of BatchNorm2d by yourself.
For example, define the function

def recursion_change_bn(module):
    if isinstance(module, torch.nn.BatchNorm2d):
        module.track_running_stats = 1
    else:
        for i, (name, module1) in enumerate(module._modules.items()):
            module1 = recursion_change_bn(module1)
    return module

and
use it when you load model

check_point = torch.load(check_point_file_path)
model = check_point['net']
for i, (name, module) in enumerate(model._modules.items()):
    module = recursion_change_bn(model)
model.eval()

I have ran 0.3.1 model in pytorch0.4.1 and pytorch1.0.0.
you can use this way to update you model from 0.3.x to 0.4+

[chandiniRNair]

@heroinlin : I added the above code and a new error creeped in:
File "/home/chandini/.local/lib/python3.6/site-packages/torch/nn/modules/module.py", line 535, in getattr
type(self).name, name))
AttributeError: 'Upsample' object has no attribute 'name'

I couldn't find any leads on this in google search.
Any suggestions please ?

[wuzb19951013]

@chanduKichu I have met the same problem. Did you solve it?

[gauharbains]

@chandiniRNair @wuzb19951013 I have the same error. Were you able to figure it out ?

[tianlinxu312]

A working solution is to load the state_dict of the model with argument "strict=False".

Example code as below:

ckpt = torch.load(model_path)
model = Model(model_args)
model.load_state_dict(ckpt.state_dict(), strict=False)
model.eval()

[yyunhh]

I got the error:

AttributeError: 'Upsample' object has no attribute 'align_corners'

There is no answer to my error on google.
But I try to modify the lib\site-packages\torch\nn\modules\upsampling.py", line 141:

def forward(self, input: Tensor) -> Tensor:
        return F.interpolate(input, self.size, self.scale_factor, self.mode)#self.align_corners

It works...
Hope it helps someone encounter the same error, but please correct me if I am doing the wrong way out.