Error: 'DataParallel' object has no attribute 'items'

I am trying to convert PyTorch model to Onnx model.

I keep ending up with the following error:

Traceback (most recent call last):
  File "/tmp/pycharm_project_896/utils/", line 15, in <module>
    for k, v in state_dict.items():
  File "/usr/local/lib/python2.7/dist-packages/torch/nn/modules/", line 398, in __getattr__
    type(self).__name__, name))
AttributeError: 'DataParallel' object has no attribute 'items'

I have been using the following code:

import torch
from torch.autograd import Variable

# Load the trained model from file

path = "model.pth"

#pytorch_network = torch.load(path)
state_dict = torch.load("model.pth")

# Create new OrderedDict that does not contain the module
from collections import OrderedDict
new_state_dict = OrderedDict()
for k, v in state_dict.items():
    name = k[7:] # remove module
    new_state_dict[name] = v


# Export the trained model to ONNX
#pytorch_network = torch.nn.DataParallel(pytorch_network, device_ids=range(torch.cuda.device_count()))
dummy_input = Variable(torch.randn(1, 3, 224, 224))
torch.onnx.export(pytorch_network, dummy_input, "model.onnx")

Can anyone tell me where I might be going wrong. I keep ended up with the same error.
I have referred this Issue too but I don’t seem to end with a solution.


I guess it is similar with the problem you posted before.
The .pth file saves the entire model rather than the weights. Your state_dict is actually not an OrderedDict but a DataParallel object. That is why you get the error.