KeyError: 'unexpected key "module.conv1_1.weight" in state_dict'

[nelaturuharsha]

Traceback (most recent call last):
File "predict.py", line 18, in
model = create_model(opt)
File "/home/Documents//enlighten/EnlightenGAN/models/models.py", line 36, in create_model
model.initialize(opt)
File "/home/Documents/enlighten/EnlightenGAN/models/single_model.py", line 72, in initialize
self.load_network(self.netG_A, 'G_A', which_epoch)
File "/home/Documents/enlighten/EnlightenGAN/models/base_model.py", line 53, in load_network
network.load_state_dict(torch.load(save_path))
File "/home/anaconda3/envs/enlighten/lib/python3.5/site-packages/torch/nn/modules/module.py", line 522, in load_state_dict
.format(name))
KeyError: 'unexpected key "module.conv1_1.weight" in state_dict'

Python: 3.5.6
PyTorch : 0.3.1
CPU

I'm facing this error when I run the test command in script/scripts.py, I would appreciate any help!

Thank you in advance!

[nelaturuharsha]

Fix :

In models/base_model.py replace the load_network with the following:

def load_network(self, network, network_label, epoch_label):
save_filename = '%s_net_%s.pth' % (epoch_label, network_label)
save_path = os.path.join(self.save_dir, save_filename)
state_dict = torch.load(save_path)
new_state_dict = OrderedDict()
for k, v in state_dict.items():
name = k[7:] # remove module.
new_state_dict[name] = v
network.load_state_dict(new_state_dict)

[Dear-Mr]

您好，在经过上述修改后，
并将 new_state_dict = OrderedDict() 修改为 new_state_dict = collections.OrderedDict() 后，
出现KeyError: 'unexpected key "conv10.bias" in state_dict'，该如何解决？

[nelaturuharsha]

Is your python 3.5, pytorch 0.3.1? And are you trying to run via CPU or GPU?

[yifanjiang19]

@Dear-Mr @SreeHarshaNelaturu
This problem is about torch.nn.DataParallel(). Normally DataParallel() will cover the original model, which means you can only use model.module to access the original model. And also, if you directly save the model covered by DataParallel, the key of saved model will be module.conv.weight instead of conv.weight.
You can simply use DataParallel to cover the model before you load the pre-trained model to solve this problem.

Eg.

model = ResNet()
parallel_model = torch.nn.DataParallel(model)
print(parallel_model.module)
# Here parallel_model.module is equal to model

[yifanjiang19]

I've pushed a new version. You can directly pull the new version to avoid this problem.

[Dear-Mr]

Is your python 3.5, pytorch 0.3.1? And are you trying to run via CPU or GPU?

Thanks for your reply.
My workspace is python 3.5, PyTorch 0.3.1, torchvision 0.2.0. And i tried to run via GPU.