RuntimeError: Error(s) in loading state_dict for DataParallel: Missing key(s) in state_dict

1

I’m new to PyTorch, working on some image processing-related projects. I’m currently using ResNet101 and here I am facing this problem:

RuntimeError: Error(s) in loading state_dict for DataParallel:
Missing key(s) in state_dict: “module.context_encoding.stages.0.2.bn.weight”, “module.context_encoding.stages.0.2.bn.bias”, “module.context_encoding.stages
.0.2.bn.running_mean”, “module.context_encoding.stages.0.2.bn.running_var”, “module.context_encoding.stages.1.2.bn.weight”, “module.context_encoding.stages.1.2.bn.
bias”, “module.context_encoding.stages.1.2.bn.running_mean”, “module.context_encoding.stages.1.2.bn.running_var”, “module.context_encoding.stages.2.2.bn.weight”, "
module.context_encoding.stages.2.2.bn.bias", “module.context_encoding.stages.2.2.bn.running_mean”, “module.context_encoding.stages.2.2.bn.running_var”, “module.con
text_encoding.stages.3.2.bn.weight”, “module.context_encoding.stages.3.2.bn.bias”, “module.context_encoding.stages.3.2.bn.running_mean”, “module.context_encoding.s
tages.3.2.bn.running_var”, “module.context_encoding.bottleneck.1.bn.weight”, “module.context_encoding.bottleneck.1.bn.bias”, “module.context_encoding.bottleneck.1.
bn.running_mean”, “module.context_encoding.bottleneck.1.bn.running_var”, “module.edge.conv1.1.bn.weight”, “module.edge.conv1.1.bn.bias”, “module.edge.conv1.1.bn.ru
nning_mean”, “module.edge.conv1.1.bn.running_var”, “module.edge.conv2.1.bn.weight”, “module.edge.conv2.1.bn.bias”, “module.edge.conv2.1.bn.running_mean”, “module.e
dge.conv2.1.bn.running_var”, “module.edge.conv3.1.bn.weight”, “module.edge.conv3.1.bn.bias”, “module.edge.conv3.1.bn.running_mean”, “module.edge.conv3.1.bn.running
_var”, “module.decoder.conv1.1.bn.weight”, “module.decoder.conv1.1.bn.bias”, “module.decoder.conv1.1.bn.running_mean”, “module.decoder.conv1.1.bn.running_var”, “mo
dule.decoder.conv2.1.bn.weight”, “module.decoder.conv2.1.bn.bias”, “module.decoder.conv2.1.bn.running_mean”, “module.decoder.conv2.1.bn.running_var”, “module.decod
er.conv3.1.bn.weight”, “module.decoder.conv3.1.bn.bias”, “module.decoder.conv3.1.bn.running_mean”, “module.decoder.conv3.1.bn.running_var”, “module.decoder.conv3.3
.bn.weight”, “module.decoder.conv3.3.bn.bias”, “module.decoder.conv3.3.bn.running_mean”, “module.decoder.conv3.3.bn.running_var”, “module.fushion.1.bn.weight”, “module.fushion.1.bn.bias”, “module.fushion.1.bn.running_mean”, “module.fushion.1.bn.running_var”.
Unexpected key(s) in state_dict: “module.context_encoding.stages.0.2.weight”, “module.context_encoding.stages.0.2.bias”, “module.context_encoding.stages.0.
2.running_mean”, “module.context_encoding.stages.0.2.running_var”, “module.context_encoding.stages.1.2.weight”, “module.context_encoding.stages.1.2.bias”, “module.
context_encoding.stages.1.2.running_mean”, “module.context_encoding.stages.1.2.running_var”, “module.context_encoding.stages.2.2.weight”, “module.context_encoding.
stages.2.2.bias”, “module.context_encoding.stages.2.2.running_mean”, “module.context_encoding.stages.2.2.running_var”, “module.context_encoding.stages.3.2.weight”,
“module.context_encoding.stages.3.2.bias”, “module.context_encoding.stages.3.2.running_mean”, “module.context_encoding.stages.3.2.running_var”, “module.context_en
coding.bottleneck.1.weight”, “module.context_encoding.bottleneck.1.bias”, “module.context_encoding.bottleneck.1.running_mean”, “module.context_encoding.bottleneck.
1.running_var”, “module.edge.conv1.1.weight”, “module.edge.conv1.1.bias”, “module.edge.conv1.1.running_mean”, “module.edge.conv1.1.running_var”, “module.edge.conv2
.1.weight”, “module.edge.conv2.1.bias”, “module.edge.conv2.1.running_mean”, “module.edge.conv2.1.running_var”, “module.edge.conv3.1.weight”, “module.edge.conv3.1.b
ias”, “module.edge.conv3.1.running_mean”, “module.edge.conv3.1.running_var”, “module.decoder.conv1.1.weight”, “module.decoder.conv1.1.bias”, “module.decoder.conv1.
1.running_mean”, “module.decoder.conv1.1.running_var”, “module.decoder.conv2.1.weight”, “module.decoder.conv2.1.bias”, “module.decoder.conv2.1.running_mean”, “modu
le.decoder.conv2.1.running_var”, “module.decoder.conv3.1.weight”, “module.decoder.conv3.1.bias”, “module.decoder.conv3.1.running_mean”, “module.decoder.conv3.1.run
ning_var”, “module.decoder.conv3.3.weight”, “module.decoder.conv3.3.bias”, “module.decoder.conv3.3.running_mean”, “module.decoder.conv3.3.running_var”, “module.fushion.1.weight”, “module.fushion.1.bias”, “module.fushion.1.running_mean”, “module.fushion.1.running_var”.

I have only little knowledge on PyTorch. Based on what I understand from the error messages, it seems like the bn layer is not loaded. Can anyone help me with this? Thanks!

2

@ptrblck I’ve seen you helping out others with their problems which are similar as mine. Could you please help? Thank you very very very very very much :pleading_face:

3

I guess you are trying to load a state_dict stored by a plain model into an nn.DataParallel model, as the load_state_dict method is complaining about the missing .module attributes.
If so, then try to load the state_dict into the model before wrapping it into nn.DataParallel:

model = MyModel()
model.load_state_dict(torch.load(path))
model = nn.DataParallel(model)
4

Thanks for replying!

I have changed my code from:
model = network(num_classes=num_classes, pretrained=None)
model = nn.DataParallel(model)
state_dict = torch.load(args.restore_weight)[‘state_dict’]
model.load_state_dict(state_dict)
model.eval()

To:
model = network(num_classes=num_classes, pretrained=None)
model = nn.DataParallel(model)
state_dict = torch.load(args.restore_weight)[‘state_dict’]
model.load_state_dict(torch.load(‘./resnet101-imagenet.pth’))[‘state_dict’]
model.load_state_dict(state_dict)
model.eval()

Could you please guide me if I’m doing it correctly?
Now the runtime error is:
RuntimeError: Error(s) in loading state_dict for DataParallel: Missing key(s) in state_dict: “module.conv1.weight”, “module.bn1.weight”, “module.bn1.bias”, “module.bn1.running_mean”,…
Unexpected key(s) in state_dict: “conv1.weight”, “bn1.running_mean”, “bn1.running_var”, “bn1.weight”, “bn1.bias”, “layer1.0.conv1.weight”, “layer1.0.bn1.running_mean”,…

Now it says the “module.” is missing…

5

You are still wrapping the model into nn.DataParallel before loading the state_dict and the error message is thus the same.
Swap the order: i.e. load the state_dict into the model and wrap it into nn.DataParallel afterwards.

6

Now my code is:
model = network(num_classes=num_classes, pretrained=None)
state_dict = model.load_state_dict(torch.load(‘./resnet101-imagenet.pth’))
model.load_state_dict(state_dict)
model = nn.DataParallel(model)
model.eval()

RuntimeError: Error(s) in loading state_dict for ResNet: Missing key(s) in state_dict: “context_encoding.stages.0.1.weight”, “context_encoding.stages.0.2.bn.weight”, “context_encoding.stages.0.2.bn.bias”, “context_encoding.stages.0.2.bn.running_mean”, “context_encoding.stages.0.2.bn.running_var”, “context_encoding.stages.1.1.weight”, “context_encoding.stages.1.2.bn.weight”, “context_encoding.stages.1.2.bn.bias”, “context_encoding.stages.1.2.bn.running_mean”, “context_encoding.stages.1.2.bn.running_var”, “context_encoding.stages.2.1.weight”, “context_encoding.stages.2.2.bn.weight”, “context_encoding.stages.2.2.bn.bias”, “context_encoding.stages.2.2.bn.running_mean”, “context_encoding.stages.2.2.bn.running_var”, “context_encoding.stages.3.1.weight”, “context_encoding.stages.3.2.bn.weight”, “context_encoding.stages.3.2.bn.bias”, “context_encoding.stages.3.2.bn.running_mean”, “context_encoding.stages.3.2.bn.running_var”, “context_encoding.bottleneck.0.weight”, “context_encoding.bottleneck.1.bn.weight”, “context_encoding.bottleneck.1.bn.bias”, “context_encoding.bottleneck.1.bn.running_mean”, “context_encoding.bottleneck.1.bn.running_var”, “edge.conv1.0.weight”, “edge.conv1.1.bn.weight”, “edge.conv1.1.bn.bias”, “edge.conv1.1.bn.running_mean”, “edge.conv1.1.bn.running_var”, “edge.conv2.0.weight”, “edge.conv2.1.bn.weight”, “edge.conv2.1.bn.bias”, “edge.conv2.1.bn.running_mean”, “edge.conv2.1.bn.running_var”, “edge.conv3.0.weight”, “edge.conv3.1.bn.weight”, “edge.conv3.1.bn.bias”, “edge.conv3.1.bn.running_mean”, “edge.conv3.1.bn.running_var”, “edge.conv4.weight”, “edge.conv4.bias”, “edge.conv5.weight”, “edge.conv5.bias”, “decoder.conv1.0.weight”, “decoder.conv1.1.bn.weight”, “decoder.conv1.1.bn.bias”, “decoder.conv1.1.bn.running_mean”, “decoder.conv1.1.bn.running_var”, “decoder.conv2.0.weight”, “decoder.conv2.1.bn.weight”, “decoder.conv2.1.bn.bias”, “decoder.conv2.1.bn.running_mean”, “decoder.conv2.1.bn.running_var”, “decoder.conv3.0.weight”, “decoder.conv3.1.bn.weight”, “decoder.conv3.1.bn.bias”, “decoder.conv3.1.bn.running_mean”, “decoder.conv3.1.bn.running_var”, “decoder.conv3.2.weight”, “decoder.conv3.3.bn.weight”, “decoder.conv3.3.bn.bias”, “decoder.conv3.3.bn.running_mean”, “decoder.conv3.3.bn.running_var”, “decoder.conv4.weight”, “decoder.conv4.bias”, “fushion.0.weight”, “fushion.1.bn.weight”, “fushion.1.bn.bias”, “fushion.1.bn.running_mean”, “fushion.1.bn.running_var”, “fushion.3.weight”, “fushion.3.bias”. Unexpected key(s) in state_dict: “fc.weight”, “fc.bias”.

Did I get you correctly for first loading the state_dict into model then only wrap it?

7

Yes, this was my suggestion. The error message now changed but your model is also unable to map this state_dict to the modules.
You would have to check how the internal modules in your model are defined and in particular check what the state_dict contains.
In particular, your model seems to contain model.context_encoding, model.edge, etc. which are all missing in the state_dict.

8

Alright, no problem. Thanks a lot for your help! I will continue debugging the code and see what is the issue. Sincere thanks! Have a nice day :heart:

9

Sure, let me know if you get stuck.
PS: I would start by checking the module names of the model and compare it against the keys in the state_dict. Something like this could give you more information what’s causing the mismatch:

model = models.resnet18()
state_dict = models.vgg16().state_dict()

for name, _ in model.named_modules():
    print(name)
    
for key in state_dict:
    print(key)

This would of course also create a mismatch when you are trying to load the VGG16 state_dict into a ResNet and would show that the actual module names are completely different.

10

Hi. With the below code:
model = network(num_classes=num_classes, pretrained=None)
state_dict = torch.load(args.restore_weight)[‘state_dict’]
model.load_state_dict(state_dict)
model = nn.DataParallel(model)
model.eval()
I face the problem of having an extra ‘module.’ in the keys for state_dict, while the ‘bn.’ in the keys is still present. So, I have thought of removing ‘module.’ with the below code:
model = network(num_classes=num_classes, pretrained=None)
state_dict = torch.load(args.restore_weight)[‘state_dict’]
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
model.load_state_dict(new_state_dict)
model = nn.DataParallel(model)
model.eval()
With this, now I finally have ‘module.’ removed. However, another problem I face which is the ‘bn.’ is removed as well. Do you have any idea on this? Thanks in advance.

11

Remove the .module string from the key conditionally if it’s found only by adding an if condition before calling: name = k[7:] so that other parameter keys won’t be changed.