Convert existing model/weights fromData Parallel to generic?

deployment
1

Sorry if this isn’t the correct topic. I did try to look for my problem, but I don’t I think I’m using the right search terms, which is why I’m here.

So, I trained a model using nn.DataParallel and it worked fine. I was able to inference fine on my machine, but only my machine. I finally figured out that I mistakenly did NOT save my model as generic as described here: https://pytorch.org/tutorials/beginner/saving_loading_models.html#saving-torch-nn-dataparallel-models

Is there a way to convert my existing model to a generic form so it can be used by anyone? Hopefully in a automated way and not me sludging through the weights manually?

2

Are you still able to load the checkpoint? DataParallel just hides the actual model away under the attribute module. So you can do something like this:

my_model = ModelClass(...)
my_model = DataParallel(my_model, ...)

state_dict = torch.load(...)
my_model.load_state_dict(state_dict)

actual_model = my_model.module
new_state_dict = actual_model

Haven’t tested this, but I bet something like this would work as well:

state_dict = torch.load(...)

new_state_dict = {}
for k, v in state_dict.items():
    k = k.replace('module.', '')  # Each weight k will be prefix with the word "module"
    new_state_dict[k] = v

model = SomeModel(...)
model.load_state_dict(new_state_dict)
1 Like
3

Thanks, we tried something similar to the second method and it “works”, but unsure if the model is accurately predicting or not. Still investigating… Thanks again!