Loading weights from a pretrained model to a new model with added layer

Hi, I have model A which is pretrained and Model B which is new. Model B has all of the layers of model A + an extra layer.

I want to load the weights from Model A → B for the layers they have in common.

I implemented the following:

pretrained_path = torch.load(path to pretrained model)

new_model_dict = model.state_dict()

pretrained_weights = { k:v for k , v in pretrained_path.items() if k in new_model_dict}


model.load_state_dict(pretrained_weights, strict = False)

I had to use strict = False because the extra layer in model B was giving me an error. But this has made my results pretty bad.

How do I fix this? I want to avoid strict = False

In case you are using the same layer names, you could iterate the named_modules in both models and load the state_dict per layer. If that’s not the case, you could create a mapping between the names of the pretrained modelA layers and the modelB layers and do the same. This would avoid the strict=False usage.

Thanks, this approach worked!