How to crop weight file to use as pre-trained model?

1

hi everyone,
I have created autoencoder model with use ResNet backbone. Then I trained this model. I want to subtract the weight of the encoder part to load it into my normal ResNet model. But I don’t know how to do.
I tried crop autoencoder’s encoder parts model after load autoencoder weight file then I saved this model. But layers name changed and I could not load weight into normal ResNet model.

How can I do ?

2

By subtract the weights you mean, you want to remove the weights from the autoencoder or you want to take those weights out and place them into your ResNet model?

3

I want to take those weights out and place them into your ResNet model (just encoder part weights)

4

You can take out the model state dict from your encoder part and add then take out the other part from the ResNet model.

Encoder = encoder.__dict__

you can then take out the first layer like:

first_layer = Encoder._modules['layer1']

and then you can either take out the layers that you need from your resnet or you can just change the layers.

resnet = torchvision.models.__dict__(resnet50)(pretrained=true)
final_layer = resnet._modules['layer4']

and then pass combine these two. Hope this helps :slight_smile:

5

I will try. Thanks for your help