How to load the fc weights into its corresponding convolution layer weights?

vision
1

I know how to convert the fc layer to convolution layer.
But its weights are initialized randomly.
I want to load the pretrained fc weights to the convolution layer transformed from this fc layer.
How should I do?

2

For the FC weights, you can use view. For example, let old_fc6_weight be the fc6 weight loaded from the state_dict of the pre-trained model:
new_fc6_param.data = old_fc6_weight.view(new_fc6_param.size())

1 Like
3

Thank you for your kind answer.
It works!
For convenient reference for others, I post my concrete code here.

self.fc6_to_conv = nn.Conv2d(512, 4096, kernel_size = 7, stride = 1, padding = 0)
self.fc6_to_conv.weight.data = vgg16.fc6.weight.data.view(self.fc6_to_conv.weight.size())
1 Like
4

However, it seems difficult to train.
An I find that it’s not necessary to train with converted conv layer.
I just need to convert fc layer to conv layer when testing. Am I right?

5

If your output is spatial (as in an FCN), then the fully connected layers must be made fully convolutional in order to preserve spatial dimensions (during training and testing).

6

Does loading the pretrained fc weights to weights to conv layers work? Seems unintuitive