Copy weights from Conv2D with bias=True to Conv2D with bias=False

vision
1

Hi,

I have a model that is a sequence of Conv2D with bias=True, I have another model which also is a sequence of Conv2D but with bias=False.

How can I copy the weights from the first model (incorporating the bias) to the second model that just has weights?

Any pointers are greatly appreciated.

Thanks!

2

Yes, you can skip the additional parameters by using strict=False in load_state_dict as shown here:

conv_bias = nn.Conv2d(3, 3, 3, 1, 1, bias=True)
conv_nobias = nn.Conv2d(3, 3, 3, 1, 1, bias=False)

conv_nobias.load_state_dict(conv_bias.state_dict(), strict=False)
> _IncompatibleKeys(missing_keys=[], unexpected_keys=['bias'])