I use Resnet-50 as backbone in my person re-identificator. However, I have a problem. The problem is that stride of the last down sample layer is 2 but I want to change it to one and still use pre-trained weights and during training I want to learn new weights.
So if I understood it correctly, the image is being downsampled by a factor of 2 and you want to change it 1 so as to keep the input shape preserved. I’m not sure where the last downsample layer is however, it would be before some fully connected layer. If you are planning on reusing the said fully connected layers’ weights, you’ll have to re-train the network. You won’t be able to reuse the weights from the downsample layer onwards.
I want to use pre-trained weights before fully connected, I use resnet as a backbone, so I don’t need FC layer.