Hi,
I was wondering if there’s a way I can reverse the ResNet architecture to increase the output size instead of reducing it. I want to use for an encoder-decoder problem. I know fastai has DynamicUnet but for learning purposes, I wanted to ask.
Please let me know.
If u wanna increase your input size you have to use a layer called ConvTranspose …u can create an upsampling resnet with quite easily with the current architecture. But fastai’s dynamic unet has some activations hooked from the encoder of the unet and concatenated to the upsampling path. And there are also some advance upsampling layers like PixelShuffleICNR which is used by default in the dynamic unet.