I would like to test FeatUp as an upsampling module in the decoder of a UNet like architecture instead of transposed convolution or upsampling + convolution.
Is it feasible at the moment ? I feel like the current code only allow to use the upsamplers on a pretrained and locked backbone. In my case the upsampler need to be optimized at the same time as the entire network.
If it is feasible, which upsampler is adapted to this kind of task ? From what I understood, it seems that I must choose JBU as the Implicit upsampler is trained on each image at inference time
I would like to test FeatUp as an upsampling module in the decoder of a UNet like architecture instead of transposed convolution or upsampling + convolution.
Is it feasible at the moment ? I feel like the current code only allow to use the upsamplers on a pretrained and locked backbone. In my case the upsampler need to be optimized at the same time as the entire network.
If it is feasible, which upsampler is adapted to this kind of task ? From what I understood, it seems that I must choose JBU as the Implicit upsampler is trained on each image at inference time