What exactly is a ‘dynamic unit’. It is mentioned in Dynamic Unet and it is also in the fast.ai library
Hey Rahul, you may check it from the source under models directory.
I did check that out. It says “A dynamic implementation of Unet architecture, because calculating connections and channels suck!. When an encoder is passed, this network will automatically construct a decoder after the first single forward pass for any given encoder architecture.”
But I didn’t understand it. I have implemented a unet for brain tumor segmentation and want to try and better it.
All it does is: it takes pretrained networks (or any custom ones) like vgg,inception, resnet as backbone and automatically create the decoder part to make it like a unet. If you implemented a unet it shouldn’t be hard to grasp the idea since it’s essentially the same thing only the encoder is a pretrained network which essentially allows you to apply transfer learning. You may also check https://github.com/ternaus/TernausNet