The architecture used in the deep image prior

user_11670761 0
Viewed 90

From Deep Image Prior, there is a figure of the architecture used in the experiment, shown as follows: decoder-encoder architecture,

The figure mentions clearly what layers are to be used for each downsampling, upsampling, and skip networks.

My question is, are these layers in each network really represents the layers in the programming code?

For an easy understanding, let's take a look at the skip network. In the figure, it says the skip network uses convolution, BN, and LeakyReLU layers. But does the code really uses only these three layers to perform the skip network?

pytorch deep-learning computer-vision image-processing python
1 Answers

Their code is indeed not the easiest to read. If you look closely you'll notice that they use slightly different networks to accomplish each task.
If you wish to look at code and models, you might find the code of "Double DIP", a follow-up work recently published, to be more readable.

In my experience, the "inductive bias" effect shown by Deep Image Prior is quite robust to the specific details of the net architecture. As long as you have convolution layers at multiple scales you will observe this effect.

user_17144100
Related