Adds in basic control net infrastructure and some tests #174
Conversation
|
What's left to do here? Should I run some GPU tests to see if everything carries through? |
| CliMAgen.ControlledNoiseConditionalScoreNetwork | ||
|
|
||
| The struct containing the parameters and layers | ||
| of the Noise Conditional Score Network architecture, |
There was a problem hiding this comment.
of the Controlled Noise Conditional Score Network
| end | ||
|
|
||
| # Mean processing of noised variable channels | ||
| if net.mean_bypass |
There was a problem hiding this comment.
should the embedded control/condition also be used by the mean bypass layer?
There was a problem hiding this comment.
(we do use the context in the mean bypass in our other version, though it's only the mean of the context that is used)
| "The layers of the network" | ||
| layers::NamedTuple | ||
| "A control network to condition the output of the U-net" | ||
| control_net::N |
There was a problem hiding this comment.
It seems like the output of the control net needs to be a vector of length embed_dim (same as for the time embedding). Is this correct? Since the control net is made exterior to the NCSN, where embed_dim is defined, we should maybe just add some documentation stating this, or maybe an @Assert statement somewhere, if we can?
|
oh, one other question. the control net itself is either pretrained or else can be trained with the rest of the network, it looks like. is that correct? |
We would like a basic control net setup that allows for custom conditioning of a pretrained network. Here, we build towards that end.