0
$\begingroup$

I want to use a hypernetwork on an entire vision backbone (39m parameters).

The hypernetwork structure looks like:

512 -> 512 -> 512 -> 39m

Unfortunately, the last layer means the hypernetwork has billions of parameters.

What could I do to reduce the parameter count of the hypernetwork?

(The 512-dimensional input vector is a noise vector that acts like a key that describes what details to focus on).

Improve this question
$\endgroup$
3
  • $\begingroup$ why do you need a hypernetwork for such a large model? what kind of task are you trying to solve? $\endgroup$
    – Alberto
    Commented Oct 28 at 21:24
  • $\begingroup$ 2 convnet encoders for multimodal embedding and 1 decoder for one of the modalities. The entire network is conditioned by a noise vector for embedding diversity under uncertainty. $\endgroup$
    – Richie Bendall
    Commented Oct 29 at 0:22
  • $\begingroup$ just pass the noise as part of the input $\endgroup$
    – Alberto
    Commented Oct 29 at 10:38

0

Reset to default

You must log in to answer this question.

Browse other questions tagged .