I've got a JAX version if you're interested in checking it out!https://github.com/ericjang/nf-jax
I implemented some normalizing flows yesterday (NICE, RealNVP, MAF, IAF), tried to make core of it somewhat clean in case helpful https://github.com/karpathy/pytorch-normalizing-flows … I like how flow layers can be structured similar to backprop, each needs an invert() and emits a log det J "regularization"pic.twitter.com/F8AgiCb2RP
-
-
-
hey Eric, yep seeing your nf-jax is what finally made me pop the item off my todo list. I don't think I've really "warmed up" to the jax style of doing things, it seems wordier, clunkier and a bit less obvious/explicit. I do want to part of the hip camp though, will try harder.
- Još 2 druga odgovora
Novi razgovor -
-
-
For more PyTorch-only dependency in InvConv1x1, I think PyTorch's equivalent of scipy.linalg.qr init would be "torch.nn.init.orthogonal_(Q)" and that of http://scipy.linalg.lu is " w_p, w_l, w_u = http://torch.lu _unpack(*http://Q.lu ())"
-
fixed with commit earlier today, ty for tip!
- Još 2 druga odgovora
Novi razgovor -
-
-
Thank you! Together with the excellent exposition in http://arxiv.org/abs/1912.02762 I wonder whether it’s time for a
@distillpub article. O_oHvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
Can someone explain what is meant by "flow" here, and secondly what makes this a "normalized flow"?
-
Check out section 2.1 of http://arxiv.org/abs/1912.02762 for an explanation of both of these things.
Kraj razgovora
Novi razgovor -
-
-
Sir, As I don't precisely understand; is at issue here not simply accuracy to alpha on, [choose your variable] i.e. "consciousness". But! /Boundary/ and /sphereicization/ in self-permit output! /Sir.../ /Prediction.../ /Accuracy.../ /Disclosure.../ _Sir._ You! A Tesla employee!
– mjesto: Moorings -
"Today's http://Random.org , /atmospheric disturbance, generated integers are...!/
(
)
https://photos.app.goo.gl/XDPZpRVjazgaXte59 …
Interpret:
https://play.google.com/store/apps/details?id=me.conversio.letterlevel3&hl=en …
{
<
>
}
Kraj razgovora
Novi razgovor -
-
-
Thanks for the code! Do you have any plans to add ND case as well, maybe with an example of flatten MNIST?
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.