Book 2 (load/save) #259
Book 2 (load/save) #259
Conversation
candle-book/src/guide/hello_world.md
Outdated
| use candle::{DType, Device, Result, Tensor}; | ||
|
|
||
| struct Model { | ||
| first: Tensor, |
There was a problem hiding this comment.
Maybe it makes more sense to use candle_nn::Linear here?
Or maybe we really want to go back to basics here in which case we should mention that Linear exists but for the sake of this tutorial we won't use it here and have explicit tensors instead.
There was a problem hiding this comment.
Great idea. I was wondering how to complexify this example without delving into too much details to keep it hello world like.
I will keep this example that simple, and expand by making your own Linear, then referring to candle_nn for layers in general.
For Conv1d and Linear for instance, wdyt ?
There was a problem hiding this comment.
BTW, this comment is more about #258 (This PR I converted to draft because it's more about the save/load surface to get simple cheatsheet more complete on that front.
There was a problem hiding this comment.
Sounds good, you could even even not have backprop/gradient descent for a tensor 101 though maybe it's too simplistic. But certainly good to only introduce candle_nn later in the process (so that users can understand that there is no magic behind the hood there).
There was a problem hiding this comment.
Yup, I want to keep the training loop in a different place.
IMO after having hello world + there's no magic you have several options:
- Run a real model (90% of people)
- Train a model (10% of people)
It's an old stat, we're closer to 99% vs 1% now that AI is much more mainstream (but there's still a lot of finetuners out there, mostly using script already written by people)
Narsil
changed the title
Narsil
changed the title
Depends on #258 Drafted until merge.