this post was submitted on 25 Nov 2023
1 points (100.0% liked)
Machine Learning
1 readers
1 users here now
Community Rules:
- Be nice. No offensive behavior, insults or attacks: we encourage a diverse community in which members feel safe and have a voice.
- Make your post clear and comprehensive: posts that lack insight or effort will be removed. (ex: questions which are easily googled)
- Beginner or career related questions go elsewhere. This community is focused in discussion of research and new projects that advance the state-of-the-art.
- Limit self-promotion. Comments and posts should be first and foremost about topics of interest to ML observers and practitioners. Limited self-promotion is tolerated, but the sub is not here as merely a source for free advertisement. Such posts will be removed at the discretion of the mods.
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Libraries like PyTorch and Jax are already high level libraries in my view. The low level stuff is C++/CUDA/XLA.
I don’t really see the useful extra abstractions in Keras that would lure me to it.
If you use Jax with Keras you are eseentialy doing: keras->jax->jaxpr->llvm->cuda/xla, with probably many more intermediate levels
This.
What about not having to write your own training loop? Keras takes away a lot of boilerplate code, it makes your code more readable and less likely to contain bugs. I would compare it to scikit-learn: Sure, you can implement your own Random Forest, but why bother?
The reality is that you nearly always need to break into that training abstraction, and so it is useless.
As the others said, it's a pain to reimplement common layers in JAX (specifically). PyTorch is much higher level in it's nn API, but personally I despise rewriting the amazing training loop for every implementation. That's why even JAX uses Flax for common layers, because why use an error prone operator like jax.lax.conv_from_dilated or whatever and fill its 10 arguments every time? I would rather use flax.linen.Conv2D or keras_core.layers.Conv2D in my Sequential layer and prevent debugging a million times. For PyTorch, model.fit() can just quickly suffice and later customized.