functorch
Flux.jl
functorch | Flux.jl | |
---|---|---|
11 | 22 | |
1,372 | 4,393 | |
0.4% | 0.4% | |
0.0 | 8.7 | |
1 day ago | 4 days ago | |
Jupyter Notebook | Julia | |
BSD 3-clause "New" or "Revised" License | GNU General Public License v3.0 or later |
Stars - the number of stars that a project has on GitHub. Growth - month over month growth in stars.
Activity is a relative number indicating how actively a project is being developed. Recent commits have higher weight than older ones.
For example, an activity of 9.0 indicates that a project is amongst the top 10% of the most actively developed projects that we are tracking.
functorch
-
What is the most efficient approach to ensemble a pytorch actor-critic model?
I would suggest checking https://pytorch.org/functorch/ and https://github.com/metaopt/torchopt for efficient inference and training with ensembles (e.g., t be independent actors in a multi-agent setting or multiple critics).
-
[P] Multidimensional array batch indexing for pytorch and numpy
There were some bugs still with advanced indexing in an older release of functorch, I believe they should be fixed now though: https://github.com/pytorch/functorch/pull/862
- Functorch: Jax-like composable function transforms for PyTorch
-
Jax vs. Julia (Vs PyTorch)
Tangentially related but there is an effort to get some of the features of JAX into PyTorch: https://pytorch.org/functorch/
-
[D] Current State of JAX vs Pytorch?
Fwiw, composable vmap and stuff like that have also been implemented in PyTorch now - see functorch :) https://github.com/pytorch/functorch
-
[D] Ideal deep learning library
Fwiw, it’s not like Pytorch’s design prevents function transformations from being implemented. See functorch for an example of grad/vmap function transforms: https://github.com/pytorch/functorch
-
[P] Made Some Pytorch Modules For Agent Systems
You may find vmap from functorch to be quite useful: https://github.com/pytorch/functorch
-
[D] Are you using PyTorch or TensorFlow going into 2022?
If you're interested in function transformations in PyTorch, try out functorch :) https://github.com/pytorch/functorch
- PyTorch: Where we are headed and why it looks a lot like Julia (but not exactly)
-
Show HN: How does Jax allocate memory on a TPU? An interactive C++ walkthrough
The pytorch programming model is just really hard to adapt to an XLA-like compiler. Imperative python code doesn't translate to an ML graph compiler particularly well; Jax's API is functional, so it's easier to translate to the XLA API. By contrast, torch/xla uses "lazy tensors" that record the computation graph and compile when needed. The trouble is, if the compute graph changes from run to run, you end up recompiling a lot.
I guess in Jax you'd just only apply `jax.jit` to the parts where the compute graph is static? I'd be curious to see examples of how this works in practice. Fwiw, there's an offshoot of pytorch that is aiming to provide this sort of API (see https://github.com/pytorch/functorch and look at eager_compilation.py).
(Disclaimer: I worked on this until quite recently.)
Flux.jl
- Julia 1.10 Released
-
What Apple hardware do I need for CUDA-based deep learning tasks?
If you are really committed to running on Apple hardware then take a look at Tensorflow for macOS. Another option is the Julia programming language which has very basic Metal support at a CUDA-like level. FluxML would be the ML framework in Julia. I’m not sure either option will be painless or let you do everything you could do with a Nvidia GPU.
-
[D] ClosedAI license, open-source license which restricts only OpenAI, Microsoft, Google, and Meta from commercial use
Flux dominance!
-
What would be your programming language of choice to implement a JIT compiler ?
I’m no compiler expert but check out flux and zygote https://fluxml.ai/ https://fluxml.ai/
-
Any help or tips for Neural Networks on Computer Clusters
I would suggest you to look into Julia ecosystem instead of C++. Julia is almost identical to Python in terms of how you use it but it's still very fast. You should look into flux.jl package for Julia.
-
[D] Why are we stuck with Python for something that require so much speed and parallelism (neural networks)?
Give Julia a try: https://fluxml.ai
-
Deep Learning With Flux: Loss Doesn't Converge
2) Flux treats softmax a little different than most other activation functions (see here for more details) such as relu and sigmoid. When you pass an activation function into a layer like Dense(3, 32, relu), Flux expects that the function is broadcast over the layer's output. However, softmax cannot be broadcast as it operates over vectors rather than scalars. This means that if you want to use softmax as the final activation in your model, you need to pass it into Chain() like so:
-
“Why I still recommend Julia”
Can you point to a concrete example of one that someone would run into when using the differential equation solvers with the default and recommended Enzyme AD for vector-Jacobian products? I'd be happy to look into it, but there do not currently seem to be any correctness issues in the Enzyme issue tracker that are current (3 issues are open but they all seem to be fixed, other than https://github.com/EnzymeAD/Enzyme.jl/issues/278 which is actually an activity analysis bug in LLVM). So please be more specific. The issue with Enzyme right now seems to moreso be about finding functional forms that compile, and it throws compile-time errors in the event that it cannot fully analyze the program and if it has too much dynamic behavior (example: https://github.com/EnzymeAD/Enzyme.jl/issues/368).
Additional note, we recently did a overhaul of SciMLSensitivity (https://sensitivity.sciml.ai/dev/) and setup a system which amounts to 15 hours of direct unit tests doing a combinatoric check of arguments with 4 hours of downstream testing (https://github.com/SciML/SciMLSensitivity.jl/actions/runs/25...). What that identified is that any remaining issues that can arise are due to the implicit parameters mechanism in Zygote (Zygote.params). To counteract this upstream issue, we (a) try to default to never default to Zygote VJPs whenever we can avoid it (hence defaulting to Enzyme and ReverseDiff first as previously mentioned), and (b) put in a mechanism for early error throwing if Zygote hits any not implemented derivative case with an explicit error message (https://github.com/SciML/SciMLSensitivity.jl/blob/v7.0.1/src...). We have alerted the devs of the machine learning libraries, and from this there has been a lot of movement. In particular, a globals-free machine learning library, Lux.jl, was created with fully explicit parameters https://lux.csail.mit.edu/dev/, and thus by design it cannot have this issue. In addition, the Flux.jl library itself is looking to do a redesign that eliminates implicit parameters (https://github.com/FluxML/Flux.jl/issues/1986). Which design will be the one in the end, that's uncertain right now, but it's clear that no matter what the future designs of the deep learning libraries will fully cut out that part of Zygote.jl. And additionally, the other AD libraries (Enzyme and Diffractor for example) do not have this "feature", so it's an issue that can only arise from a specific (not recommended) way of using Zygote (which now throws explicit error messages early and often if used anywhere near SciML because I don't tolerate it).
So from this, SciML should be rather safe and if not, please share some details and I'd be happy to dig in.
- Flux: The Elegant Machine Learning Stack
-
Jax vs. Julia (Vs PyTorch)
> In his item #1, he links to https://discourse.julialang.org/t/loaderror-when-using-inter... The issue is actually a Zygote bug, a Julia package for auto-differentiation, and is not directly related to Julia codebase (or Flux package) itself. Furthermore, the problematic code is working fine now, because DiffEqFlux has switched to Enzyme, which doesn't have that bug. He should first confirm whether the problem he is citing is actually a problem or not.
> Item #2, again another Zygote bug.
If flux chose a buggy package as a dependency, that's on them, and users are well justified in steering clear of Flux if it has buggy dependencies. As of today, the Project.toml for both Flux and DiffEqFlux still lists Zygote as a dependency. Neither list Enzyme.
https://github.com/FluxML/Flux.jl/blob/master/Project.toml
What are some alternatives?
jax - Composable transformations of Python+NumPy programs: differentiate, vectorize, JIT to GPU/TPU, and more
Pytorch - Tensors and Dynamic neural networks in Python with strong GPU acceleration
nn - 🧑🏫 60 Implementations/tutorials of deep learning papers with side-by-side notes 📝; including transformers (original, xl, switch, feedback, vit, ...), optimizers (adam, adabelief, sophia, ...), gans(cyclegan, stylegan2, ...), 🎮 reinforcement learning (ppo, dqn), capsnet, distillation, ... 🧠
Knet.jl - Koç University deep learning framework.
onnx-simplifier - Simplify your onnx model
tensorflow - An Open Source Machine Learning Framework for Everyone
torch2trt - An easy to use PyTorch to TensorRT converter
Transformers.jl - Julia Implementation of Transformer models
BinaryBuilder.jl - Binary Dependency Builder for Julia
Torch.jl - Sensible extensions for exposing torch in Julia.
py2many - Transpiler of Python to many other languages
Lux.jl - Explicitly Parameterized Neural Networks in Julia