Arraymancer
Arraymancer copied to clipboard
mratsim
A fast, ergonomic and portable tensor library in Nim with a deep learning focus for CPU, GPU and embedded devices via OpenMP, Cuda and OpenCL backends
Nim added hot-code reloading support at the language level. Arraymancer should be tested with it to: - show how to use it in a hot code reload setting - ensure...
Now that Travis supports ARM, it would be useful to test it to ensure Arraymancer is usable on low-powered devices and smartphones.
See: https://www.reddit.com/r/MachineLearning/comments/dla75r/d_retrain_your_models_the_adam_optimizer_in/ https://github.com/pytorch/pytorch/pull/22628
from #384 > While am at adding new features to PCA, I'm considering a couple of additional options: > > - Adding a preprocess enum to the signature: > ```Nim...
Instead of using `ValueError` and `IndexError` everywhere it is probably more informative to introduce more descriptive error types. For example: - IncompatibleShapeError (for tensors that can't be multiplied together) -...
As mentioned in https://github.com/nim-lang/Nim/issues/12036#issuecomment-524890898 emit interpolation does not use backtick, unlike quote do and template identifier injection. A thorough review of emit interpolation should be done to avoid future breakage....
Edited: OpenMP -> Archlinux, I never run the full test suite on Arch due to the impossibility to do "nimble test -d:blas=cblas" as contrary to Debian/Ubuntu and Travis, the BLAS...
Two variants of already implemented Arraymancer optimizers have been gaining traction recently, specifically the AdamW and SGDW optimizers. Both of these were proposed in the 2017 paper [_Decoupled Weight Decay...
Currently only stochastic gradient descent is supported, at the very minimum it would be nice to support: - [ ] RMSProp - [x] Adam - [x] SGD with Momentum -...
After #363 we have einsum, but we can't backprop through it. Einsum backprop implementation: - https://github.com/rsokl/MyGrad/blob/5ea7b929108c90a83c37c3d619cef0a6f6596e28/mygrad/linalg/ops.py#L161-L280 - https://github.com/tensorflow/tensorflow/blob/eb9b2659f831c5ac60eb833dc4d04a161e588833/tensorflow/python/ops/special_math_ops.py#L137-L157 I'm not too sure how it works in PyTorch, seems magical: https://github.com/pytorch/pytorch/pull/6307/files