jax

Environment info

• transformers version: 4.11.3
• Platform: Linux-4.19.128-microsoft-standard-x86_64-with-glibc2.2.5
• Python version: 3.8.12
• PyTorch version (GPU?): 1.9.1+cu102 (False)
• Tensorflow version (GPU?): not install

numpy 1.20 and later supports several alternate normalizations for FFT functions, including "backward", "ortho", and "forward". We should support these in jax.numpy.fft as well.

Pyro's HMC and NUTS implementations are feature-complete and well-tested, but they are quite slow in models like the one in our Bayesian regression tutorial that operate on small tensors for reasons that are largely beyond our control (mostly having to do with the design and implementation of torch.autograd), which is unfortunate because these

Is there a way to train a bidirectional RNN (like LSTM or GRU) on trax nowadays?

TruncatedDistribution has both low and high. Why do TruncatedNormal and TruncatedCauchy only have low?

Probably similar to the other box colliders, where you vmap over the corners of the box and test if they are within a sphere/collider.

We should add binary_sparsemax_loss, the binary classification counterpart of multiclass_sparsemax_loss, see section 4.4 of https://arxiv.org/abs/1901.02324. This loss is also known modified Huber loss. The associated mapping is the projection on the unit cube so let's add projection_unit_cube, which is easy to implement using projection_box. Once done, they need to be added to the documen

PR #1554 introduced some docstring examples that are not contiguous and so break the usefulness of the copy button. These are good examples, and so should be turned into fully copyable examples.

Originally posted by @matthewfeickert in scikit-hep/pyhf#1554 (comment)

See wikipedia table for example patterns: https://en.wikipedia.org/wiki/List_of_integrals_of_Gaussian_functions