In operations_broadcast_test.go there are some tests that are not yet filled in. The point is to test that broadcasting works for different shapes. The semantics of broadcast probably isn't clear, so please do send me a message for anything.

This is a good first issue for anyone looking to get interested

I'm using TF 2.0, and I get this error when I import tangent, due to a list of non-differentiable functions that includes tf.to_float (line 60), which is deprecated:

https://www.tensorflow.org/versions/r1.14/api_docs/python/tf/to_float

I found that function mod2pi is not implemented yet, but mod works. Is there any list of implemented functions? Minimal working example is:

using Zygote

# This is working
gradient(x -> mod(x, 2pi), 1.)

# This is not
gradient(x -> mod2pi(x), 1.)

Summary:

The functions for the categorical distribution only accept a column vector, it would be great if it could accept also row vectors.

Description:

I use the categorical distribution to go over a matrix N_obs x N_probabilities, so it's more natural for me to use row vectors than column vectors.

Current functions:

real categorical_lpmf(ints y | vector theta)
real

Debugging Kotlin∇ code within IntelliJ IDEA can be somewhat cumbersome due to the functional API structure (lots of deeply-nested stack traces and context switching). To facilitate more user-friendly debugging, we should add support for visual debugging by exposing Kaliningraph’s built-in graph visualization capabilities. For example, the use

The init module has been deprecated, and the recommend approach for generating initial weights is to use the Template.shape method:

>>> from pennylane.templates import StronglyEntanglingLayers
>>> qml.init.strong_ent_layers_normal(n_layers=3, n_wires=2) # deprecated
>>> np.random.random(StronglyEntanglingLayers.shape(n_layers=3, n_wires=2))  # new approach

We should upd

The function can be easily implemented, but for the sake of getting closer to the numpy API we should add the function to aesara.tensor:

https://numpy.org/doc/stable/reference/generated/numpy.logaddexp.html

At the moment profiles.h (in pkg/profiles) lacks many (any?) comments. Also lots of variables are declared somewhat separately from where they are associated with heap storage.

Both these make it a bit hard to read.

It would be nicer if it was called PROFILES.h too.

Some of them can be ported over from Zygote.

cf. FluxML/Zygote.jl#906
https://github.com/FluxML/Zygote.jl/blob/956cbcf3c572c0eb09c146189bb38b1b434634ff/src/lib/array.jl#L130