NumPy random generators #1267
NumPy random generators #1267
Conversation
|
@albi3ro, I'll attach the |
Codecov Report
@@ Coverage Diff @@
## master #1267 +/- ##
=======================================
Coverage 98.14% 98.14%
=======================================
Files 147 147
Lines 11223 11242 +19
=======================================
+ Hits 11015 11034 +19
Misses 208 208
Continue to review full report at Codecov.
|
albi3ro
added
review-ready 👌
albi3ro
removed
the
review-ready 👌
pennylane/numpy/random.py
Outdated
| @@ -15,7 +15,44 @@ | |||
| This package provides a wrapped version of autograd.numpy.random, such that | |||
| it works with the PennyLane :class:`~.tensor` class. | |||
| """ | |||
| from packaging.version import parse as parse_version | |||
There was a problem hiding this comment.
Is the packaging library a part of the Python standard library? I was a bit confused, since it's available on PyPI, but at the same time, you didn't need to add it to the requirements.txt 🤔
This is nice to know though, in other parts of the library we depend on the semantic_version==2.6 package for this!
| from autograd.numpy import random as _random | ||
| from .wrapper import wrap_arrays | ||
| from numpy import __version__ as np_version | ||
| from numpy.random import MT19937, PCG64, Philox, SFC64 # pylint: disable=unused-import |
There was a problem hiding this comment.
How come this is imported if not used?
There was a problem hiding this comment.
We want these in the namespace, so someone can call qml.numpy.random.MT19937, instead of having to call qml.numpy.random._random.MT19937
|
|
||
| for name in dir(_random.Generator): | ||
| if name[0] != "_": | ||
| self.__dict__[name] = tensor_wrapper(getattr(super(), name)) |
There was a problem hiding this comment.
Nice! I tested this out, it works really well.
There was a problem hiding this comment.
I had a look through all methods - it looks like most of them do return ndarray objects, so it should be good 👍 Are we inadvertently wrapping any that don't return arrays?
There was a problem hiding this comment.
The methods will return a single scalar if not called with a size. If a size is provided, then it returns an ndarray. I've thought about explicitly writing out the names of all the generators that need to be wrapped, but decided against it.
| mat1_2 = rng1.normal(size=size) | ||
| mat2_2 = rng2.normal(size=size) | ||
|
|
||
| assert np.all(mat1_2 == mat2_2) |
Co-authored-by: Josh Izaac <josh146@gmail.com>
Currently, the recommended way of creating random numbers with NumPy involves creating a local generator object, and then querying that object for random numbers
Currently, we cannot use this method with PennyLane wrapped NumPy. This PR adds a wrapped generator object and
default_rnggenerator constructor to PennyLane's NumPy random. It also import the bitstring generators into the random namespace. Now we can calland get an array with a
requires_gradattribute,