What is a simple code example using uarray that demonstrates "replatforming SciPy"?

[hameerabbasi]

Can the simple example be used to illustrate:

• dispatch to an optimized implementation
• "symbolic numpy"
• implementation over getitem and iterate

[hameerabbasi]

Here's my take on it:

import uarray as ua
import uarray.numpy as np
import uarray.scipy.linalg as spl

# Create array `x` in TensorFlow/PyTorch/Dask/CuPy

a, b, c = spl.svd(x)

# a, b, c is a TensorFlow graph/CuPy array/whatever

with ua.use_graph(True): # Other options are False and None
    # a, b, c are graphs
    a, b, c = spl.svd(x)

# a, b, c are arrays

# Backends

# Can be used as context manager or to set default
with ua.use_backend(ua.backends.PyTorch, coerce=True):
    # Everything here will use PyTorch and coerce all arrays.

with ua.use_backend(ua.backends.PyTorch, ua.backends.NumPy):
   # Attempts to use PyTorch, and failing that, NumPy

with ua.use_backend(ua.backends.PyTorch, ua.backends.NumPy, coerce=ua.backends.PyTorch):
    # When multiple backends are specified, you must specify the one to coerce to on failure of type check.

Can the simple example be used to illustrate:

• dispatch to an optimized implementation
  • Yes. Dispatch is performed over type or other characteristics.
• "symbolic numpy"
  • Yes, a graph can be built and reduced, but this can be another backend or a first-class feature.
• implementation over getitem and iterate
  • Yet another backend.

[teoliphant]

# Create array `x` in TensorFlow/PyTorch/Dask/CuPy

a, b, c = spl.svd(x)

Yes, yes!  This is the fundamental use-case.  All functions return the kind of array they pull in (if they are all the same.  If they aren't all the same, then it raises an error). 

# a, b, c is a TensorFlow graph/CuPy array/whatever

with ua.use_graph(True): # Other options are False and None
    # a, b, c are graphs
    a, b, c = spl.svd(x)

 # I like this. This adds additional value where you can get a graph.  A decorator around a function that  
 # will produce a graph on the return value could also work.

# a, b, c are arrays

# Backends

# Can be used as context manager or to set default
with ua.use_backend(ua.backends.PyTorch, coerce=True):
    # Everything here will use PyTorch and coerce all arrays.

with ua.use_backend(ua.backends.PyTorch, ua.backends.NumPy):
   # Attempts to use PyTorch, and failing that, NumPy

with ua.use_backend(ua.backends.PyTorch, ua.backends.NumPy, coerce=ua.backends.PyTorch):
    # When multiple backends are specified, you must specify the one to coerce to on failure of type check.

 # These are interesting ideas.  I like them.

[teoliphant]

Yes, this is what the API should look like. This is what we need.

[ashwinvis]

@hameerabbasi Thank you for that simple example demonstrating the vision... I believe it is already outdated now that unumpy is a separate package. Some documentation would be nice for outsiders to understand what is working now and what's on the roadmap.

[hameerabbasi]

@ashwinvis For now I have this simple notebook up. I've attempted to get most ufuncs and ufunc-related reductions (sum, prod, any, all, min, max) working so far with Xnd, PyTorch and NumPy.

You can take a look at https://github.com/Quansight-Labs/uarray/blob/master/unumpy/tests/test_numpy.py to see what's supported so far, as far as the user-facing API goes.

As far as the dev-facing API, here's a notebook demonstrating it: https://nbviewer.jupyter.org/github/Quansight-Labs/uarray/blob/master/notebooks/basic_tutorial.ipynb

I do plan to write docs once I have enough of the API implemented.

[hameerabbasi]

Docs are up at https://uarray.readthedocs.io/en/latest/