expressive 3.2.20250425

A library for quickly applying symbolic expressions to NumPy arrays

expressive

A library for quickly applying symbolic expressions to NumPy arrays

By enabling callers to front-load sample data, developers can move the runtime cost of Numba's JIT to the application's initial loading (or an earlier build) and also avoid exec during runtime, which is otherwise needed when lambdifying symbolic expressions

Inspired in part by this Stack Overflow Question Using numba.autojit on a lambdify'd sympy expression

installation

via pip https://pypi.org/project/expressive/

pip install expressive

usage

refer to tests for examples for now

generally follow a workflow like

• create instance expr = Expressive("a + log(b)")
• build instance expr.build(sample_data)
• instance is now callable expr(full_data)

the data should be provided as dict of NumPy arrays

data_sample = {  # simplified data to build and test expr
   "a": numpy.array([1,2,3,4], dtype="int64"),
   "b": numpy.array([4,3,2,1], dtype="int64"),
}
data = {  # real data user wants to process
   "a": numpy.array(range(1_000_000), dtype="int64"),
   "b": numpy.array(range(1_000_000), dtype="int64"),
}
E = Expressive("expr")  # string or SymPy expr
E.build(data_sample)  # data's types used to compile a fast version
E(data)  # very fast callable

simple demo

import time
import contextlib
import numpy
import matplotlib.pyplot as plt
from expressive import Expressive

# simple projectile motion in a plane
E_position = Expressive("y = v0*t*sin(a0) + 1/2(g*t^2)")

# expr is built early in the process runtime by user
def build():
    # create some sample data and build with it
    # the types are used to compile a fast version for full data
    data_example = {
        "v0": 100,  # initial velocity m/s
        "g": -9.81, # earth gravity m/s/s
        "a0": .785,  # starting angle ~45° in radians
        "t": numpy.linspace(0, 15, dtype="float64"),  # 15 seconds is probably enough
    }
    assert len(data_example["t"]) == 50  # linspace default
    time_start = time.perf_counter()
    E_position.build(data_example)  # verify is implied with little data
    time_run = time.perf_counter() - time_start

    # provide some extra display details
    count = len(data_example["t"])
    print(f"built in {time_run*1000:.2f}ms on {count:,} points")
    print(f"  {E_position}")

def load_data(
    point_count=10**8,  # 100 million points (*count of angles), maybe 4GiB here
    initial_velocity=100,  # m/s
):
    # manufacture lots of data, which would be loaded in a real example
    time_array = numpy.linspace(0, 15, point_count, dtype="float64")
    # collect the results
    data_collections = []
    # process much more data than the build sample
    for angle in (.524, .785, 1.047):  # initial angles (30°, 45°, 60°)
        data = {  # data is just generated in this case
            "v0": initial_velocity,  # NOTE type must match example data
            "g": -9.81, # earth gravity m/s/s
            "a0": angle,  # radians
            "t": time_array,  # just keep re-using the times for this example
        }
        data_collections.append(data)

    # data collections are now loaded (created)
    return data_collections

# later during the process runtime
# user calls the object directly with new data
def runtime(data_collections):
    """ whatever the program is normally up to """

    # create equivalent function for numpy compare
    def numpy_cmp(v0, g, a0, t):
        return v0*t*numpy.sin(a0) + 1/2*(g*t**2)

    # TODO also compare numexpr demo

    # call already-built object directly on each data
    results = []
    for data in data_collections:
        # expressive run
        t_start_e = time.perf_counter()  # just to show time, prefer timeit for perf
        results.append(E_position(data))
        t_run_e = time.perf_counter() - t_start_e

        # simple numpy run
        t_start_n = time.perf_counter()
        result_numpy = numpy_cmp(**data)
        t_run_n = time.perf_counter() - t_start_n

        # provide some extra display details
        angle = data["a0"]
        count = len(data["t"])
        t_run_e = t_run_e * 1000  # convert to ms
        t_run_n = t_run_n * 1000
        print(f"initial angle {angle}rad ran in {t_run_e:.2f}ms on {count:,} points (numpy:{t_run_n:.2f}ms)")

    # decimate to avoid very long matplotlib processing
    def sketchy_downsample(ref, count=500):
        offset = len(ref) // count
        return ref[::offset]

    # display results to show it worked
    for result, data in zip(results, data_collections):
        x = sketchy_downsample(data["t"])
        y = sketchy_downsample(result)
        plt.scatter(x, y)
    plt.xlabel("time (s)")
    plt.ylabel("position (m)")
    plt.show()

def main():
    build()
    data_collections = load_data()
    runtime(data_collections)

main()

compatibility matrix

generally this strives to only rely on high-level support from SymPy and Numba, though Numba has stricter requirements for NumPy and llvmlite

Python	Numba	NumPy	SymPy	commit	coverage	trun
3.7.17	0.56.4	1.21.6	1.6	845e68c	{'codegen.py': '🟠 99% m 387'} 🟢 100% (9path)	86s
3.8.20	0.58.1	1.24.4	1.7	845e68c	{'codegen.py': '🟠 99% m 387'} 🟢 100% (9path)	83s
3.9.19	0.53.1	1.23.5	1.7	845e68c	{'codegen.py': '🟠 99% m 387'} 🟢 100% (9path)	79s
3.9.19	0.60.0	2.0.1	1.13.2	845e68c	🟢 100% (10path)	84s
3.10.16	0.61.0	2.1.3	1.13.3	845e68c	🟢 100% (10path)	81s
3.11.11	0.61.0	2.1.3	1.13.3	845e68c	🟢 100% (10path)	81s
3.12.7	0.59.1	1.26.4	1.13.1	845e68c	🟢 100% (10path)	74s
3.12.8	0.61.0	2.1.3	1.13.3	845e68c	🟢 100% (10path)	87s
3.13.1	0.61.0	2.1.3	1.13.3	845e68c	🟢 100% (10path)	89s

NOTE differences in test run times are not an indicator of built expr speed, more likely the opposite and more time spent represents additional build step effort, likely improving runtime execution! please consider the values arbitrary and just for development reasons

further compatibility notes

these runs build the package themselves internally, while my publishing environment is currently Python 3.11.2

though my testing indicates that this works under a wide variety of quite old versions of Python/Numba/SymPy, upgrading to the highest dependency versions you can will generally be best

• Python 3 major version status https://devguide.python.org/versions/
• https://numba.readthedocs.io/en/stable/release-notes-overview.html

NumPy 1.x and 2.0 saw some major API changes, so older environments may need to adjust or discover working combinations themselves

• some versions of Numba rely on numpy.MachAr, which has been deprecated since at least NumPy 1.22 and may result in warnings

TBD publish multi-version test tool

testing

Only docker is required in the host and used to generate and host testing

sudo apt install docker.io  # debian/ubuntu
sudo usermod -aG docker $USER
sudo su -l $USER  # login shell to self (reboot for all shells)

Run the test script from the root of the repository and it will build the docker test environment and run itself inside it automatically

./test/runtests.sh

build + install locally

Follows the generic build and publish process

• https://packaging.python.org/en/latest/tutorials/packaging-projects/#generating-distribution-archives
• build (builder) https://pypi.org/project/build/

python3 -m build
python3 -m pip install ./dist/*.whl

contributing

The development process is currently private (though most fruits are available here!), largely due to this being my first public project with the potential for other users than myself, and so the potential for more public gaffes is far greater

Please refer to CONTRIBUTING.md and LICENSE.txt and feel free to provide feedback, bug reports, etc. via Issues, subject to the former

additional future intentions for contributing

• improve internal development history as time, popularity, and practicality allows
• move to parallel/multi-version/grid CI over all-in-1, single-version dev+test container (partially done with 2.0.0!)
• greatly relax dependency version requirements to improve compatibility
• publish majority of ticket ("Issue") history

version history

v3.2.20250425

• improved smaller types handling
  • automatic dtype determination with Pow() is improved
  • give a dedicated warning when an exception related to setting dtype_result to a type with a small width that a function (such as Pow()) automatically promotes occurs
• improve autobuilding experience with new config tunables
  • easily enable autobuild globally builder.autobuild.allow_autobuild
  • option to disable build-time usage warning builder.autobuild.usage_warn_nag
• minor version is now a datestamp

v3.1.0

• instances of Expressive now have individual configurations
• further config changes
  • all configuration keys are now flattened and .-separated
  • warn and still handle legacy keys
  • include per-instance builder settings
  • new UNSET value singleton

v3.0.0

• cutover splitting project into numerous files (dbd89cd+)
  • improved MathJax reference copy handling
  • split out changelog too (truncated view in README)
• add tunables for parallelization and numba.prange() support
• improved a bug where the name literally "row" couldn't be used in Sum() (now has a dedicated error and uses "rowscalar" name, a future version should avoid this entirely via by-ref handling and/or name mangling)
• testing changes
  • various pathing and import changes to accomodate file changes
  • downgrade to non-root user in test containers
  • essential argument features (verbose, trap EXIT to shell, subset to single TestCase or TestCase.test_)

v2.2.0

• added support for the Sum function (SymPy unevaluated summation)
  • attempts to evaluate/decompose Sum into an algebraic expression during building .build()
  • creates a custom function to manage Sum instances which can't be simplified
  • spawn a thread to warn user when attempting to simplify a Sums is taking an excessive amount of time (duration and even halting are unknown, so the user may not know where the issue is .. 20s default)
• added basic configuration system CONFIG
  • API is unstable and largely featureless, but needed to control/disable Sum simplifying
  • currently a singleton dict shared by all Expressive instances, but a future version/design will accept per-instance configurations and combine them with global defaults
• generally much better handling for scalars in data
  • scalar values are no longer coerced into a 0-dim array
  • NumPy scalars (not just Python numbers) are now allowed

v2.1.0 (unreleased)

• added a new signature_automatic() which (ab)uses the Numba JIT to determine the result's dtype even for indexed exprs

v2.0.0

• enabled matrix/tensor support
• improved/reduced warnings from verify
• tested + greatly reduced dependency version requirements
• added a basic usage example (uses new docs repo https://gitlab.com/expressive-py/docs/ )

complete at CHANGELOG.md