Speed up lambdifying

import logging

import ampform
import graphviz
import qrules
import sympy as sp
from ampform.dynamics.builder import create_relativistic_breit_wigner_with_ff
from IPython.display import HTML, SVG

from tensorwaves.model import (
    LambdifiedFunction,
    SympyModel,
    optimized_lambdify,
    split_expression,
)

logger = logging.getLogger()
logger.setLevel(logging.ERROR)

Split expression

Lambdifying a SymPy expression can take rather long when an expression is complicated. Fortunately, TensorWaves offers a way to speed up the lambdify process. The idea is to split up an an expression into sub-expressions, separate those separately, and then recombining them. Let’s illustrate that idea with the following simplified example:

x, y, z = sp.symbols("x:z")
expr = x**z + 2 * y + sp.log(y * z)
expr

This expression can be represented in a tree of mathematical operations.

dot = sp.dotprint(expr)
graphviz.Source(dot)

The function split_expression() can now be used to split up this expression tree into a ‘top expression’ plus definitions for each of the sub-expressions into which it was split:

top_expr, sub_expressions = split_expression(expr, max_complexity=3)

top_expr

sub_expressions

The original expression can easily be reconstructed with subs() or xreplace():

top_expr.xreplace(sub_expressions)

Each of the expression trees are now smaller than the original:

dot = sp.dotprint(top_expr)
graphviz.Source(dot)

for symbol, definition in sub_expressions.items():
    dot = sp.dotprint(definition)
    graph = graphviz.Source(dot)
    graph.render(filename=f"sub_expr_{symbol.name}", format="svg")

html = "<table>\n"
html += "  <tr>\n"
html += "".join(
    '    <th style="text-align:center;'
    f' background-color:white">{symbol.name}</th>\n'
    for symbol in sub_expressions
)
html += "  </tr>\n"
html += "  <tr>\n"
for symbol in sub_expressions:
    svg = SVG(f"sub_expr_{symbol.name}.svg").data
    html += f'    <td style="background-color:white">{svg}</td>\n'
html += "  </tr>\n"
html += "</table>"
HTML(html)

Optimized lambdify

Generally, the lambdify time scales exponentially with the size of an expression tree. With larger expression trees, it’s therefore much faster to lambdify these sub-expressions separately and to recombine them. TensorWaves offers a function that does this for you: optimized_lambdify(). We’ll use an HelicityModel to illustrate this:

result = qrules.generate_transitions(
    initial_state=("J/psi(1S)", [+1]),
    final_state=["gamma", "pi0", "pi0"],
    allowed_intermediate_particles=["f(0)"],
)
model_builder = ampform.get_builder(result)
for name in result.get_intermediate_particles().names:
    model_builder.set_dynamics(name, create_relativistic_breit_wigner_with_ff)
model = model_builder.generate()

expression = model.expression.doit()
sorted_symbols = sorted(expression.free_symbols, key=lambda s: s.name)

%%time
lambdified_optimized = optimized_lambdify(
    sorted_symbols,
    expression,
    max_complexity=100,
)

%%time
sp.lambdify(sorted_symbols, expression)

Specifying complexity

In the usually workflow (see Usage), TensorWaves uses SymPy’s own lambdify() by default. You can change this behavior with the max_complexity argument of SympyModel:

sympy_model = SympyModel(
    expression=model.expression,
    parameters=model.parameter_defaults,
    max_complexity=100,
)

If max_complexity is specified (i.e., is not None), LambdifiedFunction uses TensorWaves’s optimized_lambdify().

%%time
intensity = LambdifiedFunction(sympy_model, backend="jax")