Merge pull request #185 from lincc-frameworks/math_eval

Create a BasicMathNode that can evaluate expressions given as a string

src/tdastro/math_nodes/basic_math_node.py

@@ -0,0 +1,185 @@
+"""Nodes that perform basic math operations that can be specified as strings.
+
+The goal of this library is to save users from needing to create a bunch of
+small FunctionNodes to perform basic math.
+"""
+
+import ast
+
+# Disable unused import because we need all of these imported
+# so they can be used during evaluation of the node.
+import math  # noqa: F401
+
+import jax.numpy as jnp  # noqa: F401
+import numpy as np  # noqa: F401
+
+from tdastro.base_models import FunctionNode
+
+
+class BasicMathNode(FunctionNode):
+    """A node that evaluates basic mathematical functions.
+
+    The BasicMathNode wraps Python's eval() function to sanitize the input string
+    and thus prevent the execution of arbitrary code. It also allows the user to write
+    the expression once and execute using math, numpy, or JAX. The names of the
+    variables in the expression must match the input variables provided by kwargs.
+
+    Example:
+        my_node = BasicMathNode(
+            "redshift + 10.0 * sin(phase)",
+            redshift=host.redshift,
+            phase=source.phase,
+        )
+
+    Attributes
+    ----------
+    expression : `str`
+        The expression to evaluate.
+    backend : `str`
+        The math libary to use. Must be one of: math, numpy, or jax.
+
+    Parameters
+    ----------
+    expression : `str`
+        The expression to evaluate.
+    backend : `str`
+        The math libary to use. Must be one of: math, numpy, or jax.
+    node_label : `str`, optional
+        An identifier (or name) for the current node.
+    **kwargs : `dict`, optional
+        Any additional keyword arguments. Every variable in the expression
+        must be included as a kwarg.
+    """
+
+    # A list of supported Python operations. Used to prevent eval from
+    # running arbitrary python expressions. The Call and Name types are special
+    # cased so we can do checks and translations.
+    _supported_ast_nodes = (
+        ast.Module,  # Top level object when parsed as exec.
+        ast.Expression,  # Top level object when parsed as eval.
+        ast.Expr,  # Math expressions.
+        ast.Constant,  # Constant values.
+        ast.Load,  # Load a variable - must come from an approved function or variable.
+        ast.Store,  # Store value - must come from an approved function or variable.
+        ast.BinOp,  # Binary operations
+        ast.Add,
+        ast.Sub,
+        ast.Mult,
+        ast.Div,
+        ast.FloorDiv,
+        ast.Mod,
+        ast.Pow,
+        ast.UnaryOp,  # Uninary operations
+        ast.UAdd,
+        ast.USub,
+        ast.Invert,
+    )
+
+    # A map from a very limited set of supported math constant/function names to
+    # the corresponding names in [math, numpy, jax]. This is needed because
+    # a very few functions have different names in different libraries.
+    _math_map = {
+        "abs": ["abs", "np.abs", "jnp.abs"],  # Special handling for math.
+        "acos": ["math.acos", "np.acos", "jnp.acos"],
+        "acosh": ["math.acosh", "np.acosh", "jnp.acosh"],
+        "asin": ["math.asin", "np.asin", "jnp.asin"],
+        "asinh": ["math.asinh", "np.asinh", "jnp.asinh"],
+        "atan": ["math.atan", "np.atan", "jnp.atan"],
+        "atan2": ["math.atan2", "np.atan2", "jnp.atan2"],
+        "cos": ["math.cos", "np.cos", "jnp.cos"],
+        "cosh": ["math.cosh", "np.cosh", "jnp.cosh"],
+        "ceil": ["math.ceil", "np.ceil", "jnp.ceil"],
+        "degrees": ["math.degrees", "np.degrees", "jnp.degrees"],
+        "deg2rad": ["math.radians", "np.deg2rad", "jnp.deg2rad"],  # Special handling for math
+        "e": ["math.e", "np.e", "jnp.e"],
+        "exp": ["math.exp", "np.exp", "jnp.exp"],
+        "fabs": ["math.fabs", "np.fabs", "jnp.fabs"],
+        "floor": ["math.floor", "np.floor", "jnp.floor"],
+        "log": ["math.log", "np.log", "jnp.log"],
+        "log10": ["math.log10", "np.log10", "jnp.log10"],
+        "log2": ["math.log2", "np.log2", "jnp.log2"],
+        "max": ["max", "np.max", "jnp.max"],  # Special handling for math
+        "min": ["min", "np.min", "jnp.min"],  # Special handling for math
+        "pi": ["math.pi", "np.pi", "jnp.pi"],
+        "pow": ["math.pow", "np.power", "jnp.power"],  # Special handling for numpy
+        "power": ["math.pow", "np.power", "jnp.power"],  # Special handling for math
+        "radians": ["math.radians", "np.radians", "jnp.radians"],
+        "rad2deg": ["math.degrees", "np.rad2deg", "jnp.rad2deg"],  # Special handling for math
+        "sin": ["math.sin", "np.sin", "jnp.sin"],
+        "sinh": ["math.sinh", "np.sinh", "jnp.sinh"],
+        "sqrt": ["math.sqrt", "np.sqrt", "jnp.sqrt"],
+        "tan": ["math.tan", "np.tan", "jnp.tan"],
+        "tanh": ["math.tanh", "np.tanh", "jnp.tanh"],
+        "trunc": ["math.trunc", "np.trunc", "jnp.trunc"],
+    }
+
+    def __init__(self, expression, backend="numpy", node_label=None, **kwargs):
+        if backend not in ["jax", "math", "numpy"]:
+            raise ValueError(f"Unsupported math backend {backend}")
+        self.backend = backend
+
+        # Check the expression is pure math and translate it into the correct backend.
+        self.expression = expression
+        self._prepare(**kwargs)
+
+        # Create a function from the expression. Note the expression has
+        # already been sanitized and validated via _prepare().
+        def eval_func(**kwargs):
+            return eval(self.expression, globals(), kwargs)
+
+        super().__init__(eval_func, node_label=node_label, **kwargs)
+
+    def __call__(self, **kwargs):
+        """Evaluate the expression."""
+        return eval(self.expression, globals(), kwargs)
+
+    def _prepare(self, **kwargs):
+        """Rewrite a python expression that consists of only basic math to use
+        the prespecified math library. Santizes the string to prevent
+        arbitrary code execution.
+
+        Parameters
+        ----------
+        **kwargs : `dict`, optional
+            Any additional keyword arguments, including the variable
+            assignments.
+
+        Returns
+        -------
+        tree : `ast.*`
+            The root node of the parsed syntax tree.
+        """
+        tree = ast.parse(self.expression)
+
+        # Walk the tree and confirm that it only contains the basic math.
+        for node in ast.walk(tree):
+            if isinstance(node, self._supported_ast_nodes):
+                # Nothing to do, this is a valid operation for the ast.
+                continue
+            elif isinstance(node, ast.Call):
+                # Check that function calls are only using items on the allow list.
+                if node.func.id not in self._math_map:
+                    raise ValueError(f"Unsupported function {node.func.id}")
+            elif isinstance(node, ast.Name):
+                if node.id in kwargs:
+                    # This is a user supplied variable.
+                    continue
+                elif node.id in self._math_map:
+                    # This is a math function or constant. Overwrite
+                    if self.backend == "math":
+                        node.id = self._math_map[node.id][0]
+                    elif self.backend == "numpy":
+                        node.id = self._math_map[node.id][1]
+                    elif self.backend == "jax":
+                        node.id = self._math_map[node.id][2]
+                else:
+                    raise ValueError(
+                        f"Unrecognized named variable or function {node.id}. "
+                        "This could be because the function is not supported or "
+                        "you forgot to include the variable as an argument."
+                    )
+            else:
+                raise ValueError(f"Invalid part of expression {type(node)}")
+
+        # Convert the expression back into a string.
+        self.expression = ast.unparse(tree)

tests/tdastro/math_nodes/test_basic_math_node.py

@@ -0,0 +1,138 @@
+import math
+
+import jax
+import pytest
+from tdastro.math_nodes.basic_math_node import BasicMathNode
+from tdastro.math_nodes.single_value_node import SingleVariableNode
+
+
+def test_basic_math_node():
+    """Test that we can perform computations via a BasicMathNode."""
+    node_a = SingleVariableNode("a", 10.0)
+    node_b = SingleVariableNode("b", -5.0)
+    node = BasicMathNode("a + b", a=node_a.a, b=node_b.b, node_label="test", backend="math")
+    state = node.sample_parameters()
+    assert state["test"]["function_node_result"] == 5.0
+
+    # Try with a math function.
+    node_c = SingleVariableNode("c", 1000.0)
+    node = BasicMathNode("a + b - log10(c)", a=10.0, b=5.0, c=node_c.c, node_label="test", backend="math")
+    state = node.sample_parameters()
+    assert state["test"]["function_node_result"] == 12.0
+
+    # Try with a second math function.
+    node = BasicMathNode(
+        "sqrt(a) + b - log10(c)", a=16.0, b=4.0, c=node_c.c, node_label="test", backend="math"
+    )
+    state = node.sample_parameters()
+    assert state["test"]["function_node_result"] == 5.0
+
+    # Test that we can reproduce the power function.
+    node_d = SingleVariableNode("d", 5.0)
+    node = BasicMathNode("a ** b", a=node_d.d, b=2.5, node_label="test", backend="math")
+    state = node.sample_parameters()
+    assert state["test"]["function_node_result"] == pytest.approx(math.pow(5.0, 2.5))
+
+
+def test_basic_math_node_special_cases():
+    """Test that we can handle some of the special cases for a BasicMathNode."""
+    node_a = SingleVariableNode("a", 180.0)
+    node = BasicMathNode("sin(deg2rad(x) + pi / 2.0)", x=node_a.a, node_label="test", backend="math")
+    state = node.sample_parameters()
+    assert state["test"]["function_node_result"] == pytest.approx(-1.0)
+
+
+def test_basic_math_node_fail():
+    """Test that we perform the needed checks for a math node."""
+    # Imports not allowed
+    with pytest.raises(ValueError):
+        _ = BasicMathNode("import os")
+
+    # Ifs not allowed (won't work with JAX)
+    with pytest.raises(ValueError):
+        _ = BasicMathNode("x if 1.0 else 1.0", x=2.0)
+
+    # We only allow functions on the allow list.
+    with pytest.raises(ValueError):
+        _ = BasicMathNode("fake_delete_everything_no_confirm('./')")
+    with pytest.raises(ValueError):
+        _ = BasicMathNode("median(10, 20)")
+
+    # All variables must be defined.
+    with pytest.raises(ValueError):
+        _ = BasicMathNode("x + y", x=1.0)
+
+
+def test_basic_math_node_numpy():
+    """Test that we can perform computations via a BasicMathNode."""
+    node_a = SingleVariableNode("a", 10.0)
+    node_b = SingleVariableNode("b", -5.0)
+    node = BasicMathNode("a + b", a=node_a.a, b=node_b.b, node_label="test", backend="numpy")
+    state = node.sample_parameters()
+    assert state["test"]["function_node_result"] == 5.0
+
+    # Try with a math function.
+    node_c = SingleVariableNode("c", 1000.0)
+    node = BasicMathNode("a + b - log10(c)", a=10.0, b=5.0, c=node_c.c, node_label="test", backend="numpy")
+    state = node.sample_parameters()
+    assert state["test"]["function_node_result"] == 12.0
+
+    # Try with a second math function.
+    node = BasicMathNode(
+        "sqrt(a) + b - log10(c)", a=16.0, b=4.0, c=node_c.c, node_label="test", backend="numpy"
+    )
+    state = node.sample_parameters()
+    assert state["test"]["function_node_result"] == 5.0
+
+    # Test that we can reproduce the power function.
+    node_d = SingleVariableNode("d", 5.0)
+    node = BasicMathNode("a ** b", a=node_d.d, b=2.5, node_label="test", backend="math")
+    state = node.sample_parameters()
+    assert state["test"]["function_node_result"] == pytest.approx(math.pow(5.0, 2.5))
+
+
+def test_basic_math_node_jax():
+    """Test that we can perform computations via a BasicMathNode."""
+    node_a = SingleVariableNode("a", 10.0)
+    node_b = SingleVariableNode("b", -5.0)
+    node = BasicMathNode("a + b", a=node_a.a, b=node_b.b, node_label="test", backend="jax")
+    state = node.sample_parameters()
+    assert state["test"]["function_node_result"] == 5.0
+
+    # Try with a math function.
+    node_c = SingleVariableNode("c", 1000.0)
+    node = BasicMathNode("a + b - log10(c)", a=10.0, b=5.0, c=node_c.c, node_label="test", backend="jax")
+    state = node.sample_parameters()
+    assert state["test"]["function_node_result"] == 12.0
+
+    # Try with a second math function.
+    node = BasicMathNode(
+        "sqrt(a) + b - log10(c)", a=16.0, b=4.0, c=node_c.c, node_label="test", backend="jax"
+    )
+    state = node.sample_parameters()
+    assert state["test"]["function_node_result"] == 5.0
+
+    # Test that we can reproduce the power function.
+    node_d = SingleVariableNode("d", 5.0)
+    node = BasicMathNode("a ** b", a=node_d.d, b=2.5, node_label="test", backend="math")
+    state = node.sample_parameters()
+    assert state["test"]["function_node_result"] == pytest.approx(math.pow(5.0, 2.5))
+
+
+def test_basic_math_node_autodiff_jax():
+    """Test that we can do auto-differentiation with JAX."""
+    node_a = SingleVariableNode("a", 16.0, node_label="a_node")
+    node_b = SingleVariableNode("b", 1000.0, node_label="b_node")
+
+    # Create a basic math function and create tghe pytree.
+    node = BasicMathNode(
+        "sqrt(a) + 1.0 - log10(b)", a=node_a.a, b=node_b.b, node_label="diff_test", backend="jax"
+    )
+    state = node.sample_parameters()
+    pytree = node.build_pytree(state)
+
+    gr_func = jax.value_and_grad(node.resample_and_compute)
+    values, gradients = gr_func(pytree)
+    assert values == 2.0
+    assert gradients["a_node"]["a"] > 0.0
+    assert gradients["b_node"]["b"] < 0.0