Evaluation & Derivatives

Evaluation

Given an expression tree specified with a Node type, you may evaluate the expression over an array of data with the following command:

eval_tree_array(tree::AbstractExpressionNode, cX::AbstractMatrix{T}, operators::OperatorEnum; turbo::Union{Bool,Val}=Val(false), bumper::Union{Bool,Val}=Val(false))

def eval(current_node)
    if current_node is leaf
        return current_node.value
    elif current_node is degree 1
        return current_node.operator(eval(current_node.left_child))
    else
        return current_node.operator(eval(current_node.left_child), eval(current_node.right_child))

Assuming you are only using a single OperatorEnum, you can also use the following shorthand by using the expression as a function:

    (tree::AbstractExpressionNode)(X::AbstractMatrix{T}, operators::OperatorEnum; turbo::Union{Bool,Val}=false, bumper::Union{Bool,Val}=Val(false))

Evaluate a binary tree (equation) over a given input data matrix. The
operators contain all of the operators used. This function fuses doublets
and triplets of operations for lower memory usage.

# Arguments
- `tree::AbstractExpressionNode`: The root node of the tree to evaluate.
- `cX::AbstractMatrix{T}`: The input data to evaluate the tree on.
- `operators::OperatorEnum`: The operators used in the tree.
- `turbo::Union{Bool,Val}`: Use LoopVectorization.jl for faster evaluation.
- `bumper::Union{Bool,Val}`: Use Bumper.jl for faster evaluation.

# Returns
- `output::AbstractVector{T}`: the result, which is a 1D array.
    Any NaN, Inf, or other failure during the evaluation will result in the entire
    output array being set to NaN.

For example,

using DynamicExpressions

operators = OperatorEnum(; binary_operators=[+, -, *], unary_operators=[cos])
tree = Node(; feature=1) * cos(Node(; feature=2) - 3.2)

tree([1 2 3; 4 5 6.], operators)

3-element Vector{Float64}:
  0.6967067435533686
 -0.4544040965128262
 -2.8266669740855317

This is possible because when you call OperatorEnum, it automatically re-defines (::Node)(X) to call the evaluation operation with the given operators loaded. It also re-definesprint,show, and the various operators, to work with theNode` type.

You can also work with arbitrary types, by defining a GenericOperatorEnum instead. The notation is the same for eval_tree_array, though it will return nothing when it can't find a method, and not do any NaN checks:

eval_tree_array(tree::AbstractExpressionNode, cX::AbstractMatrix, operators::GenericOperatorEnum; throw_errors::Bool=true)

function eval(current_node)
    if current_node is leaf
        return current_node.value
    elif current_node is degree 1
        return current_node.operator(eval(current_node.left_child))
    else
        return current_node.operator(eval(current_node.left_child), eval(current_node.right_child))

Likewise for the shorthand notation:

    (tree::Node)(X::AbstractMatrix, operators::GenericOperatorEnum; throw_errors::Bool=true)

# Arguments
- `X::AbstractArray`: The input data to evaluate the tree on.
- `operators::GenericOperatorEnum`: The operators used in the tree.
- `throw_errors::Bool=true`: Whether to throw errors
    if they occur during evaluation. Otherwise,
    MethodErrors will be caught before they happen and
    evaluation will return `nothing`,
    rather than throwing an error. This is useful in cases
    where you are unsure if a particular tree is valid or not,
    and would prefer to work with `nothing` as an output.

# Returns
- `output`: the result of the evaluation.
    If evaluation failed, `nothing` will be returned for the first argument.
    A `false` complete means an operator was called on input types
    that it was not defined for. You can change this behavior by
    setting `throw_errors=false`.

Derivatives

DynamicExpressions.jl can efficiently compute first-order derivatives of expressions with respect to variables or constants. This is done using either eval_diff_tree_array, to compute derivative with respect to a single variable, or with eval_grad_tree_array, to compute the gradient with respect all variables (or, all constants). Both use forward-mode automatic, but use Zygote.jl to compute derivatives of each operator, so this is very efficient.

eval_diff_tree_array(tree::AbstractExpressionNode{T}, cX::AbstractMatrix{T}, operators::OperatorEnum, direction::Integer; turbo::Union{Bool,Val}=Val(false))

eval_grad_tree_array(tree::AbstractExpressionNode{T}, cX::AbstractMatrix{T}, operators::OperatorEnum; variable::Union{Bool,Val}=Val(false), turbo::Union{Bool,Val}=Val(false))

You can compute gradients this with shorthand notation as well (which by default computes gradients with respect to input matrix, rather than constants).

    (tree::Node{T})'(X::AbstractMatrix{T}, operators::OperatorEnum; turbo::Bool=false, variable::Bool=true)

Compute the forward-mode derivative of an expression, using a similar
structure and optimization to eval_tree_array. `variable` specifies whether
we should take derivatives with respect to features (i.e., X), or with respect
to every constant in the expression.

# Arguments
- `X::AbstractMatrix{T}`: The data matrix, with each column being a data point.
- `operators::OperatorEnum`: The operators used to create the `tree`.
- `variable::Bool`: Whether to take derivatives with respect to features (i.e., `X` - with `variable=true`),
    or with respect to every constant in the expression (`variable=false`).
- `turbo::Bool`: Use `LoopVectorization.@turbo` for faster evaluation.

# Returns

- `(evaluation, gradient, complete)::Tuple{AbstractVector{T}, AbstractMatrix{T}, Bool}`: the normal evaluation,
    the gradient, and whether the evaluation completed as normal (or encountered a nan or inf).

Alternatively, you can compute higher-order derivatives by using ForwardDiff on the function differentiable_eval_tree_array, although this will be slower.

differentiable_eval_tree_array(tree::AbstractExpressionNode, cX::AbstractMatrix, operators::OperatorEnum)

Enzyme

DynamicExpressions.jl also supports automatic differentiation with Enzyme.jl. Note that this is extremely experimental. You should expect to see occasional incorrect gradients. Be sure to explicitly verify gradients are correct for a particular space of operators (e.g., with finite differences).

Let's look at an example. First, let's create a tree:

using DynamicExpressions

operators = OperatorEnum(binary_operators=(+, -, *, /), unary_operators=(cos, sin))

x1 = Node{Float64}(feature=1)
x2 = Node{Float64}(feature=2)

tree = 0.5 * x1 + cos(x2 - 0.2)

Now, say we want to take the derivative of this expression with respect to x1 and x2. First, let's evaluate it normally:

X = [1.0 2.0 3.0; 4.0 5.0 6.0]  # 2x3 matrix (2 features, 3 rows)

tree(X, operators)

Now, let's use Enzyme.jl to compute the derivative of the outputs with respect to x1 and x2, using reverse-mode autodiff:

using Enzyme

function my_loss_function(tree, X, operators)
    # Get the outputs
    y = tree(X, operators)
    # Sum them (so we can take a gradient, rather than a jacobian)
    return sum(y)
end


dX = begin
    storage=zero(X)
    autodiff(
        Reverse,
        my_loss_function,
        Active,
        ## Actual arguments to function:
        Const(tree),
        Duplicated(X, storage),
        Const(operators),
    )
    storage
end

This will get returned as

 2×3 Matrix{Float64}:
  0.5       0.5       0.5
  0.611858  0.996165  0.464602

which one can confirm is the correct gradient!

This will take a while the first time you run it, as Enzyme needs to take the gradients of the actual LLVM IR code. Subsequent runs won't spend any time compiling and be much faster.

Some general notes about this:

1. We want to take a reverse-mode gradient, so we pass Reverse to autodiff.
2. Since we want to take the gradient of the output of my_loss_function, we declare Active as the third argument.
3. Following this, we pass our actual arguments to the function.
   • Objects which we don't want to take gradients with respect to, and also don't temporarily store any data during the computation (such as tree and operators here) should be wrapped with Const.
   • Objects which we wish to take derivatives with respect to, we need to use Duplicated, and explicitly create a copy of it, with all numerical values set to zero. Enzyme will then store the derivatives in this object.

Note that you should never use anything other than turbo=Val(false) with Enzyme, as Enzyme and LoopVectorization are not compatible, and will cause a segfault. Even using turbo=false will not work, because it would cause Enzyme to trace the (unused) LoopVectorization code!