DelegateState

Helpers & Utilities · AiDotNet.Serialization

Saves and restores a delegate a layer was constructed with.

How It Works

A delegate is the one kind of construction state that cannot simply be written down, and the established answers are both unsatisfying. Python’s pickle refuses a lambda outright and stores a module-level function as a name reference. Keras goes further and marshals the Lambda layer’s Python bytecode into the model file, which is why loading one is arbitrary code execution and why safe_mode=True blocks it by default – it was still bypassable (CVE-2025-9906). PyTorch avoids the question: copy.deepcopy treats a function as atomic and returns the same object, so a clone aliases the delegate and a save never round-trips it.

.NET makes better available, because a delegate is not opaque here: it is a MethodInfo plus a target. So this tries progressively weaker descriptions and keeps the first that fits, rather than marshalling code:

When none fits, the delegate is reported as unsaveable rather than written as something that will not come back. Cloning a layer in memory does not go through here at all: it hands the live delegate over, which is what PyTorch’s deepcopy does and is always correct in-process.

Methods

Method Summary
TDelegate Load(string saved, string layerName, string key) Rebuilds a delegate from its saved form.
string Save(Delegate value) Describes value well enough to rebuild it, or returns empty when no description fits.
string Save(Delegate value, int[] inputShape) Describes value well enough to rebuild it, or returns empty when no description fits.

Fields

Field Summary
string ExpressionScheme Marks a saved value as a serialized expression tree.
string GraphScheme Marks a saved value as a legacy traced computation graph (written by AiDotNet versions before the legacy autodiff was removed).
string MethodScheme Marks a saved value as a reference to a named method.
string ProgramScheme Marks a saved value as a portable tensor program: what a Func<Tensor<T>, Tensor<T>> computed, recorded as AiDotNet.Tensors engine calls (one program per input rank it was traced at).