Implementation:Alibaba MNN Express Optimizer

Metadata

Source Repository	https://github.com/alibaba/MNN
Source File	express/Optimizer.cpp (29 lines)
Language	C++
Namespace	MNN::Express
Domains	Training, Optimization
Last Updated	2026-02-10

Summary

Optimizer provides the base implementation of the optimizer class for the MNN Express framework. This file defines the inner Parameters class for managing floating-point parameter arrays and a factory method Optimizer::create() for constructing optimizer instances. In its current form, the factory method is a stub/placeholder that returns nullptr, serving as the foundation for concrete optimizer implementations (e.g., SGD, Adam) to be built upon.

Note: This is currently a minimal scaffolding implementation. The create() factory method does not yet instantiate any concrete optimizer. Concrete optimizer subclasses are expected to override or extend this base.

Import

#include <MNN/expr/Optimizer.hpp>

I/O Contract

Input	Output
int n (parameter count)	Parameters object owning a float[n] array
Optimizer::Config config	std::shared_ptr<Optimizer> (currently nullptr)

Key Class: Optimizer

Optimizer::Parameters::Parameters(int n) (L13-17)

Optimizer::Parameters::Parameters(int n) {
    MNN_ASSERT(n > 0);
    mValue = new float[n];
    mSize  = n;
}

Allocates a raw floating-point array of size n to store optimizer parameters (e.g., learning rates, momentum values, weight arrays). The MNN_ASSERT guard ensures the allocation size is always positive.

Optimizer::Parameters::~Parameters() (L18-22)

Optimizer::Parameters::~Parameters() {
    if (nullptr != mValue) {
        delete[] mValue;
    }
}

Deallocates the parameter array. The null check prevents double-free in cases where the array was moved or never allocated.

Optimizer::create(Config config) (L23-26)

std::shared_ptr<Optimizer> Optimizer::create(Config config) {
    // Do nothing
    return nullptr;
}

Factory method intended to construct optimizer instances based on the provided configuration. Currently returns nullptr, indicating this is a placeholder for future implementation. Concrete optimizer types (SGD, Adam, etc.) are expected to be instantiated through this factory once their implementations are available.

Internal Dependencies

#include <MNN/expr/Optimizer.hpp>
#include "core/Backend.hpp"

The inclusion of core/Backend.hpp suggests the optimizer is designed to be backend-aware, potentially allowing optimization computations to be dispatched to different hardware backends (CPU, GPU, etc.).

Design Notes

• The Parameters inner class uses raw new[]/delete[] for memory management rather than smart pointers or standard containers, likely for performance reasons and direct memory control in training workloads.
• The factory pattern via Optimizer::create() allows runtime selection of optimizer type based on configuration, following the same pattern used throughout MNN for backend and session creation.
• The stub nature of this implementation means it serves primarily as an API contract definition, establishing the interface that concrete optimizers must fulfill.

Related Pages

• Alibaba_MNN_Express_Gradient_Optimizer_Interface -- The principle of gradient-based optimization in the MNN Express framework