Mila::Dnn::NetworkFactory Class Reference

Factory registry for Network deserialization. More...

Public Types
template<DeviceType TDeviceType, TensorDataType TPrecision>
using	NetworkFactoryFunc

Static Public Member Functions
template<Compute::DeviceType TDeviceType, TensorDataType TPrecision>
static std::unique_ptr< Network< TDeviceType, TPrecision > >	create (ModelArchive &archive, std::shared_ptr< ExecutionContext< TDeviceType > > exec_context)
	Create a network instance from archive using registered factories.
template<DeviceType TDeviceType, TensorDataType TPrecision>
static void	registerNetwork (const std::string &network_type, NetworkFactoryFunc< TDeviceType, TPrecision > factory)

Detailed Description

Factory registry for Network deserialization.

Provides type-safe network reconstruction from archives using registered factory functions. Each concrete network type registers its own Load() method to enable polymorphic deserialization.

Design Pattern:

• Registration: Concrete networks register factory functions at startup
• Dispatch: Factory reads metadata to determine network type and precision
• Construction: Invokes appropriate registered factory function

Usage:

// Registration (typically in network implementation file)
NetworkFactory::registerNetwork<DeviceType::Cpu, TensorDataType::FP32>(
    "MnistClassifier",
    [](ModelArchive& archive, auto exec_ctx) {
        return MnistClassifier::Load(archive, exec_ctx->getDeviceId());
    });
 
// Deserialization
auto network = NetworkFactory::create<DeviceType::Cpu, TensorDataType::FP32>(
    archive, exec_context);

Examples

/home/runner/work/Mila/Mila/Mila/Src/Dnn/Core/NetworkFactory.ixx.

Member Typedef Documentation

NetworkFactoryFunc

template<DeviceType TDeviceType, TensorDataType TPrecision>

using Mila::Dnn::NetworkFactory::NetworkFactoryFunc

Initial value:

std::function<std::unique_ptr<Network<TDeviceType, TPrecision>>(
           ModelArchive&,
           std::shared_ptr<ExecutionContext<TDeviceType>>
       )>

Member Function Documentation

create()

template<Compute::DeviceType TDeviceType, TensorDataType TPrecision>

std::unique_ptr< Network< TDeviceType, TPrecision > > Mila::Dnn::NetworkFactory::create ( ModelArchive & archive, std::shared_ptr< ExecutionContext< TDeviceType > > exec_context )

inlinestatic

Create a network instance from archive using registered factories.

Reads network metadata from the archive to determine the network type and dispatches to the appropriate registered factory function for reconstruction.

Metadata Requirements:

• network/meta.json must contain:
  • "network_type": String identifying the concrete network class
  • "template_precision": String identifying the precision type

Template Parameters

TDeviceType	Device type for the network
TPrecision	Precision type for the network

Parameters

archive	Archive containing serialized network
exec_context	Execution context for the reconstructed network

Returns

Unique pointer to reconstructed network

Exceptions

std::runtime_error	if network metadata is missing or malformed
std::runtime_error	if no factory is registered for the network type
std::runtime_error	if network reconstruction fails

auto exec_ctx = std::make_shared<ExecutionContext<DeviceType::Cpu>>(...);
auto network = NetworkFactory::create<DeviceType::Cpu, TensorDataType::FP32>(
    archive, exec_ctx);

Examples

/home/runner/work/Mila/Mila/Mila/Src/Dnn/Core/NetworkFactory.ixx.

---

The documentation for this class was generated from the following file:

• Mila/Src/Dnn/Core/NetworkFactory.ixx