Implementation:Alibaba MNN MNNConvert Info

Field	Value
Implementation Name	MNNConvert_Info
Type	API Doc
Category	Model_Conversion_Pipeline
Source	tools/converter/source/common/cli.cpp:L99-140 (dumpModelInfo), tools/converter/source/common/cli.cpp:L258-290 (--info/--JsonFile parsing), tools/converter/source/common/cli.cpp:L1087-1214 (mnn2json/json2mnn)
External Dependencies	MNN core library (for Module::load), flatbuffers

Summary

MNNConvert provides two model inspection modes: a quick info dump that prints model metadata to stdout, and a full JSON export that serializes the entire model structure (operators, tensors, sub-graphs) to a JSON file. Both modes operate on existing .mnn model files and are accessed through the same MNNConvert CLI.

API

Model Info Dump

MNNConvert --info --modelFile model.mnn -f MNN

JSON Export

MNNConvert --JsonFile output.json --modelFile model.mnn -f MNN

JSON to MNN (Reverse Conversion)

MNNConvert -f JSON --modelFile model.json --MNNModel output.mnn

Key Parameters

Parameter	Type	Description
--info	flag	Dump model metadata (inputs, outputs, version, format) to stdout. Only valid with -f MNN.
--JsonFile	string	Export MNN model to the specified JSON file path. Only valid with -f MNN.
--modelFile	string	Path to the MNN model file (.mnn) to inspect.
-f	string	Must be MNN for info dump or JSON export; JSON for JSON-to-MNN conversion.

Inputs

• MNN model file (.mnn) -- The compiled MNN model to inspect
• JSON model file (.json) -- For reverse conversion from JSON back to MNN binary

Outputs

--info Mode Output

Prints to stdout:

Model default dimensionFormat is NCHW
Model Inputs:
[ input ]: dimensionFormat: NCHW, size: [ 1,3,224,224 ], type is float
Model Outputs:
[ output ]
Model Version: 2.8.0

Information reported:

• Default dimension format -- NCHW, NHWC, or NC4HW4
• Input details -- For each input: name, dimension format, shape, data type
• Output names -- Names of all model outputs
• Model version -- MNN version string (or "< 2.0.0" for legacy models)
• Metadata -- Any custom key-value metadata pairs embedded in the model

--JsonFile Mode Output

Produces a JSON file containing the complete FlatBuffers-serialized model structure, including:

• All operators (oplists) with their types, names, parameters, input/output indices
• Tensor names (tensorName)
• Extra tensor descriptors (quantization info, formats)
• Sub-graphs for control flow
• Weight data (unless flag-based truncation is active)

Internal Implementation

dumpModelInfo (cli.cpp:L99-140)

// tools/converter/source/common/cli.cpp:L99-140
static int dumpModelInfo(const char* modelName) {
    std::vector<std::string> empty;
    std::shared_ptr<MNN::Express::Module> module(
        MNN::Express::Module::load(empty, empty, modelName));
    if (nullptr == module.get()) {
        MNN_ERROR("Load MNN from %s Failed\n", modelName);
        return 1;
    }
    auto info = module->getInfo();
    // Print default dimensionFormat (NCHW/NHWC/NC4HW4)
    // Print each input: name, dimensionFormat, shape, type
    // Print each output name
    // Print model version
    // Print metadata key-value pairs if present
    return 0;
}

The info dump uses MNN's Express API to load the model as a Module, then calls getInfo() to retrieve:

• info->defaultFormat -- The model's default dimension format
• info->inputNames / info->inputs -- Input names and variable info (dimensions, order, type)
• info->outputNames -- Output tensor names
• info->version -- Model version string
• info->metaData -- Custom metadata map

mnn2json (cli.cpp:L1087-1190)

// tools/converter/source/common/cli.cpp:L1087-1190
bool Cli::mnn2json(const char* modelFile, const char* jsonFile, int flag) {
    // Read binary model file into buffer
    // If flag > 3: strip large data (conv weights, biases, blob data, etc.)
    // If flag > 4: dump each sub-graph to a separate file
    // Convert FlatBuffers binary to string using FlatBufferToString()
    // Write to output JSON file
}

The JSON export reads the raw MNN binary and uses FlatBuffers' FlatBufferToString() to produce a human-readable representation. When flag > 3, large tensors are cleared before export:

• Convolution weights and biases
• Blob data (for tensors with > 20 elements)
• MatMul weights and biases
• PReLU slopes
• Extra op info buffers
• LSTM weight matrices

json2mnn (cli.cpp:L1192-1214)

// tools/converter/source/common/cli.cpp:L1192-1214
bool Cli::json2mnn(const char* jsonFile, const char* modelFile) {
    // Parse JSON using rapidjson
    // Convert JSON to FlatBuffers using Json2Flatbuffer::writeJsonToFlatbuffer()
    // Write binary output to model file
}

CLI Argument Routing (cli.cpp:L258-290)

The --info and --JsonFile arguments are processed in initializeMNNConvertArgs():

// tools/converter/source/common/cli.cpp:L439-452
if (result.count("MNNModel")) {
    // Standard conversion output
} else if (result.count("JsonFile")) {
    modelPath.mnn2json = true;
    modelPath.MNNModel = JsonFilePath;
} else if (result.count("info") && modelPath.model == modelConfig::MNN) {
    modelPath.dumpInfo = true;
    return true;
}

In convertModel(), the dumpInfo flag triggers the info dump path:

// tools/converter/source/common/cli.cpp:L673-677
bool Cli::convertModel(modelConfig& modelPath) {
    if (modelPath.dumpInfo) {
        dumpModelInfo(modelPath.modelFile.c_str());
        return true;
    }
    // ... normal conversion flow
}

Usage Examples

Quick Model Inspection

./MNNConvert -f MNN --modelFile resnet18.mnn --info

Example output:

Model default dimensionFormat is NCHW
Model Inputs:
[ input.1 ]: dimensionFormat: NCHW, size: [ 1,3,224,224 ], type is float
Model Outputs:
[ 497 ]
Model Version: 2.8.0

Full JSON Export

./MNNConvert -f MNN --modelFile resnet18.mnn --JsonFile resnet18.json

JSON Export Without Weights (Structure Only)

The mnn2json function's flag parameter controls weight inclusion. This is accessible programmatically via the C++ API:

// flag > 3 strips weights for readability
MNN::Cli::mnn2json("model.mnn", "model_structure.json", 4);

// flag > 4 additionally separates sub-graphs into individual files
MNN::Cli::mnn2json("model.mnn", "model_structure.json", 5);

Convert JSON Back to MNN

./MNNConvert -f JSON --modelFile model.json --MNNModel model_rebuilt.mnn

This enables a workflow where models can be exported to JSON, manually edited (e.g., adjusting parameters or removing ops), and then re-serialized to MNN binary format.

Practical Validation Workflow

A recommended post-conversion validation sequence:

# 1. Convert the model
./MNNConvert -f ONNX --modelFile model.onnx --MNNModel model.mnn --bizCode MNN

# 2. Inspect model metadata
./MNNConvert -f MNN --modelFile model.mnn --info

# 3. Export full structure for review (optional)
./MNNConvert -f MNN --modelFile model.mnn --JsonFile model_debug.json

# 4. Run numerical verification
python tools/script/testMNNFromOnnx.py model.onnx

Related Pages

• Principle:Alibaba_MNN_Model_Inspection
• Environment:Alibaba_MNN_CPU_Build_Environment
• Implementation:Alibaba_MNN_MNNConvert_CLI