lite-converter

name	lite-converter
description	Model conversion pipeline, parser development, optimization passes and quantization. Use when converting models to .ms, writing parser code, implementing optimizer passes, or configuring quantization.
paths	["mindspore-lite/tools/converter/**","mindspore-lite/tools/optimizer/**","mindspore-lite/schema/**","mindspore-lite/tools/schema_gen/**"]

MindSpore Lite Model Conversion and Optimization

Conversion Pipeline

Input Model (MindIR/TF/Caffe/ONNX/TFLite/PyTorch)
  -> Parse (framework-specific Parser) -> Unified MindIR (ANF Graph)
  -> Import -> Internal graph representation
  -> Optimize (Constant Folding, Op Fusion, Format Transform, Parallel Split, Redundant elimination)
  -> Quantize (optional: Weight / Full / Mixed Precision)
  -> Export (.ms for LiteRT or .mindir for ExtendRT)

Conversion Tool Usage

# MindIR -> .ms (device-side)
./converter_lite --fmk=MINDIR --modelFile=model.mindir --outputFile=model

# TensorFlow / TFLite / Caffe / ONNX / PyTorch -> .ms
./converter_lite --fmk=TF --modelFile=model.pb --outputFile=model
./converter_lite --fmk=TFLITE --modelFile=model.tflite --outputFile=model
./converter_lite --fmk=CAFFE --modelFile=model.prototxt --weightFile=model.caffemodel --outputFile=model
./converter_lite --fmk=ONNX --modelFile=model.onnx --outputFile=model
./converter_lite --fmk=PYTORCH --modelFile=model.pt --outputFile=model

# With quantization
./converter_lite --fmk=MINDIR --modelFile=model.mindir --outputFile=model \
  --quantType=WeightQuant --bitNum=8

# Cloud-side optimization

# Ascend-specific
./converter_lite --fmk=ONNX --modelFile=model.ONNX --outputFile=model --optimize=ascend_oriented

# CPU
./converter_lite --fmk=ONNX --modelFile=model.ONNX --outputFile=model

Key Parameters

Parameter	Description
--fmk	Input framework: MINDIR/TF/TFLITE/CAFFE/ONNX/PYTORCH
--modelFile	Input model file path
--weightFile	Caffe weight file (Caffe only)
--outputFile	Output file path (without extension)
--quantType	WeightQuant / FullQuant / NoQuant
--bitNum	Quantization bits: 1-8 (default 8)
--optimize	ascend_oriented / general / none
--configFile	Quantization or runtime config file
--inputShape	Dynamic shape input (e.g., input1:1,3,224,224;input2:1,3,256,256)

Parser Development

mindspore-lite/tools/converter/parser/
  onnx/             # onnx_model_parser.cc + per-operator parsers
  tf/               # tf_model_parser.cc + per-operator parsers
  tflite/           # tflite_model_parser.cc + per-operator parsers
  caffe/            # caffe_model_parser.cc + per-operator parsers

Adding a New Framework Parser

1. Create directory under tools/converter/parser/
2. Implement XxxModelParser with Parse() method (original model -> ANF Graph)
3. Operator mapping: original operator -> MindSpore Primitive
4. Register FMK type in tools/converter/converter_flags.cc
5. Add unit tests

Operator Mapping

Each Parser maps original operators to MindSpore internals:

• Attribute conversion: e.g., TF padding format to MindSpore format
• Data layout: NHWC <-> NCHW conversion
• Missing operators: simulate with composites (e.g., ReduceMean for GlobalAvgPool)

Converter Directory Structure

mindspore-lite/tools/converter/
  adapter/          # Format adapters
  config_parser/    # Config file parsing
  converter_lite/   # CLI tool entry point
  cxx_api/          # Converter C++ API
  decomposer/       # Operator decomposition
  import/           # Model import (mindir_importer, primitive_adjust)
  legacy_optimizer/ # Legacy optimization passes
  micro/            # Micro code generation
  ops/              # Operator utilities
  parser/           # Framework-specific parsers
  preprocess/       # Data preprocessing pipeline
  quantizer/        # Quantization implementations
    calibrator.cc        # Calibration data processing
    full_quant_quantizer/  # Full quantization
    weight_quantizer/      # Weight-only quantization
    gptq_quantizer/        # GPTQ quantization
    fse_encoder/           # FSE encoding
    huffman_encode/        # Huffman encoding
  registry/         # Extension registration
  session/          # Conversion session

Optimizer Pass Development

mindspore-lite/tools/optimizer/
  common/           # Common optimization utilities
  const_fold/       # Constant folding passes
  fiss on/          # Operator fission passes (note: directory is "fisson")
  format/           # Format transform passes
  fusion/           # Operator fusion passes (conv_bn, conv_activation, matmul_add, etc.)
  graph/            # Graph-level optimization
  parallel/         # Parallel split passes

Writing an Optimization Pass

class MyFusionPass : public Pass {
 public:
  bool Run(const FuncGraphPtr &graph) override {
    auto node_list = TopoSort(graph->get_return());
    for (auto &node : node_list) {
      if (!CheckPattern(node)) continue;
      DoFusion(node);
      changed = true;
    }
    return changed;
  }
};

Passes execute during converter phase in a fixed order chain.

Quantization

Post-training Quantization

1. Prepare calibration dataset (100-500 representative samples)
2. Write config file:

[common_quant_param]
quant_type=WEIGHT_QUANT
bit_num=8

[data_preprocess]
calibrate_path=/path/to/calibration/images/
calibrate_size=100

[input_format]
input_type=IMAGE
resize_height=224
resize_width=224

3. Run: ./converter_lite --fmk=MINDIR --modelFile=model.mindir --outputFile=model_quant --configFile=quant_config.cfg

Quantization Types

Type	Description	Use Case
WeightQuant	Weight-only	Reduce model size, minimal accuracy loss
FullQuant	Weight + activation	Maximum compression, needs calibration data
Mixed Precision	Partial INT8 + FP32	Balance accuracy and performance

Export Formats

Format	Target Runtime	Serialization
.ms	LiteRT (device-side)	FlatBuffers, zero-copy deserialization
.mindir	ExtendRT (cloud-side)	Protobuf, supports large models

Schema files: mindspore-lite/schema/ops.fbs (~1.3K lines), model.fbs, ops_types.fbs

Common Conversion Issues

1. Unsupported operator: Check schema/ops.fbs operator list
2. Shape inference failure: Use --inputShape to specify input shapes
3. Quantization accuracy drop: Try mixed precision or more calibration data
4. Caffe missing BatchNorm params: Ensure .caffemodel weight file path is correct