8. 配置文件详细说明
本节将对 pulsar2 build 中的 config 文件进行详细介绍.
8.1. 配置文件概述
工具链支持的全部编译参数定义请参考 《proto 配置定义》 ，基础数据结构为
BuildConfig；用户可以根据参数规格编写
prototxt / 宽松 json / yaml / toml格式的配置文件，通过命令行参数--config指向配置文件；宽松的
json格式：支持包含js-style或者python-style注释的json文件；
部分编译参数支持命令行传入，且优先级高于配置文件，通过
pulsar2 build -h查看支持的命令行编译参数，比如命令行参数--quant.calibration_method相当于配置了QuantConfig结构体的calibration_method字段.
8.2. 完整的 json 配置参考
{
// input model file path. type: string. required: true.
"input": "/path/to/lenet5.onnx",
// axmodel output directory. type: string. required: true.
"output_dir": "/path/to/output_dir",
// rename output axmodel. type: string. required: false. default: compiled.axmodel.
"output_name": "compiled.axmodel",
// temporary data output directory. type: string. required: false. default: same with ${output_dir}.
"work_dir": "",
// input model type. type: enum. required: false. default: ONNX. option: ONNX, QuantAxModel, QuantONNX.
"model_type": "ONNX",
// target hardware. type: enum. required: false. default: AX650. option: AX650, AX620E, M76H.
"target_hardware": "AX650",
// npu mode. while ${target_hardware} is AX650, npu mode can be NPU1 / NPU2 / NPU3. while ${target_hardware} is AX620E, npu mode can be NPU1 / NPU2. type: enum. required: false. default: NPU1.
"npu_mode": "NPU1",
// modify model input shape of input model, this feature will take effect before the `input_processors` configuration. format: input1:1x3x224x224;input2:1x1x112x112. type: string. required: false. default: .
"input_shapes": "input:1x1x28x28",
"onnx_opt": {
// disable onnx optimization. type: bool. required: false. default: false.
"disable_onnx_optimization": false,
// enable onnx simplify by https://github.com/daquexian/onnx-simplifier. type: bool. required: false. default: false.
"enable_onnxsim": false,
// enable model check. type: bool. required: false. default: false.
"model_check": false,
// disable transformation check. type: bool. required: false. default: false.
"disable_transformation_check": false,
// save tensors data to optimize memory footprint. type: bool. required: false. default: false.
"save_tensors_data": false
},
"quant": {
"input_configs": [
{
// input tensor name in origin model. "DEFAULT" means input config for all input tensors. type: string. required: true.
"tensor_name": "input",
// quantize calibration dataset archive file path. type: string. required: true. limitation: tar, tar.gz, zip.
"calibration_dataset": "/path/to/dataset",
// quantize calibration data format. type: enum. required: false. default: Image. option: Image, Numpy, Binary.
"calibration_format": "Image",
// quantize calibration data size is min(${calibration_size}, size of ${calibration_dataset}), "-1" means load all dataset. type: int. required: false. default: 32.
"calibration_size": 32,
// quantize mean parameter of normlization. type: float array. required: false. default: [].
"calibration_mean": [127],
// quantize std parameter of normlization. type: float array. required: false. default: [].
"calibration_std": [1]
}
],
"layer_configs": [
{
// set layer quantize precision. type: string. required: must choose between `layer_name` and `op_type`. default: .
"layer_name": "Conv_0",
// quantize data type. type: enum. required: false. default: U8. option: U8, S8, U16, S16, FP32.
"data_type": "U8",
// quantize data type for Conv. type: enum. required: false. default: U8. option: U8, S8, U16, S16, FP32.
"output_data_type": "U8",
// quantize weight type for Conv. type: enum. required: false. default: S8. option: S8, FP32.
"weight_data_type": "S8"
},
{
// set quantize precision by operator type. type: string. required: must choose between `layer_name` and `op_type`. default: .
"op_type": "MaxPool",
// quantize data type. type: enum. required: false. default: U8. option: U8, S8, U16, S16, FP32.
"data_type": "U8"
},
{
// start tensor names of subgraph quantization config. type: string array. required: false. default: [].
"start_tensor_names": ["13"],
// end tensor names of subgraph quantization config. type: string array. required: false. default: [].
"end_tensor_names": ["15"],
// quantize data type. type: enum. required: false. default: U8. option: U8, S8, U16, S16, FP32.
"data_type": "U16"
}
],
// quantize calibration method. type: enum. required: false. default: MinMax. option: MinMax, Percentile, MSE.
"calibration_method": "MinMax",
// enable quantization precision analysis. type: bool. required: false. default: false.
"precision_analysis": true,
// precision analysis method. type: enum. required: false. default: PerLayer. option: PerLayer, EndToEnd.
"precision_analysis_method": "PerLayer",
// precision analysis mode. type: enum. required: false. default: Reference. option: Reference, NPUBackend.
"precision_analysis_mode": "Reference",
// input sample data dir for precision analysis. type: string. required: false. default: .
"input_sample_dir": "",
// enable highest mix precision quantization. type: bool. required: false. default: false.
"highest_mix_precision": false,
// conv bias data type. type: enum. required: false. default: S32. option: S32, FP32.
"conv_bias_data_type": "S32",
// LayerNormalization scale data type. type: enum. required: false. default: FP32. option: FP32, S32, U32.
"ln_scale_data_type": "FP32",
// refine weight threshold, should be a legal float number, like 1e-6. -1 means disable this feature. type: float. required: false. default: 1e-6. limitation: 0 or less than 0.0001.
"refine_weight_threshold": 1e-6,
// enalbe smooth quant strategy for conv 1x1. type: bool. required: false. default: false.
"enable_smooth_quant": false,
// tranformer opt level. type: int. required: false. default: 0. limitation: 0~2.
"transformer_opt_level": 0,
// quant check level, 0: no check; 1: check node dtype. type: int. required: false. default: 0.
"check": 0,
// refine weight scale and input scale, type: bool. required: false. default: false.
"disable_auto_refine_scale": false
},
"input_processors": [
{
// input tensor name in origin model. "DEFAULT" means processor for all input tensors. type: string. required: true.
"tensor_name": "input",
// input tensor format in origin model. type: enum. required: false. default: AutoColorSpace. option: AutoColorSpace, BGR, RGB, GRAY.
"tensor_format": "AutoColorSpace",
// input tensor layout in origin model. type: enum. required: false. default: NCHW. option: NHWC, NCHW.
"tensor_layout": "NCHW",
// input format in runtime. type: enum. required: false. default: AutoColorSpace. option: AutoColorSpace, GRAY, BGR, RGB, YUYV422, UYVY422, YUV420SP, YVU420SP.
"src_format": "AutoColorSpace",
// input layout in runtime; if `src_format` is YUV/YVU, `src_layout` will be changed to NHWC. type: enum. required: false. default: NCHW. option: NHWC, NCHW.
"src_layout": "NHWC",
// input data type in runtime. type: enum. required: false. default: FP32. option: U8, S8, U16, S16, U32, S32, FP16, FP32.
"src_dtype": "U8",
// extra compiler shapes for this input. src_extra_shapes size of every input should be the same. shape at the same index of every input will be treated as a input group which can inference independently at runtime. type: list of Shape. required: false. default [].
"src_extra_shapes": [],
// color space mode. type: enum. required: false. default: NoCSC. option: NoCSC, Matrix, FullRange, LimitedRange.
"csc_mode": "NoCSC",
// color space conversion matrix, 12 elements array that represents a 3x4 matrix. type: float array. required: false. default: [].
"csc_mat": [1.1, 1.2, 1.3, 1.4, 2.1, 2.2, 2.3, 2.4, 3.1, 3.2, 3.3, 3.4],
// mean parameter of normlization in runtime. type: float array. required: false. default: same with ${quant.input_configs.calibration_mean}.
"mean": [],
// std parameter of normlization in runtime. type: float array. required: false. default: same with ${quant.input_configs.calibration_std}.
"std": []
}
],
"output_processors": [
{
// output tensor name in origin model. "DEFAULT" means processor for all output tensors. type: string. required: true.
"tensor_name": "output",
// permute the output tensor. type: int32 array. required: false. default: [].
"dst_perm": [0, 1],
// output data type. type: enum. required: false. default: FP32. option: FP32, U8.
"output_dtype": "FP32"
}
],
"const_processors": [
{
// const tensor name in origin model. type: string. required: true.
"name": "fc2.bias",
// const tensor data array. type: list of double. required: false.
"data": [0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
// const tensor data file path, support .bin / .npy / .txt. type: string. required: false.
"data_path": "replaced_data_file_path"
}
],
"quant_op_processors": [
{
// operator name in origin model. type: string. required: true.
"op_name": "MaxPool_3",
// operator attributes to be patched. type: dict. default: {}. required: true.
"attrs": {
"ceil_mode": 0
}
},
{
"op_name": "Flatten_4", // AxReshape
"attrs": {
"shape": [0, 800]
}
}
],
"compiler": {
// static batch sizes. type: int array. required: false. default: [].
"static_batch_sizes": [],
// max dynamic batch. type: int, required: false. default: 0.
"max_dynamic_batch_size": 0,
// disable ir fix, only work in multi-batch compilation. type: bool. required: false. default: false.
"disable_ir_fix": false,
// compiler check level, 0: no check; 1: simulate compile result; 2: simulate and check compile result (for debug). type: int. required: false. default: 0.
"check": 0,
// compiler debug level. type: int. required: false. default: 0.
"debug": 0,
// input sample data dir for compiler check. type: string. required: false. default: .
"input_sample_dir": ""
}
}
8.3. 量化参数说明
input_configs中的tensor_name需要根据模型的实际输入/输出节点名称进行设置。input_configs中的tensor_name可以设置为DEFAULT代表量化配置应用于全部输入。模型输入的色彩空间由预处理
input_processors配置中的tensor_format参数来表达。工具链读取量化校准集时，会根据
input_processors中的tensor_format参数自动转换校准集数据的色彩空间。layer_configs中的layer_name及op_type选项不可以同时配置。transformer_opt_level设置Transformer模型的优化选项。
8.4. 量化精度分析参数说明
精度分析计算方法，
precision_analysis_mode字段。Reference可以运行编译器支持的全部模型（支持包含 CPU 及 NPU 子图的模型），但是计算结果相比于最终上板结果会有少量误差（基本上差距在正负 1 内，且无系统性误差）。NPUBackend可以运行仅包含 NPU 子图的模型，但是计算结果与上板结果比特对齐。
精度分析方法，
precision_analysis_method字段。PerLayer意味着每一层都采用浮点模型对应的层输入，计算每一层的输出与浮点模型输出的相似度。EndToEnd代表首层采用浮点模型输入，然后进行完整模型的仿真，计算最终输出结果与浮点模型输出的相似度。
8.5. 预处理、后处理参数说明
input_processors/output_processors配置说明tensor_name需要根据模型的实际输入/输出节点名称进行设置。tensor_name可以设置为DEFAULT代表配置应用于全部输入或者输出。前缀为
tensor_的参数代表原始模型中的输入输出属性。前缀为
src_的参数代表着运行时实际的输入输出属性。工具链会根据用户的配置自动添加算子，以完成运行时输入输出与原始模型输入输出之间的转换。
例如：当
tensor_layout为NCHW，且src_layout为NHWC时，工具链会在原始模型输入之前自动添加一个perm属性为 [0, 3, 1, 2] 的Transpose算子。
色彩空间转换预处理
当
csc_mode为LimitedRange或者FullRange且src_format为YUV 色彩空间时，工具链会根据内置的模板参数，在原始的输入前添加一个色彩空间转换算子，此时csc_mat配置无效；当
csc_mode为Matrix且src_format为YUV 色彩空间时，工具链会根据用户配置的csc_mat矩阵，在原始的输入前添加一个色彩空间转换算子，以实现在运行时将输入的YUV数据转换为模型计算所需的BGR或者RGB数据；当
csc_mode为Matrix时，计算流程为，先将YUV / YVU 色彩空间输入统一转换为YUV444格式，然后再乘以csc_mat系数矩阵。当
csc_mode为Matrix时，bias(csc_mat[3] / csc_mat[7] / csc_mat[11]) 数值范围为 (-9, 8)。其余参数 (csc_mat[0-2] / csc_mat[4-6] / csc_mat[8-10]) 数值范围为 (-524289, 524288)。
归一化预处理
input_processors中的mean/std参数，默认为用户在量化配置中calibration_mean/calibration_std参数所配置的值。如果用户希望在运行时采用不同的归一化参数，那么可以显示的配置 中的
mean/std参数以覆盖默认值。
8.6. proto 配置定义
syntax = "proto3";
package common;
enum ColorSpace {
AutoColorSpace = 0;
GRAY = 1;
BGR = 2;
RGB = 3;
RGBA = 4;
YUV420SP = 6; // Semi-Planner, NV12
YVU420SP = 7; // Semi-Planner, NV21
YUYV422 = 8; // Planner, YUYV
UYVY422 = 9; // Planner, UYVY
}
enum Layout {
DefaultLayout = 0;
NHWC = 1;
NCHW = 2;
}
enum DataType {
DefaultDataType = 0;
U8 = 1;
S8 = 2;
U16 = 3;
S16 = 4;
U32 = 5;
S32 = 6;
U64 = 7;
S64 = 8;
FP16 = 9;
FP32 = 10;
}
enum NPUMode {
NPU1 = 0;
NPU2 = 1;
NPU3 = 2;
}
enum HardwareType {
AX650 = 0;
AX620E = 1;
M76H = 2;
}
syntax = "proto3";
import "path/to/common.proto";
import "google/protobuf/struct.proto";
package pulsar2.build;
enum ModelType {
ONNX = 0;
QuantAxModel = 1;
QuantONNX = 3;
}
enum QuantMethod {
MinMax = 0;
Percentile = 1;
MSE = 2;
}
enum PrecisionAnalysisMethod {
PerLayer = 0;
EndToEnd = 1;
}
enum PrecisionAnalysisMode {
Reference = 0;
NPUBackend = 1;
}
enum DataFormat {
Image = 0;
Numpy = 1;
Binary = 2;
}
enum CSCMode {
NoCSC = 0;
Matrix = 1;
FullRange = 2;
LimitedRange = 3;
}
message InputQuantConfig {
// input tensor name in origin model. "DEFAULT" means input config for all input tensors. type: string. required: true.
string tensor_name = 1;
// quantize calibration dataset archive file path. type: string. required: true. limitation: tar, tar.gz, zip.
string calibration_dataset = 2;
// quantize calibration data format. type: enum. required: false. default: Image. option: Image, Numpy, Binary.
DataFormat calibration_format = 3;
// quantize calibration data size is min(${calibration_size}, size of ${calibration_dataset}), "-1" means load all dataset. type: int. required: false. default: 32.
int32 calibration_size = 4;
// quantize mean parameter of normlization. type: float array. required: false. default: [].
repeated float calibration_mean = 5;
// quantize std parameter of normlization. type: float array. required: false. default: [].
repeated float calibration_std = 6;
}
message LayerConfig {
// set layer quantize precision. type: string. required: must choose between `layer_name` and `op_type`. default: .
string layer_name = 1;
// set quantize precision by operator type. type: string. required: must choose between `layer_name` and `op_type`. default: .
string op_type = 2;
// start tensor names of subgraph quantization config. type: string array. required: false. default: [].
repeated string start_tensor_names = 3;
// end tensor names of subgraph quantization config. type: string array. required: false. default: [].
repeated string end_tensor_names = 4;
// quantize data type. type: enum. required: false. default: U8. option: U8, S8, U16, S16, FP32.
common.DataType data_type = 5;
// quantize weight type for Conv. type: enum. required: false. default: S8. option: S8, FP32.
common.DataType weight_data_type = 6;
// quantize data type for Conv. type: enum. required: false. default: U8. option: U8, S8, U16, S16, FP32.
common.DataType output_data_type = 10;
}
message OnnxOptimizeOption {
// disable onnx optimization. type: bool. required: false. default: false.
bool disable_onnx_optimization = 1;
// enable onnx simplify by https://github.com/daquexian/onnx-simplifier. type: bool. required: false. default: false.
bool enable_onnxsim = 2;
// enable model check. type: bool. required: false. default: false.
bool model_check = 3;
// disable transformation check. type: bool. required: false. default: false.
bool disable_transformation_check = 4;
}
message QuantConfig {
repeated InputQuantConfig input_configs = 1;
repeated LayerConfig layer_configs = 2;
// quantize calibration method. type: enum. required: false. default: MinMax. option: MinMax, Percentile, MSE.
QuantMethod calibration_method = 3;
// enable quantization precision analysis. type: bool. required: false. default: false.
bool precision_analysis = 4;
// precision analysis method. type: enum. required: false. default: PerLayer. option: PerLayer, EndToEnd.
PrecisionAnalysisMethod precision_analysis_method = 5;
// precision analysis mode. type: enum. required: false. default: Reference. option: Reference, NPUBackend.
PrecisionAnalysisMode precision_analysis_mode = 6;
// enable highest mix precision quantization. type: bool. required: false. default: false.
bool highest_mix_precision = 7;
// conv bias data type. type: enum. required: false. default: S32. option: S32, FP32.
common.DataType conv_bias_data_type = 8;
// refine weight threshold, should be a legal float number, like 1e-6. -1 means disable this feature. type: float. required: false. default: 1e-6. limitation: 0 or less than 0.0001.
float refine_weight_threshold = 9;
// enalbe smooth quant strategy for conv 1x1. type: bool. required: false. default: false.
bool enable_smooth_quant = 10;
// tranformer opt level. type: int. required: false. default: 0. limitation: 0~2.
int32 transformer_opt_level = 20;
// input sample data dir for precision analysis. type: string. required: false. default: .
string input_sample_dir = 30;
// LayerNormalization scale data type. type: enum. required: false. default: FP32. option: FP32, S32, U32.
common.DataType ln_scale_data_type = 40;
// quant check level, 0: no check; 1: check node dtype. type: int. required: false. default: 0.
int32 check = 50;
// refine weight scale and input scale, type: bool. required: false. default: false.
bool disable_auto_refine_scale = 60;
}
message InputProcessor {
// input tensor name in origin model. "DEFAULT" means processor for all input tensors. type: string. required: true.
string tensor_name = 1;
// input tensor format in origin model. type: enum. required: false. default: AutoColorSpace. option: AutoColorSpace, BGR, RGB, GRAY.
common.ColorSpace tensor_format = 2;
// input tensor layout in origin model. type: enum. required: false. default: NCHW. option: NHWC, NCHW.
common.Layout tensor_layout = 3;
// input format in runtime. type: enum. required: false. default: AutoColorSpace. option: AutoColorSpace, GRAY, BGR, RGB, YUYV422, UYVY422, YUV420SP, YVU420SP.
common.ColorSpace src_format = 4;
// input layout in runtime; if `src_format` is YUV/YVU, `src_layout` will be changed to NHWC. type: enum. required: false. default: NCHW. option: NHWC, NCHW.
common.Layout src_layout = 5;
// input data type in runtime. type: enum. required: false. default: FP32. option: U8, S8, U16, S16, U32, S32, FP16, FP32.
common.DataType src_dtype = 6;
// extra compiler shapes for this input. src_extra_shapes size of every input should be the same. shape at the same index of every input will be treated as a input group which can inference independently at runtime. type: list of Shape. required: false. default [].
repeated common.Shape src_extra_shapes = 11;
// color space mode. type: enum. required: false. default: NoCSC. option: NoCSC, Matrix, FullRange, LimitedRange.
CSCMode csc_mode = 7;
// color space conversion matrix, 12 elements array that represents a 3x4 matrix. type: float array. required: false. default: [].
repeated float csc_mat = 8;
// mean parameter of normlization in runtime. type: float array. required: false. default: same with ${quant.input_configs.calibration_mean}.
repeated float mean = 9;
// std parameter of normlization in runtime. type: float array. required: false. default: same with ${quant.input_configs.calibration_std}.
repeated float std = 10;
}
message OutputProcessor {
// output tensor name in origin model. "DEFAULT" means processor for all output tensors. type: string. required: true.
string tensor_name = 1;
common.Layout tensor_layout = 2;
// permute the output tensor. type: int32 array. required: false. default: [].
repeated int32 dst_perm = 3;
// output data type. type: enum. required: false. default: FP32. option: FP32, U8.
common.DataType output_dtype = 4;
}
message OpProcessor {
// operator name in origin model. type: string. required: true.
string op_name = 1;
// operator attributes to be patched. type: dict. default: {}. required: true.
.google.protobuf.Struct attrs = 2;
}
message ConstProcessor {
// const tensor name in origin model. type: string. required: true.
string name = 1;
// const tensor data array. type: list of double. required: false.
repeated double data = 2;
// const tensor data file path, support .bin / .npy / .txt. type: string. required: false.
string data_path = 3;
}
message CompilerConfig {
// static batch sizes. type: int array. required: false. default: [].
repeated int32 static_batch_sizes = 1;
// max dynamic batch. type: int, required: false. default: 0.
int32 max_dynamic_batch_size = 2;
// disable ir fix, only work in multi-batch compilation. type: bool. required: false. default: false.
bool disable_ir_fix = 3;
// compiler check level, 0: no check; 1: simulate compile result; 2: simulate and check compile result (for debug). type: int. required: false. default: 0.
int32 check = 5;
// compiler debug level. type: int. required: false. default: 0.
int32 debug = 6;
// input sample data dir for compiler check. type: string. required: false. default: .
string input_sample_dir = 30;
}
message BuildConfig {
// input model file path. type: string. required: true.
string input = 1;
// axmodel output directory. type: string. required: true.
string output_dir = 2;
// rename output axmodel. type: string. required: false. default: compiled.axmodel.
string output_name = 3;
// temporary data output directory. type: string. required: false. default: same with ${output_dir}.
string work_dir = 4;
// input model type. type: enum. required: false. default: ONNX. option: ONNX, QuantAxModel, QuantONNX.
ModelType model_type = 5;
// target hardware. type: enum. required: false. default: AX650. option: AX650, AX620E, M76H.
common.HardwareType target_hardware = 6;
// npu mode. while ${target_hardware} is AX650, npu mode can be NPU1 / NPU2 / NPU3. while ${target_hardware} is AX620E, npu mode can be NPU1 / NPU2. type: enum. required: false. default: NPU1.
common.NPUMode npu_mode = 7;
// modify model input shape of input model, this feature will take effect before the `input_processors` configuration. format: input1:1x3x224x224;input2:1x1x112x112. type: string. required: false. default: .
string input_shapes = 8;
OnnxOptimizeOption onnx_opt = 10;
QuantConfig quant = 20;
repeated InputProcessor input_processors = 31;
repeated OutputProcessor output_processors = 32;
repeated ConstProcessor const_processors = 33;
repeated OpProcessor op_processors = 34;
repeated OpProcessor quant_op_processors = 35;
CompilerConfig compiler = 40;
}