当前位置： 首页 > news >正文

【Paddle2ONNX】为Paddle2ONNX升级Opset版本到18

news 来源：原创 2024/9/28 23:13:13

1 简介

在深度学习模型的部署过程中，模型格式的转换是一个至关重要的环节。Paddle2ONNX是一个开源工具，用于将飞桨（PaddlePaddle）模型转换为ONNX格式，以便在各种不同的深度学习框架中进行推理和部署。随着ONNX标准的不断更新，保持Paddle2ONNX与最新版本的兼容性显得尤为重要。本篇文章《【Paddle2ONNX】为Paddle2ONNX升级ONNX版本》将详细介绍如何为Paddle2ONNX项目升级其依赖的ONNX版本。

2 升级ONNX依赖实现过程

2.1 修改 .gitmodules

修改onnx的branch参数到最新的commit id

[submodule "third_party/onnx"]path = third_party/onnxurl = https://github.com/onnx/onnx.gitbranch = ad834eb73ee0cd9b6fa9ea892caeed5fa17d7dc0

2.2 修改 CMakeLists.txt

Paddle2ONNX在CMakeLists.txt中规定了最大opset的数值，我们需要修改为18，关于opset与onnx版本的对应关系，可以查看 onnx版本与opset版本对应表

# if you build from other version of onnx this should be modified
# refer to https://github.com/onnx/onnx/blob/main/docs/Versioning.md#released-versions
add_definitions(-DMAX_ONNX_OPSET_VERSION=18)
add_definitions(-DPADDLE2ONNX_LIB)

2.3 添加对opset18转换的支持

在 paddle2onnx/mapper/mapper.h 中添加对opset18的支持

class Mapper {public:void Run() {int32_t opset_version = helper_->GetOpsetVersion();Assert(opset_version >= 7 && opset_version <= MAX_ONNX_OPSET_VERSION,"[Paddle2ONNX] Only support opset_version in range of [7, " +std::to_string(MAX_ONNX_OPSET_VERSION) + "].");if (IsExportAsCustomOp()) {return ExportAsCustomOp();}if (opset_version == 18) {Opset18();} else if (opset_version == 17) {Opset17();} else if (opset_version == 16) {Opset16();} else if (opset_version == 15) {Opset15();} else if (opset_version == 14) {Opset14();} else if (opset_version == 13) {Opset13();} else if (opset_version == 12) {Opset12();} else if (opset_version == 11) {Opset11();} else if (opset_version == 10) {Opset10();} else if (opset_version == 9) {Opset9();} else if (opset_version == 8) {Opset8();} else {Opset7();}}virtual void Opset18() { Opset17(); }
};

3 升级算子对Opset18的支持

3.1 升级 LayerNormalize 算子

当opset升级到17后，LayerNormalize算子会出现问题，这里做的针对性的修复，主要是Scale和Bias两个参数的Shape需要与 input_shape[norm_axis:] 一致

void LayerNormMapper::Opset17() {auto input_info = GetInput("X");auto output_info = GetOutput("Y");constexpr std::array<P2ODataType, 3> T = {P2ODataType::FP16, P2ODataType::FP32, P2ODataType::FP64};auto input_name = input_info[0].name;auto input_type = input_info[0].dtype;auto input_shape = input_info[0].shape;if (std::find(T.begin(), T.end(), input_type) == T.end()) {input_name = helper_->AutoCast(input_name, input_info[0].dtype, P2ODataType::FP32);input_type = P2ODataType::FP32;}bool has_input_Bias = HasInput("Bias");bool has_input_Scale = HasInput("Scale");if (has_input_Bias && has_input_Scale) {auto scale_info = GetInput("Scale");auto scale_name = scale_info[0].name;auto scale_type = scale_info[0].dtype;if (std::find(T.begin(), T.end(), scale_type) == T.end()) {scale_name = helper_->AutoCast(scale_name, scale_type, P2ODataType::FP32);scale_type = P2ODataType::FP32;}auto bias_info = GetInput("Bias");auto bias_name = bias_info[0].name;auto bias_type = bias_info[0].dtype;if (std::find(T.begin(), T.end(), bias_type) == T.end()) {bias_name = helper_->AutoCast(bias_name, bias_type, P2ODataType::FP32);bias_type = P2ODataType::FP32;}auto layer_norm_node = helper_->MakeNode("LayerNormalization",{input_name, scale_name, bias_name},{output_info[0].name});AddAttribute(layer_norm_node, "axis", begin_norm_axis_);AddAttribute(layer_norm_node, "epsilon", epsilon_);return;}if (has_input_Scale) {auto scale_info = GetInput("Scale");auto scale_name = scale_info[0].name;auto scale_type = scale_info[0].dtype;if (std::find(T.begin(), T.end(), scale_type) == T.end()) {scale_name = helper_->AutoCast(scale_name, scale_type, P2ODataType::FP32);scale_type = P2ODataType::FP32;}auto layer_norm_node = helper_->MakeNode("LayerNormalization",{input_name, scale_name},{output_info[0].name});AddAttribute(layer_norm_node, "axis", begin_norm_axis_);AddAttribute(layer_norm_node, "epsilon", epsilon_);return;}std::vector<int64_t> normalized_shape;for (int64_t i = begin_norm_axis_;i < input_shape.size();i++) {normalized_shape.emplace_back(input_shape[i]);}if (has_input_Bias) {auto bias_info = GetInput("Bias");auto bias_name = bias_info[0].name;auto bias_type = bias_info[0].dtype;if (std::find(T.begin(), T.end(), bias_type) == T.end()) {bias_name = helper_->AutoCast(bias_name, bias_type, P2ODataType::FP32);bias_type = P2ODataType::FP32;}std::string scale_name = helper_->Constant(normalized_shape, GetOnnxDtype(P2ODataType::FP32), static_cast<float>(1.0));auto layer_norm_node = helper_->MakeNode("LayerNormalization",{input_name, scale_name, bias_name},{output_info[0].name});AddAttribute(layer_norm_node, "axis", begin_norm_axis_);AddAttribute(layer_norm_node, "epsilon", epsilon_);return;}if (!has_input_Bias && !has_input_Scale) {std::string scale_name = helper_->Constant(normalized_shape, GetOnnxDtype(P2ODataType::FP32), static_cast<float>(1.0));auto layer_norm_node = helper_->MakeNode("LayerNormalization",{input_name, scale_name},{output_info[0].name});AddAttribute(layer_norm_node, "axis", begin_norm_axis_);AddAttribute(layer_norm_node, "epsilon", epsilon_);}
}

3.2 升级并拆分Reduce算子

在OP版本升级到18后，Reduce算子的输入发生了变化，因此需要对Reduce系列的算子进行全面的升级，为了不将所有的代码耦合在一起，我将他们做了拆分，具体代码如下:

#include "paddle2onnx/mapper/tensor/reduce_logsumexp.h"namespace paddle2onnx {
REGISTER_MAPPER(logsumexp, ReduceLogSumExpMapper)int32_t ReduceLogSumExpMapper::GetMinOpset(bool verbose) {constexpr int op_version = 11;Logger(verbose, op_version) << RequireOpset(op_version) << std::endl;return op_version;
}void ReduceLogSumExpMapper::Opset18() {GetAttr("keepdim", &keep_dim_);GetAttr("reduce_all", &reduce_all_);GetAttr("axis", &dim_);auto x_info = GetInput("X");std::string dims;if (!reduce_all_) {dims = helper_->Constant(ONNX_NAMESPACE::TensorProto::INT64, dim_);} else {dims = helper_->Constant(ONNX_NAMESPACE::TensorProto::INT64, Arange(0, x_info[0].Rank()));}std::string input_name = x_info[0].name;auto input_tpye = x_info[0].dtype;if (x_info[0].dtype == P2ODataType::BOOL) {input_name = helper_->AutoCast(input_name, input_tpye, P2ODataType::INT32);input_tpye = P2ODataType::INT32;}auto reduce_node = helper_->MakeNode("ReduceLogSumExp", {input_name, dims});// Add attributeAddAttribute(reduce_node, "keepdims", static_cast<int64_t>(keep_dim_));auto out_node_name = reduce_node->output(0);bool reduce_all_axes = dim_.size() == x_info[0].Rank();if (reduce_all_) {reduce_all_axes = true;}if (!keep_dim_ && reduce_all_axes) {out_node_name = helper_->Reshape(out_node_name, {-1});}auto out_info = GetOutput("Out");helper_->AutoCast(out_node_name, out_info[0].name, input_tpye, out_info[0].dtype);
}
}  // namespace paddle2onnx

#include "paddle2onnx/mapper/tensor/reduce_max.h"namespace paddle2onnx {
REGISTER_MAPPER(reduce_max, ReduceMaxMapper)
REGISTER_MAPPER(reduce_any, ReduceMaxMapper)int32_t ReduceMaxMapper::GetMinOpset(bool verbose) {int op_version = 11;auto x_info = GetInput("X");if (x_info[0].dtype == P2ODataType::FP64) {op_version = 12;}Logger(verbose, op_version) << RequireOpset(op_version) << std::endl;return op_version;
}void ReduceMaxMapper::Opset18() {GetAttr("keep_dim", &keep_dim_);GetAttr("reduce_all", &reduce_all_);GetAttr("in_dtype", &in_dtype_);GetAttr("out_dtype", &out_dtype_);GetAttr("dim", &dim_);auto x_info = GetInput("X");std::string dims;if (!reduce_all_) {dims = helper_->Constant(ONNX_NAMESPACE::TensorProto::INT64, dim_);} else {dims = helper_->Constant(ONNX_NAMESPACE::TensorProto::INT64, Arange(0, x_info[0].Rank()));}std::string input_name = x_info[0].name;auto input_tpye = x_info[0].dtype;if (x_info[0].dtype == P2ODataType::BOOL) {input_name = helper_->AutoCast(input_name, input_tpye, P2ODataType::INT32);input_tpye = P2ODataType::INT32;}auto reduce_node = helper_->MakeNode("ReduceMax", {input_name, dims});// Add attributeAddAttribute(reduce_node, "keepdims", static_cast<int64_t>(keep_dim_));auto out_node_name = reduce_node->output(0);bool reduce_all_axes = dim_.size() == x_info[0].Rank();if (reduce_all_) {reduce_all_axes = true;}if (!keep_dim_ && reduce_all_axes) {out_node_name = helper_->Reshape(out_node_name, {-1});}auto out_info = GetOutput("Out");helper_->AutoCast(out_node_name, out_info[0].name, input_tpye, out_info[0].dtype);
}void ReduceMaxMapper::Opset12() {// The implementation logic of Opset12 is the same as that of Opset11, with the difference being that Opset12 supports input data types as double.Opset11();
}
} // namespace paddle2onnx

#include "paddle2onnx/mapper/tensor/reduce_min.h"namespace paddle2onnx {
REGISTER_MAPPER(reduce_min, ReduceMinMapper)
REGISTER_MAPPER(reduce_all, ReduceMinMapper)int32_t ReduceMinMapper::GetMinOpset(bool verbose) {int op_version = 11;auto x_info = GetInput("X");if (x_info[0].dtype == P2ODataType::FP64) {op_version = 12;}Logger(verbose, op_version) << RequireOpset(op_version) << std::endl;return op_version;
}void ReduceMinMapper::Opset18() {GetAttr("keep_dim", &keep_dim_);GetAttr("reduce_all", &reduce_all_);GetAttr("in_dtype", &in_dtype_);GetAttr("out_dtype", &out_dtype_);GetAttr("dim", &dim_);auto x_info = GetInput("X");std::string dims;if (!reduce_all_) {dims = helper_->Constant(ONNX_NAMESPACE::TensorProto::INT64, dim_);} else {dims = helper_->Constant(ONNX_NAMESPACE::TensorProto::INT64, Arange(0, x_info[0].Rank()));}auto input_node_name = x_info[0].name;auto input_tpye = x_info[0].dtype;if (x_info[0].dtype == P2ODataType::BOOL) {input_node_name = helper_->AutoCast(x_info[0].name, x_info[0].dtype, P2ODataType::INT32);input_tpye = P2ODataType::INT32;}// Add attributeauto reduce_node = helper_->MakeNode("ReduceMin", {input_node_name, dims});AddAttribute(reduce_node, "keepdims", static_cast<int64_t>(keep_dim_));auto out_node_name = reduce_node->output(0);bool reduce_all_axes = dim_.size() == x_info[0].Rank();if (reduce_all_) {reduce_all_axes = true;}if (!keep_dim_ && reduce_all_axes) {out_node_name = helper_->Reshape(out_node_name, {-1});}auto out_info = GetOutput("Out");helper_->AutoCast(out_node_name, out_info[0].name, input_tpye, out_info[0].dtype);
}void ReduceMinMapper::Opset12() {// The implementation logic of Opset12 is the same as that of Opset11, with the difference being that Opset12 supports input data types as double.Opset11();
}
}  // namespace paddle2onnx

#include "paddle2onnx/mapper/tensor/reduce_mean.h"namespace paddle2onnx {
REGISTER_MAPPER(reduce_mean, ReduceMeanMapper)int32_t ReduceMeanMapper::GetMinOpset(bool verbose) {constexpr int op_version = 11;Logger(verbose, op_version) << RequireOpset(op_version) << std::endl;return op_version;
}void ReduceMeanMapper::Opset18() {auto axis_name_ = "dim";GetAttr("keep_dim", &keep_dim_);GetAttr("reduce_all", &reduce_all_);GetAttr("in_dtype", &in_dtype_);GetAttr("out_dtype", &out_dtype_);if (IsAttrVar(axis_name_)) {auto info = GetAttrVar(axis_name_);TryGetValue(info[0], &dim_);} else {GetAttr(axis_name_, &dim_);}auto x_info = GetInput("X");std::string dims;if (IsAttrVar(axis_name_)) {auto info = GetAttrVar(axis_name_);dims = helper_->AutoCast(info[0].name, info[0].dtype, P2ODataType::INT64);} else {if (!reduce_all_) {dims = helper_->Constant(ONNX_NAMESPACE::TensorProto::INT64, dim_);} else {dims = helper_->Constant(ONNX_NAMESPACE::TensorProto::INT64, Arange(0, x_info[0].Rank()));}}// Add attributeauto reduce_node = helper_->MakeNode("ReduceMean", {x_info[0].name, dims});AddAttribute(reduce_node, "keepdims", static_cast<int64_t>(keep_dim_));auto out_node_name = reduce_node->output(0);bool reduce_all_axes = dim_.size() == x_info[0].Rank();if (reduce_all_) {reduce_all_axes = true;}if (!keep_dim_ && reduce_all_axes) {out_node_name = helper_->Reshape(out_node_name, {-1});}auto out_info = GetOutput("Out");helper_->AutoCast(out_node_name, out_info[0].name, x_info[0].dtype, out_info[0].dtype);
}
}  // namespace paddle2onnx

#include "paddle2onnx/mapper/tensor/reduce_sum.h"namespace paddle2onnx {
REGISTER_MAPPER(reduce_sum, ReduceMapperSum)int32_t ReduceMapperSum::GetMinOpset(bool verbose) {constexpr int op_version = 13;Logger(verbose, op_version) << RequireOpset(op_version) << std::endl;return op_version;
}void ReduceMapperSum::Opset13() {auto axis_name_ = "dim";GetAttr("keep_dim", &keep_dim_);GetAttr("reduce_all", &reduce_all_);GetAttr("in_dtype", &in_dtype_);GetAttr("out_dtype", &out_dtype_);if (IsAttrVar(axis_name_)) {auto info = GetAttrVar(axis_name_);TryGetValue(info[0], &dim_);} else {GetAttr(axis_name_, &dim_);}auto x_info = GetInput("X");std::string dims;if (IsAttrVar(axis_name_)) {auto info = GetAttrVar(axis_name_);dims = helper_->AutoCast(info[0].name, info[0].dtype, P2ODataType::INT64);} else {if (!reduce_all_) {dims = helper_->Constant(ONNX_NAMESPACE::TensorProto::INT64, dim_);} else {dims = helper_->Constant(ONNX_NAMESPACE::TensorProto::INT64, Arange(0, x_info[0].Rank()));}}// Add attributeauto reduce_node = helper_->MakeNode("ReduceSum", {x_info[0].name, dims});AddAttribute(reduce_node, "keepdims", static_cast<int64_t>(keep_dim_));auto out_node_name = reduce_node->output(0);bool reduce_all_axes = dim_.size() == x_info[0].Rank();if (reduce_all_) {reduce_all_axes = true;}if (!keep_dim_ && reduce_all_axes) {out_node_name = helper_->Reshape(out_node_name, {-1});}auto out_info = GetOutput("Out");helper_->AutoCast(out_node_name, out_info[0].name, x_info[0].dtype, out_info[0].dtype);
}
}  // namespace paddle2onnx

由于篇幅的限制，这里仅放出了部分核心代码，更多代码请到 Paddle2ONNX PR 1250 内查看。

4 升级CI机制

原先的CI机制对所有的opset都进行验证，这在高等级opset上是不合适的，所以对原先的CI机制进行了升级，仅对指定的opset进行验证

4.1 升级 onnxbase

class APIOnnx(object):def run(self):"""1. use dygraph layer to make exp2. dygraph layer to onnx3. use onnx to make res4. compare diff"""self._mkdir()self.set_input_spec()for place in self.places:paddle.set_device(place)exp = self._mk_dygraph_exp(self._func)res_fict = {}assert len(self.ops) <= 1, "Need to make sure the number of ops in config is 1."# Save Paddle Inference modelif os.path.exists(self.name):shutil.rmtree(self.name)paddle.jit.save(self._func, os.path.join(self.name, "model"), self.input_spec)# Get PaddleInference model pathpdmodel_path = os.path.join(self.name, "model.pdmodel")pdiparams_path = os.path.join(self.name, "model.pdiparams")if len(self.ops) > 0:self.dev_check_ops(self.ops[0], pdmodel_path)original_model_file = pdmodel_pathparams_file = pdiparams_pathif not os.path.exists(params_file):params_file = ""# clip extramodel_file = os.path.join(self.name, "cliped_model.pdmodel")self.clip_extra_program_only(original_model_file, model_file)for v in self._version:onnx_model_str = c_p2o.export(model_file, params_file, v, False, True, True, True,True, {}, "onnxruntime", "", "", False)with open(os.path.join(self.name, self.name + '_' + str(v) + ".onnx"), "wb") as f:f.write(onnx_model_str)res_fict[str(v)] = self._mk_onnx_res(ver=v)for v in self._version:compare(res_fict[str(v)], exp, delta=self.delta, rtol=self.rtol)

4.2 升级单测代码

由于篇幅的限制，更多代码请到 Paddle2ONNX PR 1250 内查看。

5 参考链接

  • [ONNX][Update] Update ONNX Version to 1.13.1 by Zheng-Bicheng · Pull Request #1250 · PaddlePaddle/Paddle2ONNX
  • ONNX Operators List
  • ONNX Operators ReduceMax
  • ONNX Operators ReduceLogSum
  • ONNX Operators ReduceMin
  • ONNX Operators ReduceProd
  • ONNX Operators ReduceSum

相关文章：

  • 免费，Python蓝桥杯等级考试真题--第12级（含答案解析和代码）
  • el-table 实现嵌套表格的思路及完整功能代码
  • “胖东来”超市商业模式，为何被誉为中国零售业是神一般的存在？
  • C++的第一道门坎:类与对象（一）
  • 用队列实现栈 用栈实现队列 设计循环队列
  • Python 基于深度图、RGB图生成RGBD点云数据
  • 代码随想录算法训练营第36期DAY42
  • 深入解析：Element Plus 与 Vite、Nuxt、Laravel 的结合使用
  • 【Linux】centos7下载安装Python3.10，下载安装openssl1.1.1
  • RabbitMQ（一）概述第一个应用程序
  • 视频监控平台AS-V1000产品介绍：账户或用户数据的导入和导出功能介绍
  • ts:条件类型
  • MySQL：如果用left join的话，左边的表一定是驱动表吗
  • Diffusion Policy:基于扩散模型的机器人动作生成策略
  • CLIP源码详解：clip.py 文件
  • [js高手之路]搞清楚面向对象,必须要理解对象在创建过程中的内存表示
  • 【个人向】《HTTP图解》阅后小结
  • 【挥舞JS】JS实现继承，封装一个extends方法
  • Django 博客开发教程 16 - 统计文章阅读量
  • Effective Java 笔记（一）
  • Javascript Math对象和Date对象常用方法详解
  • k8s如何管理Pod
  • SpringCloud集成分布式事务LCN (一)
  • vue2.0项目引入element-ui
  • Vue全家桶实现一个Web App
  • webpack入门学习手记（二）
  • 从0实现一个tiny react（三）生命周期
  • 可能是历史上最全的CC0版权可以免费商用的图片网站
  • 老板让我十分钟上手nx-admin
  • 类orAPI - 收藏集 - 掘金
  • const的用法,特别是用在函数前面与后面的区别
  • MyCAT水平分库
  • ​​​​​​​sokit v1.3抓手机应用socket数据包： Socket是传输控制层协议，WebSocket是应用层协议。
  • # Maven错误Error executing Maven
  • #{} 和 ${}区别
  • #stm32驱动外设模块总结w5500模块
  • #我与Java虚拟机的故事#连载09：面试大厂逃不过的JVM
  • (cljs/run-at (JSVM. :browser) 搭建刚好可用的开发环境!)
  • (el-Transfer)操作(不使用 ts)：Element-plus 中 Select 组件动态设置 options 值需求的解决过程
  • （MTK）java文件添加简单接口并配置相应的SELinux avc 权限笔记2
  • (Spark3.2.0)Spark SQL 初探： 使用大数据分析2000万KF数据
  • （二）hibernate配置管理
  • （附源码）spring boot火车票售卖系统 毕业设计 211004
  • (附源码)计算机毕业设计SSM智能化管理的仓库管理
  • （三）Pytorch快速搭建卷积神经网络模型实现手写数字识别(代码+详细注解)
  • （数据大屏）（Hadoop）基于SSM框架的学院校友管理系统的设计与实现+文档
  • (原創) 博客園正式支援VHDL語法著色功能 (SOC) (VHDL)
  • （转） ns2/nam与nam实现相关的文件
  • ＊Django中的Ajax 纯js的书写样式1
  • .NET 2.0中新增的一些TryGet,TryParse等方法
  • .NET Core 和 .NET Framework 中的 MEF2
  • .Net Core 生成管理员权限的应用程序
  • .NET/C# 利用 Walterlv.WeakEvents 高性能地定义和使用弱事件
  • .sdf和.msp文件读取
  • @EnableConfigurationProperties注解使用