pytorch转onnx转mnn并验证

2023-12-18 02:58
文章标签 验证 pytorch onnx mnn

本文主要是介绍pytorch转onnx转mnn并验证，希望对大家解决编程问题提供一定的参考价值，需要的开发者们随着小编来一起学习吧！

pytorch训练的模型在实际使用时往往需要转换成onnx或mnn部署，训练好的模型需先转成onnx：

import sys
import argparse
import torch
import torchvision
import torch.onnxfrom  mobilenetv2  import MobileNetV2if __name__ == '__main__':model=MobileNetV2(2)model_path='./model/mobilenetv2.mdl'model.eval()model.load_state_dict(torch.load(model_path, map_location=torch.device('cpu')))dummy_input = torch.randn([1,3,32,32])   #batch,channel,height,widthtorch.onnx.export(model, dummy_input, model_path.replace('mdl', 'onnx'), verbose=True, input_names=['input'], output_names=['output'],opset_version=11)print('Done!')

转换成功后，再转mnn，通过MNN转换工具:

.MNNConvert -f ONNX --modelFile XXX.onnx --MNNModel XXX.mnn --bizCode biz

测试pytorch的结果：

import argparse
import os
from glob import glob
import cv2
import torch
import torch.backends.cudnn as cudnn
import yaml
from torch.utils.data import Dataset, DataLoader
from torchvision import transforms
from tqdm import tqdm
from  PIL import Image
from  mobilenetv2  import MobileNetV2
import numpy as npdef parse_args():parser = argparse.ArgumentParser()parser.add_argument('--image_path', default=None,help='the path of imgae')args = parser.parse_args()return argsdef main():args = parse_args()device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")start = cv2.getTickCount()# create modelmodel = MobileNetV2(2).to(device)model.load_state_dict(torch.load('models/best-mobilenetv2.mdl',map_location=torch.device('cpu')))model.eval()img = args.image_pathcut_size = 48tf = transforms.Compose([lambda x: Image.open(x).convert('RGB'),  # string path= > image datatransforms.ToTensor(),transforms.Normalize(mean=[0.485, 0.456, 0.406],std=[0.229, 0.224, 0.225])])img = tf(img).unsqueeze(0)x = img.to(device)outputs = model(x)# 输出概率最大的类别_, indices = torch.max(outputs, 1)percentage = torch.nn.functional.softmax(outputs, dim=1)[0] * 100perc = percentage[int(indices)].item()print('predicted:', perc)print('id:', int(indices))end = cv2.getTickCount()during = (end - start) / cv2.getTickFrequency()print("avg_time:", during)if __name__ == '__main__':main()

测试ONNX的结果,与pytorch结果一致：

import argparse
import os
from glob import glob
import onnxruntime
import onnx
import cv2
import torch
import torch.backends.cudnn as cudnn
import yaml
from torch.utils.data import Dataset, DataLoader
from torchvision import transforms
from tqdm import tqdm
from  PIL import Image
from  mobilenetv2  import MobileNetV2
import numpy as npdef parse_args():parser = argparse.ArgumentParser()parser.add_argument('--image_path', default=None,help='the path of imgae')args = parser.parse_args()return argsdef to_numpy(tensor):return tensor.detach().cpu.numpy() if tensor.requires_grad else tensor.cpu().numpy()def main():args = parse_args()device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")start = cv2.getTickCount()model = 'models/best-mobilenetv2.onnx'onet_seeion=onnxruntime.InferenceSession(model)img = args.image_pathcut_size = 48tf = transforms.Compose([lambda x: Image.open(x).convert('RGB'),  # string path= > image datatransforms.ToTensor(),transforms.Normalize(mean=[0.485, 0.456, 0.406],std=[0.229, 0.224, 0.225])])img = tf(img).unsqueeze(0)x = img.to(device)inputs={onet_seeion.get_inputs()[0].name:to_numpy(img)}outputs=onet_seeion.run(None,inputs)print(outputs)end = cv2.getTickCount()during = (end - start) / cv2.getTickFrequency()print("avg_time:", during)if __name__ == '__main__':main()

测试mnn的结果,与前面的结果一致，但是速度快了近20倍：

import argparse
import os
from glob import glob
import MNN
import cv2
import torch
import torch.backends.cudnn as cudnn
import yaml
from torch.utils.data import Dataset, DataLoader
from torchvision import transforms
from tqdm import tqdm
from  PIL import Image
from  mobilenetv2  import MobileNetV2
import numpy as npdef parse_args():parser = argparse.ArgumentParser()parser.add_argument('--image_path', default=None,help='the path of imgae')args = parser.parse_args()return argsdef to_numpy(tensor):return tensor.detach().cpu.numpy() if tensor.requires_grad else tensor.cpu().numpy()def main():args = parse_args()device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")start = cv2.getTickCount()model = 'models/best-mobilenetv2.mnn'interpreter = MNN.Interpreter(model)mnn_session = interpreter.createSession()input_tensor = interpreter.getSessionInput(mnn_session)img = args.image_pathcut_size = 48tf = transforms.Compose([lambda x: Image.open(x).convert('RGB'),  # string path= > image datatransforms.ToTensor(),transforms.Normalize(mean=[0.485, 0.456, 0.406],std=[0.229, 0.224, 0.225])])img = tf(img).unsqueeze(0)tmp_input = MNN.Tensor((1, 3, 32, 32), MNN.Halide_Type_Float, \to_numpy(img[0]), MNN.Tensor_DimensionType_Caffe)print(tmp_input.getShape())# print(tmp_input.getData())print(input_tensor.copyFrom(tmp_input))input_tensor.printTensorData()interpreter.runSession(mnn_session)output_tensor = interpreter.getSessionOutput(mnn_session, 'output')output_tensor.printTensorData()output_data = np.array(output_tensor.getData())print('mnn result is:', output_data)print("output belong to class: {}".format(np.argmax(output_tensor.getData())))end = cv2.getTickCount()during = (end - start) / cv2.getTickFrequency()print("avg_time:", during)if __name__ == '__main__':main()

用c++进行mnn重写测试,结果一致，这样就可以编库了：

// mnn_test.cpp : 定义控制台应用程序的入口点。#include "stdafx.h"
#include <iostream>
#include <opencv2/opencv.hpp>
#include <MNN/Interpreter.hpp>
#include <MNN/MNNDefine.h>
#include <MNN/Tensor.hpp>
#include <MNN/ImageProcess.hpp>
#include <memory>#define IMAGE_VERIFY_SIZE 32
#define CLASSES_SIZE 2
#define INPUT_NAME "input"
#define OUTPUT_NAME "output"cv::Mat BGRToRGB(cv::Mat img)
{cv::Mat image(img.rows, img.cols, CV_8UC3);for (int i = 0; i<img.rows; ++i) {cv::Vec3b *p1 = img.ptr<cv::Vec3b>(i);cv::Vec3b *p2 = image.ptr<cv::Vec3b>(i);for (int j = 0; j<img.cols; ++j) {p2[j][2] = p1[j][0];p2[j][1] = p1[j][1];p2[j][0] = p1[j][2];}}return image;
}int main(int argc, char* argv[]) {if (argc < 2) {printf("Usage:\n\t%s mnn_model_path image_path\n", argv[0]);return -1;}// create net and sessionconst char *mnn_model_path = argv[1];const char *image_path = argv[2];auto mnnNet = std::shared_ptr<MNN::Interpreter>(MNN::Interpreter::createFromFile(mnn_model_path));MNN::ScheduleConfig netConfig;netConfig.type = MNN_FORWARD_CPU;netConfig.numThread = 4;auto session = mnnNet->createSession(netConfig);auto input = mnnNet->getSessionInput(session, INPUT_NAME);if (input->elementSize() <= 4) {mnnNet->resizeTensor(input, { 1, 3, IMAGE_VERIFY_SIZE, IMAGE_VERIFY_SIZE });mnnNet->resizeSession(session);}std::cout << "input shape: " << input->shape()[0] << " " << input->shape()[1] << " " << input->shape()[2] << " " << input->shape()[3] << std::endl;// preprocess imageMNN::Tensor givenTensor(input, MNN::Tensor::CAFFE);// const int inputSize = givenTensor.elementSize();// std::cout << inputSize << std::endl;auto inputData = givenTensor.host<float>();cv::Mat bgr_image = cv::imread(image_path);bgr_image = BGRToRGB(bgr_image);cv::Mat norm_image;cv::resize(bgr_image, norm_image, cv::Size(IMAGE_VERIFY_SIZE, IMAGE_VERIFY_SIZE));for (int k = 0; k < 3; k++) {for (int i = 0; i < norm_image.rows; i++) {for (int j = 0; j < norm_image.cols; j++) {const auto src = norm_image.at<cv::Vec3b>(i, j)[k];auto dst = 0.0;if (k == 0) dst = (float(src) / 255.0f - 0.485) / 0.229;if (k == 1) dst = (float(src) / 255.0f - 0.456) / 0.224;if (k == 2) dst = (float(src) / 255.0f - 0.406) / 0.225;inputData[k * IMAGE_VERIFY_SIZE * IMAGE_VERIFY_SIZE + i * IMAGE_VERIFY_SIZE + j] = dst;}}}input->copyFromHostTensor(&givenTensor);double st = cvGetTickCount();// run sessionmnnNet->runSession(session);double et = cvGetTickCount() - st;et = et / ((double)cvGetTickFrequency() * 1000);std::cout << " speed: " << et << " ms" << std::endl;// get output dataauto output = mnnNet->getSessionOutput(session, OUTPUT_NAME);// std::cout << "output shape: " << output->shape()[0] << " " << output->shape()[1] << std::endl;auto output_host = std::make_shared<MNN::Tensor>(output, MNN::Tensor::CAFFE);output->copyToHostTensor(output_host.get());auto values = output_host->host<float>();// post processstd::vector<float> output_values;auto exp_sum = 0.0;auto max_index = 0;for (int i = 0; i < CLASSES_SIZE; i++) {if (values[i] > values[max_index]) max_index = i;output_values.push_back(values[i]);exp_sum += std::exp(values[i]);}std::cout << "output: " << output_values[0]<<","<< output_values[1] << std::endl;std::cout << "id: " << max_index << std::endl;std::cout << "prob: " << std::exp(output_values[max_index]) / exp_sum << std::endl;system("pause");return 0;
}

 

这篇关于pytorch转onnx转mnn并验证的文章就介绍到这儿，希望我们推荐的文章对编程师们有所帮助！


http://www.chinasem.cn/article/506796。

相关文章

Spring Security基于数据库验证流程详解

Spring Security基于数据库验证流程详解

Spring Security 校验流程图 相关解释说明（认真看哦） AbstractAuthenticationProcessingFilter 抽象类 /*** 调用 #requiresAuthentication(HttpServletRequest, HttpServletResponse) 决定是否需要进行验证操作。* 如果需要验证，则会调用 #attemptAuthentica
阅读更多...
C++ | Leetcode C++题解之第393题UTF-8编码验证

C++ | Leetcode C++题解之第393题UTF-8编码验证

题目： 题解： class Solution {public:static const int MASK1 = 1 << 7;static const int MASK2 = (1 << 7) + (1 << 6);bool isValid(int num) {return (num & MASK2) == MASK1;}int getBytes(int num) {if ((num &
阅读更多...
C语言 | Leetcode C语言题解之第393题UTF-8编码验证

C语言 | Leetcode C语言题解之第393题UTF-8编码验证

题目： 题解： static const int MASK1 = 1 << 7;static const int MASK2 = (1 << 7) + (1 << 6);bool isValid(int num) {return (num & MASK2) == MASK1;}int getBytes(int num) {if ((num & MASK1) == 0) {return
阅读更多...
easyui同时验证账户格式和ajax是否存在

easyui同时验证账户格式和ajax是否存在

accountName: {validator: function (value, param) {if (!/^[a-zA-Z][a-zA-Z0-9_]{3,15}$/i.test(value)) {$.fn.validatebox.defaults.rules.accountName.message = '账户名称不合法（字母开头，允许4-16字节，允许字母数字下划线）';return fal
阅读更多...
easyui 验证下拉菜单select

easyui 验证下拉菜单select

validatebox.js中添加以下方法： selectRequired: {validator: function (value) {if (value == "" || value.indexOf('请选择') >= 0 || value.indexOf('全部') >= 0) {return false;}else {return true;}},message: '该下拉框为必选项'}
阅读更多...
web群集--nginx配置文件location匹配符的优先级顺序详解及验证

web群集--nginx配置文件location匹配符的优先级顺序详解及验证

文章目录 前言优先级顺序优先级顺序(详解)1. 精确匹配（Exact Match）2. 正则表达式匹配（Regex Match）3. 前缀匹配（Prefix Match） 匹配规则的综合应用验证优先级 前言 location的作用 在 NGINX 中，location 指令用于定义如何处理特定的请求 URI。由于网站往往需要不同的处理方式来适应各种请求，NGINX 提供了多种匹
阅读更多...
React 笔记 父子组件传值 | 父组件调用子组件数据 | defaultProps | propsType合法性验证

React 笔记 父子组件传值 | 父组件调用子组件数据 | defaultProps | propsType合法性验证

1.通过props实现父组件像子组件传值 、方法、甚至整个父组件 传递整个父组件则   [变量名]={this} import Header from "./Header"render(){return(<Header msg={"我是props传递的数据"}/>)} import React,{Component} from "react";class Header extends
阅读更多...
Nn criterions don’t compute the gradient w.r.t. targets error「pytorch」 (debug笔记)

Nn criterions don’t compute the gradient w.r.t. targets error「pytorch」 (debug笔记)

Nn criterions don’t compute the gradient w.r.t. targets error「pytorch」 ##一、 缘由及解决方法 把这个pytorch-ddpg|github搬到jupyter notebook上运行时，出现错误Nn criterions don’t compute the gradient w.r.t. targets error。注：我用
阅读更多...
【超级干货】2天速成PyTorch深度学习入门教程，缓解研究生焦虑

【超级干货】2天速成PyTorch深度学习入门教程，缓解研究生焦虑

3、cnn基础 卷积神经网络 输入层 —输入图片矩阵 输入层一般是 RGB 图像或单通道的灰度图像，图片像素值在[0,255]，可以用矩阵表示图片 卷积层 —特征提取 人通过特征进行图像识别，根据左图直的笔画判断X，右图曲的笔画判断圆 卷积操作 激活层 —加强特征 池化层 —压缩数据 全连接层 —进行分类 输出层 —输出分类概率 4、基于LeNet
阅读更多...
Java验证辛钦大数定理

Java验证辛钦大数定理

本实验通过程序模拟采集大量的样本数据来验证辛钦大数定理。   实验环境： 本实验采用Java语言编程，开发环境为Eclipse，图像生成使用JFreeChart类。   一，验证辛钦大数定理 由辛钦大数定理描述为： 辛钦大数定理（弱大数定理）  设随机变量序列 X1, X2, … 相互独立，服从同一分布，具有数学期望E(Xi) = μ, i = 1, 2, …， 则对于任意正数ε ，
阅读更多...

Copyright China编程 23002807@qq.com

沪ICP备2023033072号-2