第一部分 前言
衆位小夥伴，好久沒更新博客了，本次爲大家帶來：如何將PyTorch訓練的網絡通過模型轉換，部署到Windows純C++下執行，落地應用。這裏並沒有將PyTorch模型轉至其他深度學習框架下，而是通過PyTorch的LibTorch來完成相關C++的部署應用。
第二部分 說明
PyTorch版本：Torch-1.4.0-cu101
LibTorch版本：LibTorch-1.4.0-cu101
Anaconda版本：Anaconda3-Python3.6
GPU：GTX1080
VS版本：VS2017（用於編譯LibTorch）
第三部分 開源鏈接
GITHUB（如果對您有幫助，還望給個star，以鼓勵博主向着開源之路繼續前進^v^）
第四部分 訓練數據
以kaggle貓狗大戰數據集爲例，數據格式如下：
1、訓練數據路徑：data/train/cat/*.jpg，data/train/dog/*.jpg
2、驗證數據路徑：data/val/cat/*.jpg，data/val/dog/*.jpg
第五部分 訓練代碼
from __future__ import print_function, division
import torch
import torch.nn as nn
import torch.optim as optim
from torch.optim import lr_scheduler
from torchvision import datasets, models, transforms
import time
import os
import copy
def train_model(model, criterion, optimizer, scheduler, num_epochs=25):
since = time.time()
best_model_wts = copy.deepcopy(model.state_dict())
best_acc = 0.0
for epoch in range(num_epochs):
print('Epoch {}/{}'.format(epoch, num_epochs - 1))
print('learning rate: {}'.format(scheduler.get_lr()[0]))
print('-' * 10)
# Each epoch has a training and validation phase
for phase in ['train', 'val']:
if phase == 'train':
model.train() # Set model to training mode
else:
model.eval() # Set model to evaluate mode
running_loss = 0.0
running_corrects = 0
# Iterate over data.
for inputs, labels in dataloaders[phase]:
inputs = inputs.to(device)
labels = labels.to(device)
# zero the parameter gradients
optimizer.zero_grad()
# forward
# track history if only in train
with torch.set_grad_enabled(phase == 'train'):
outputs = model(inputs)
_, preds = torch.max(outputs, 1)
loss = criterion(outputs, labels)
# backward + optimize only if in training phase
if phase == 'train':
loss.backward()
optimizer.step()
# statistics
running_loss += loss.item() * inputs.size(0)
running_corrects += torch.sum(preds == labels.data)
epoch_loss = running_loss / dataset_sizes[phase]
epoch_acc = running_corrects.double() / dataset_sizes[phase]
print('{} Loss: {:.4f} Acc: {:.4f}'.format(
phase, epoch_loss, epoch_acc))
# deep copy the model
if phase == 'val' and epoch_acc > best_acc:
best_acc = epoch_acc
best_model_wts = copy.deepcopy(model.state_dict())
if phase == 'train':
scheduler.step()
time_elapsed = time.time() - since
print('Training complete in {:.0f}m {:.0f}s'.format(
time_elapsed // 60, time_elapsed % 60))
print('Best val Acc: {:4f}'.format(best_acc))
# load best model weights
model.load_state_dict(best_model_wts)
return model
# Data augmentation and normalization for training
# Just normalization for validation
data_transforms = {
'train': transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]),
'val': transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
]),
}
data_dir = 'data'
image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x),data_transforms[x]) for x in ['train', 'val']}
dataloaders = {x: torch.utils.data.DataLoader(image_datasets[x], batch_size=4,shuffle=True, num_workers=0) for x in ['train', 'val']}
dataset_sizes = {x: len(image_datasets[x]) for x in ['train', 'val']}
class_names = image_datasets['train'].classes
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# model_conv = torchvision.models.resnet18(pretrained=False)
# print(model_conv)
# for param in model_conv.parameters():
# param.requires_grad = False
model_ft = models.resnet18(pretrained=True)
num_ftrs = model_ft.fc.in_features
model_ft.fc = nn.Linear(num_ftrs, 2)
model_ft = model_ft.to(device)
criterion = nn.CrossEntropyLoss()
# Observe that all parameters are being optimized
optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9)
# Decay LR by a factor of 0.1 every 7 epochs
exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1)
model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler, num_epochs=25)
torch.save(model_ft, 'model.pkl')
第六部分 測試代碼
from __future__ import print_function, division
import torch
import torch.nn.functional as F
from torchvision import transforms
from PIL import Image
import os
classes = ['cat','dog']
test_path = "data/val/"
true_count = 0
all_count = 0
for test_dir in os.listdir(test_path):
test_dir_path = test_path + test_dir + "/"
for img_names in os.walk(test_dir_path):
for img_name in img_names[2]:
img_path = test_dir_path + img_name
print(img_path)
image = Image.open(img_path)
transform = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
image_transformed = transform(image)
image_transformed = image_transformed.unsqueeze(0)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = torch.load('model.pkl')
model = model.to(device)
model.eval()
output = model(image_transformed.to(device))
output = F.softmax(output, dim=1)
predict_value, predict_idx = torch.max(output, 1)
if(classes[predict_idx.cpu().data[0].numpy()] == test_dir):
true_count += 1
all_count += 1
print("acc: {}/{}={}".format(true_count,all_count,float(true_count)/float(all_count)))
#acc: 1966/2000=0.983
第七部分 模型轉換代碼
"""
This python script converts the network into Script Module---CPU
"""
import torch
# Download and load the pre-trained model
model = torch.load("model.pkl",map_location='cpu')
model.eval()
example_input = torch.rand(1, 3, 224, 224)
script_module = torch.jit.trace(model, example_input)
script_module.save('model_cpu.pt')
#"""
#This python script converts the network into Script Module---GPU
#"""
#import torch
#
#device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
#
## Download and load the pre-trained model
#model = torch.load("model.pkl")
#
#model.eval()
#
#example_input = torch.rand(1, 3, 224, 224)
#script_module = torch.jit.trace(model, example_input.to(device))
#script_module.save('model_gpu.pt')
第八部分 LibTorch_C++實現代碼
#include <torch/torch.h>
#include <torch/script.h> // One-stop header.
#include <iostream>
#include <memory>
#include <opencv2/opencv.hpp>
using namespace std;
using namespace cv;
Mat pilResize(Mat &img, int size) {
int imgWidth = img.cols;
int imgHeight = img.rows;
if ((imgWidth <= imgHeight && imgWidth == size) || (imgHeight <= imgWidth && imgHeight == size)) {
return img;
}
Mat output;
if (imgWidth < imgHeight) {
int outWidth = size;
int outHeight = int(size * imgHeight / (float)imgWidth);
resize(img, output, Size(outWidth, outHeight));
}
else {
int outHeight = size;
int outWidth = int(size * imgWidth / (float)imgHeight);
resize(img, output, Size(outWidth, outHeight));
}
return output;
}
Mat pilCropCenter(Mat &img, int output_size) {
Rect imgRect;
imgRect.x = int(round((img.cols - output_size) / 2.));
imgRect.y = int(round((img.rows - output_size) / 2.));
imgRect.width = output_size;
imgRect.height = output_size;
return img(imgRect).clone();
}
Mat setNorm(Mat &img) {
Mat img_rgb;
cvtColor(img, img_rgb, COLOR_RGB2BGR);
Mat img_resize = pilResize(img_rgb, 256);
Mat img_crop = pilCropCenter(img_resize, 224);
Mat image_resized_float;
img_crop.convertTo(image_resized_float, CV_32F, 1.0 / 255.0);
return image_resized_float;
}
Mat setMean(Mat &image_resized_float) {
vector<float> mean = { 0.485, 0.456, 0.406 };
vector<float> std = { 0.229, 0.224, 0.225 };
vector<Mat> image_resized_split;
split(image_resized_float, image_resized_split);
for (int ch = 0; ch < image_resized_split.size(); ch++) {
image_resized_split[ch] -= mean[ch];
image_resized_split[ch] /= std[ch];
}
Mat image_resized_merge;
merge(image_resized_split, image_resized_merge);
return image_resized_merge;
}
int main() {
torch::DeviceType device_type;
if (torch::cuda::is_available()) {
std::cout << "CUDA available! Test on GPU." << std::endl;
device_type = torch::kCUDA;
}
else {
std::cout << "Test on CPU." << std::endl;
device_type = torch::kCPU;
}
torch::Device device(device_type);
// Deserialize the ScriptModule from a file using torch::jit::load().
torch::jit::script::Module model = torch::jit::load("model_cpu.pt");
model.to(device);
vector<string> classes = { "cat","dog" };
string test_path = "val/dog/";
vector<string> img_paths;
glob(test_path, img_paths);
int truth_count = 0;
for (int i = 0; i < img_paths.size(); i++) {
Mat img = imread(img_paths[i]);
clock_t start_t = clock();
//norm
Mat image_resized_float = setNorm(img);
//mean
Mat image_resized_merge = setMean(image_resized_float);
auto img_tensor = torch::from_blob(image_resized_merge.data, { 224, 224, 3 }, torch::kFloat32);
auto img_tensor_ = torch::unsqueeze(img_tensor, 0);
img_tensor_ = img_tensor_.permute({ 0, 3, 1, 2 });
// Create a vector of inputs.
vector<torch::jit::IValue> inputs;
inputs.push_back(img_tensor_.to(device));
torch::Tensor prob = model.forward(inputs).toTensor();
torch::Tensor output = torch::softmax(prob, 1);
auto predict = torch::max(output, 1);
//cout << "cost time:" << clock() - start_t << endl;
cout << img_paths[i] << "\t";
cout << "class: " << classes[get<1>(predict).item<int>()] <<
", prob: " << get<0>(predict).item<float>() << endl;
if (get<1>(predict).item<int>() == 1) {
truth_count++;
}
}
cout << truth_count << "/" << img_paths.size() << endl;
system("pause");
return 0;
}
第九部分 相關說明
1、Pytorch默認通過PIL載入圖像數據，這點很重要！
2、需要載入ResNet預訓練模型，否則訓練效果較差！
3、PyTorch訓練出來的模型格式爲pkl，需要將其轉換爲pt格式，C++方能採用torch::jit::load方式載入。
4、轉換時，有CPU和GPU兩種方式，C++實現時可通過兩種方式載入。
5、C++實現時，需將PIL格式的圖像轉換爲OpenCV的圖像，否則數據不統一，導致測試結果不正確！
任何問題請加唯一QQ2258205918（名稱samylee）！