AI学习者 · 3月15日

PyTorch ResNet实现图像分类(从模型的训练到Android部署)

1.数据集

数据集地址:10 Monkey Species

采用kaggle上的猴子数据集,包含两个文件:训练集和验证集。每个文件夹包含10个标记为n0-n9的猴子。图像尺寸为400x300像素或更大,并且为JPEG格式(近1400张图像)。

image.png

图片样本

image.png

图片类别标签,训练集,验证集划分说明

image.png

2.代码

2.1 定义需要的库

import os
import sys
import json
import time
import torch
import torch.nn as nn
import torch.optim as optim
from torchvision import transforms, datasets
from tqdm import tqdm
import torch.nn.functional as F
import numpy as np
import matplotlib.pyplot as plt
from sklearn.metrics import accuracy_score

2.2 定义训练验证函数

def train_and_val(epochs, model, train_loader, val_loader, criterion, optimizer):
    torch.cuda.empty_cache()
    train_loss = []
    val_loss = []
    train_acc = []
    val_acc = []
    best_acc = 0

    model.to(device)
    fit_time = time.time()
    for e in range(epochs):
        since = time.time()
        running_loss = 0
        training_acc = 0
        with tqdm(total=len(train_loader)) as pbar:
            for image, label in train_loader:
                # training phase

                #                 images, labels = data
                #             optimizer.zero_grad()
                #             logits = net(images.to(device))
                #             loss = loss_function(logits, labels.to(device))
                #             loss.backward()
                #             optimizer.step()

                model.train()
                optimizer.zero_grad()
                image = image.to(device)
                label = label.to(device)
                # forward
                output = model(image)
                loss = criterion(output, label)
                predict_t = torch.max(output, dim=1)[1]

                # backward
                loss.backward()
                optimizer.step()  # update weight

                running_loss += loss.item()
                training_acc += torch.eq(predict_t, label).sum().item()
                pbar.update(1)

        model.eval()
        val_losses = 0
        validation_acc = 0
        # validation loop
        with torch.no_grad():
            with tqdm(total=len(val_loader)) as pb:
                for image, label in val_loader:
                    image = image.to(device)
                    label = label.to(device)
                    output = model(image)

                    # loss
                    loss = criterion(output, label)
                    predict_v = torch.max(output, dim=1)[1]

                    val_losses += loss.item()
                    validation_acc += torch.eq(predict_v, label).sum().item()
                    pb.update(1)

            # calculatio mean for each batch
            train_loss.append(running_loss / len(train_dataset))
            val_loss.append(val_losses / len(val_dataset))

            train_acc.append(training_acc / len(train_dataset))
            val_acc.append(validation_acc / len(val_dataset))
            
            torch.save(model, "last.pth")
            if best_acc<(validation_acc / len(val_dataset)):
                torch.save(model, "best.pth")
            

            print("Epoch:{}/{}..".format(e + 1, epochs),
                  "Train Acc: {:.3f}..".format(training_acc / len(train_dataset)),
                  "Val Acc: {:.3f}..".format(validation_acc / len(val_dataset)),
                  "Train Loss: {:.3f}..".format(running_loss / len(train_dataset)),
                  "Val Loss: {:.3f}..".format(val_losses / len(val_dataset)),
                  "Time: {:.2f}s".format((time.time() - since)))
            

    history = {'train_loss': train_loss, 'val_loss': val_loss,'train_acc': train_acc, 'val_acc': val_acc}
    print('Total time: {:.2f} m'.format((time.time() - fit_time) / 60))
    
    return history

2.3定义ResNet网络

class BasicBlock(nn.Module):
    expansion = 1

    def __init__(self, in_channel, out_channel, stride=1, downsample=None, **kwargs):
        super(BasicBlock, self).__init__()
        self.conv1 = nn.Conv2d(in_channels=in_channel, out_channels=out_channel,
                               kernel_size=3, stride=stride, padding=1, bias=False)
        self.bn1 = nn.BatchNorm2d(out_channel)
        self.relu = nn.ReLU()
        self.conv2 = nn.Conv2d(in_channels=out_channel, out_channels=out_channel,
                               kernel_size=3, stride=1, padding=1, bias=False)
        self.bn2 = nn.BatchNorm2d(out_channel)
        self.downsample = downsample

    def forward(self, x):
        identity = x
        if self.downsample is not None:
            identity = self.downsample(x)

        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)

        out += identity
        out = self.relu(out)

        return out


class Bottleneck(nn.Module):
    expansion = 4

    def __init__(self, in_channel, out_channel, stride=1, downsample=None,
                 groups=1, width_per_group=64):
        super(Bottleneck, self).__init__()

        width = int(out_channel * (width_per_group / 64.)) * groups

        self.conv1 = nn.Conv2d(in_channels=in_channel, out_channels=width,
                               kernel_size=1, stride=1, bias=False)  # squeeze channels
        self.bn1 = nn.BatchNorm2d(width)
        # -----------------------------------------
        self.conv2 = nn.Conv2d(in_channels=width, out_channels=width, groups=groups,
                               kernel_size=3, stride=stride, bias=False, padding=1)
        self.bn2 = nn.BatchNorm2d(width)
        # -----------------------------------------
        self.conv3 = nn.Conv2d(in_channels=width, out_channels=out_channel*self.expansion,
                               kernel_size=1, stride=1, bias=False)  # unsqueeze channels
        self.bn3 = nn.BatchNorm2d(out_channel*self.expansion)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample

    def forward(self, x):
        identity = x
        if self.downsample is not None:
            identity = self.downsample(x)

        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)
        out = self.relu(out)

        out = self.conv3(out)
        out = self.bn3(out)

        out += identity
        out = self.relu(out)

        return out


class ResNet(nn.Module):

    def __init__(self,
                 block,
                 blocks_num,
                 num_classes=10,
                 include_top=True,
                 groups=1,
                 width_per_group=64):
        super(ResNet, self).__init__()
        self.include_top = include_top
        self.in_channel = 64

        self.groups = groups
        self.width_per_group = width_per_group

        self.conv1 = nn.Conv2d(3, self.in_channel, kernel_size=7, stride=2,
                               padding=3, bias=False)
        self.bn1 = nn.BatchNorm2d(self.in_channel)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
        self.layer1 = self._make_layer(block, 64, blocks_num[0])
        self.layer2 = self._make_layer(block, 128, blocks_num[1], stride=2)
        self.layer3 = self._make_layer(block, 256, blocks_num[2], stride=2)
        self.layer4 = self._make_layer(block, 512, blocks_num[3], stride=2)
        if self.include_top:
            self.avgpool = nn.AdaptiveAvgPool2d((1, 1))  # output size = (1, 1)
            self.fc = nn.Linear(512 * block.expansion, num_classes)

        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')

    def _make_layer(self, block, channel, block_num, stride=1):
        downsample = None
        if stride != 1 or self.in_channel != channel * block.expansion:
            downsample = nn.Sequential(
                nn.Conv2d(self.in_channel, channel * block.expansion, kernel_size=1, stride=stride, bias=False),
                nn.BatchNorm2d(channel * block.expansion))

        layers = []
        layers.append(block(self.in_channel,
                            channel,
                            downsample=downsample,
                            stride=stride,
                            groups=self.groups,
                            width_per_group=self.width_per_group))
        self.in_channel = channel * block.expansion

        for _ in range(1, block_num):
            layers.append(block(self.in_channel,
                                channel,
                                groups=self.groups,
                                width_per_group=self.width_per_group))

        return nn.Sequential(*layers)

    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool(x)

        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)

        if self.include_top:
            x = self.avgpool(x)
            x = torch.flatten(x, 1)
            x = self.fc(x)

        return x

def resnet34(num_classes=10, include_top=True):
    # https://download.pytorch.org/models/resnet34-333f7ec4.pth
    return ResNet(BasicBlock, [3, 4, 6, 3], num_classes=num_classes, include_top=include_top)


def resnet50(num_classes=10, include_top=True):
    return ResNet(Bottleneck, [3, 4, 6, 3], num_classes=num_classes, include_top=include_top)


def resnet101(num_classes=10, include_top=True):
    return ResNet(Bottleneck, [3, 4, 23, 3], num_classes=num_classes, include_top=include_top)


def resnext50_32x4d(num_classes=10, include_top=True):
    groups = 32
    width_per_group = 4
    return ResNet(Bottleneck, [3, 4, 6, 3],
                  num_classes=num_classes,
                  include_top=include_top,
                  groups=groups,
                  width_per_group=width_per_group)


def resnext101_32x8d(num_classes=10, include_top=True):
    groups = 32
    width_per_group = 8
    return ResNet(Bottleneck, [3, 4, 23, 3],
                  num_classes=num_classes,
                  include_top=include_top,
                  groups=groups,
                  width_per_group=width_per_group)

2.4 设置训练集和验证集

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print("using {} device.".format(device))

BATCH_SIZE = 16

data_transform = {
    "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])])}

train_dataset = datasets.ImageFolder("../input/10-monkey-species/training/training/", transform=data_transform["train"])  # 训练集数据
train_loader = torch.utils.data.DataLoader(dataset=train_dataset, batch_size=BATCH_SIZE, shuffle=True,
                                           num_workers=2)  # 加载数据

val_dataset = datasets.ImageFolder("../input/10-monkey-species/validation/validation/", transform=data_transform["val"])  # 测试集数据
val_loader = torch.utils.data.DataLoader(dataset=val_dataset, batch_size=BATCH_SIZE, shuffle=False,
                                         num_workers=2)  # 加载数据

2.5 开始训练

net = resnet34()
loss_function = nn.CrossEntropyLoss()  # 设置损失函数
optimizer = optim.Adam(net.parameters(), lr=0.0001)  # 设置优化器和学习率
epoch = 60

history = train_and_val(epoch, net, train_loader, val_loader, loss_function, optimizer)
执行结果
Epoch:55/60.. Train Acc: 0.813.. Val Acc: 0.860.. Train Loss: 0.038.. Val Loss: 0.029.. Time: 38.40s
100%|██████████| 69/69 [00:28<00:00,  2.38it/s]
100%|██████████| 17/17 [00:09<00:00,  1.81it/s]
Epoch:56/60.. Train Acc: 0.830.. Val Acc: 0.882.. Train Loss: 0.031.. Val Loss: 0.025.. Time: 38.84s
100%|██████████| 69/69 [00:27<00:00,  2.48it/s]
100%|██████████| 17/17 [00:09<00:00,  1.78it/s]
Epoch:57/60.. Train Acc: 0.843.. Val Acc: 0.871.. Train Loss: 0.031.. Val Loss: 0.025.. Time: 37.80s
100%|██████████| 69/69 [00:28<00:00,  2.39it/s]
100%|██████████| 17/17 [00:09<00:00,  1.86it/s]
Epoch:58/60.. Train Acc: 0.829.. Val Acc: 0.827.. Train Loss: 0.030.. Val Loss: 0.035.. Time: 38.49s
100%|██████████| 69/69 [00:28<00:00,  2.39it/s]
100%|██████████| 17/17 [00:09<00:00,  1.86it/s]
Epoch:59/60.. Train Acc: 0.852.. Val Acc: 0.853.. Train Loss: 0.029.. Val Loss: 0.031.. Time: 38.42s
100%|██████████| 69/69 [00:28<00:00,  2.39it/s]
100%|██████████| 17/17 [00:08<00:00,  1.90it/s]
Epoch:60/60.. Train Acc: 0.826.. Val Acc: 0.831.. Train Loss: 0.032.. Val Loss: 0.035.. Time: 38.25s
Total time: 38.95 m

2.6 打印准确率以及loss曲线

def plot_loss(x, history):
    plt.plot(x, history['val_loss'], label='val', marker='o')
    plt.plot(x, history['train_loss'], label='train', marker='o')
    plt.title('Loss per epoch')
    plt.ylabel('loss')
    plt.xlabel('epoch')
    plt.legend(), plt.grid()
    plt.show()


def plot_acc(x, history):
    plt.plot(x, history['train_acc'], label='train_acc', marker='x')
    plt.plot(x, history['val_acc'], label='val_acc', marker='x')
    plt.title('Score per epoch')
    plt.ylabel('score')
    plt.xlabel('epoch')
    plt.legend(), plt.grid()
    plt.show()

plot_loss(np.arange(0,epoch), history)
plot_acc(np.arange(0,epoch), history)
执行结果

loss曲线
image.png

准确率曲线
image.png

2.7 查看每一类的准确率

classes = ('n0', 'n1', 'n2', 'n3', 'n4', 'n5', 'n6', 'n7', 'n8', 'n9')

class_correct = [0.] * 10
class_total = [0.] * 10 
y_test, y_pred = [] , []
X_test = []

with torch.no_grad():
    for images, labels in val_loader:
        X_test.extend([_ for _ in images])
        outputs = model(images.to(device))
        _, predicted = torch.max(outputs, 1)
        predicted = predicted.cpu()
        c = (predicted == labels).squeeze()
        for i, label in enumerate(labels):
            class_correct[label] += c[i].item()
            class_total[label] += 1
        y_pred.extend(predicted.numpy())
        y_test.extend(labels.cpu().numpy())      
        
for i in range(10):
    print(f"Acuracy of {classes[i]:5s}: {100*class_correct[i]/class_total[i]:2.0f}%")
执行结果
Acuracy of n0   : 77%
Acuracy of n1   : 86%
Acuracy of n2   : 85%
Acuracy of n3   : 87%
Acuracy of n4   : 85%
Acuracy of n5   : 89%
Acuracy of n6   : 73%
Acuracy of n7   : 75%
Acuracy of n8   : 89%
Acuracy of n9   : 85%

2.8 查看precision,recall和f1-score

from sklearn.metrics import confusion_matrix, classification_report

ac = accuracy_score(y_test, y_pred)
cm = confusion_matrix(y_test, y_pred)
cr = classification_report(y_test, y_pred, target_names=classes)
print("Accuracy is :",ac) 
print(cr)
执行结果
Accuracy is : 0.8308823529411765
              precision    recall  f1-score   support

          n0       0.77      0.77      0.77        26
          n1       0.69      0.86      0.76        28
          n2       1.00      0.85      0.92        27
          n3       0.93      0.87      0.90        30
          n4       0.88      0.85      0.86        26
          n5       0.81      0.89      0.85        28
          n6       0.90      0.73      0.81        26
          n7       0.84      0.75      0.79        28
          n8       0.89      0.89      0.89        27
          n9       0.71      0.85      0.77        26

    accuracy                           0.83       272
   macro avg       0.84      0.83      0.83       272
weighted avg       0.84      0.83      0.83       272

2.9 查看混淆矩阵

import seaborn as sns, pandas as pd

labels = pd.DataFrame(cm).applymap(lambda v: f"{v}" if v!=0 else f"")
plt.figure(figsize=(7,5))
sns.heatmap(cm, annot=labels, fmt='s', xticklabels=classes, yticklabels=classes, linewidths=0.1 )
plt.show()

image.png

3.模型部署在Android

3.1 导出onnx模型

INPUT_DICT = './weight/best.pth'
OUT_ONNX = './weight/best.onnx'

x = torch.randn(1, 3, 224, 224)
input_names = ["input"]
out_names = ["output"]

model= torch.load(INPUT_DICT, map_location=torch.device('cpu'))
model.eval()

torch.onnx._export(model, x, OUT_ONNX, export_params=True, training=False, input_names=input_names, output_names=out_names)
print('please run: python -m onnxsim test.onnx test_sim.onnx\n')

3.2 将onnx模型简化

python -m onnxsim best.onnx best_sim.onnx

image.png

3.3 使用ncnn进行转化

首先转化为.param和.bin文件
onnx2ncnn.exe best_sim.onnx res.param res.bin

image.png

将.param和.bin文件加密
ncnn2mem.exe res.param res.bin res.id.h res.mem.h

image.png

3.4 最终效果

image.png

测试的时候发现,将图片稍微裁剪一下,猴子区域占整幅图像的比例大一点效果较好。

代码开源(仅供参考)

1.完整训练代码:https://github.com/yaoyi30/ResNet_Image_Classification_PyTorch

2.安卓代码:https://github.com/yaoyi30/ResNet_ncnn_android

3.我的CSDN:姚先生97的博客_CSDN博客

作者:YaoXiansheng
文章来源:知乎

推荐阅读

更多芯擎AI开发板干货请关注芯擎AI开发板专栏。欢迎添加极术小姐姐微信(id:aijishu20)加入技术交流群,请备注研究方向。

推荐阅读
关注数
9851
内容数
79
搭载基于安谋科技自研“周易”NPU的芯擎科技工业级“龍鹰一号”SE1000-I处理器
目录
极术微信服务号
关注极术微信号
实时接收点赞提醒和评论通知
安谋科技学堂公众号
关注安谋科技学堂
实时获取安谋科技及 Arm 教学资源
安谋科技招聘公众号
关注安谋科技招聘
实时获取安谋科技中国职位信息