一、选题的背景
为了实现对水果和蔬菜的分类识别,收集了香蕉、苹果、梨、葡萄、橙子、猕猴桃、西瓜、石榴、菠萝、芒果、黄瓜、胡萝卜、辣椒、洋葱、马铃薯、柠檬、番茄、萝卜、甜菜根、卷心菜、生菜、菠菜、大豆、花椰菜、甜椒、辣椒、萝卜、玉米、甜玉米、红薯、辣椒粉、生姜、大蒜、豌豆、茄子共36种果蔬的图像。该项目使用resnet18网络进行分类。
二、机器学习案例设计方案
1.本选题采用的机器学习案例(训练集与测试集)的来源描述
数据集来自百度AI studio平台(https://aistudio.baidu.com/aistudio/datasetdetail/119023/0),共包含36种果蔬,每一个类别包括100张训练图像,10张测试图像和10张验证图像。
2 采用的机器学习框架描述
本次使用的网络框架,主要用到了二维卷积、激活函数、最大池化、Dropout和全连接,下面将对搭建的网络模型进行解释。
首先是一个二维卷积层,输入通道数为3,输出通道数为100,卷积核大小是3*3,填充大小是1*1。输入通道数为3是因为这个是第一层卷积,输入的是RGB图像,具有三个通道,输出通道数量可以根据实际情况自定。填充是因为希望在卷积后,不要改变图像的尺寸。
在卷积层之后是一个RELU激活函数,如果不用激活函数,在这种情况下每一层输出都是上层输入的线性函数。容易验证,无论神经网络有多少层,输出都是输入的线性组合,与没有隐藏层效果相当。因此引入非线性函数作为激活函数,这样深层神经网络就有意义了(不再是输入的线性组合,可以逼近任意函数)。最早的想法是sigmoid函数或者tanh函数,输出有界,很容易充当下一层输入。
引入RELU激活函数有以下三个原因:
第一,采用sigmoid等函数,算激活函数时(指数运算),计算量大,反向传播求误差梯度时,求导涉及除法,计算量相对大,而采用Relu激活函数,整个过程的计算量节省很多。
第二,对于深层网络,sigmoid函数反向传播时,很容易就会出现 梯度消失 的情况(在sigmoid接近饱和区时,变换太缓慢,导数趋于0,这种情况会造成信息丢失),从而无法完成深层网络的训练。
第三,ReLu会使一部分神经元的输出为0,这样就造成了 网络的稀疏性,并且减少了参数的相互依存关系,缓解了过拟合问题的发生。
然后再跟一个二维卷积层,输入通道数应该和上一层卷积的输出通道数相同,所以设为100, 输出通道数同样根据实际情况设定,此处设为150,其他参数与第一层卷积相同。
后续每一个卷积层和全连接层后面都会跟一个RELU激活函数,所以后面不再叙述RELU激活函数层。
再之后添加一个2*2的最大池化层,该层用来缩减模型的大小,提高计算速度,同时提高所提取特征的鲁棒性。
再经过三次卷积后,使用Flatten将二维Tensor拉平,变为一维Tensor,然后使用全连接层,通过多个全连接层后,使用dropout层随机删除一些结点,该方法可以有效的避免网络过拟合,在最后一个全连接层的输出对应需要分类的个数。
3.涉及到的技术难点与解决思路
下载的数据集没有划分训练集、测试集和验证集,需要自己写代码完成划分。在刚开始写代码的时候对于文件路径没有搞清楚,没有弄懂os.path.join方法如何使用,导致总是读取不到图像,并且代码还没有报错误正常运行结束,但是查看划分后的文件夹里没有数据。通过debug发现文件的路径出现问题,具体是windows下的/和\\混用,导致不能正确的对路径进行处理。在排除问题后统一使用\\\\,最终问题得到解决。
三、机器学习的实现步骤
(1)划分数据集并进行缩放
1 import os 2 import glob 3 import random 4 import shutil 5 from PIL import Image 6 #对所有图片进行RGB转化,并且统一调整到一致大小,但不让图片发生变形或扭曲,划分了训练集和测试集 7 8 if __name__ == \'__main__\': 9 test_split_ratio = 0.05 #百分之五的比例作为测试集 10 desired_size = 128 # 图片缩放后的统一大小 11 raw_path = \'./raw\' 12 13 #把多少个类别算出来,包括目录也包括文件 14 dirs = glob.glob(os.path.join(raw_path, \'*\')) 15 #进行过滤,只保留目录,一共36个类别 16 dirs = [d for d in dirs if os.path.isdir(d)] 17 18 print(f\'Totally {len(dirs)} classes: {dirs}\') 19 20 for path in dirs: 21 # 对每个类别单独处理 22 23 #只保留类别名称 24 path = path.split(\'/\')[-1] 25 print(path) 26 #创建文件夹 27 os.makedirs(f\'train/{path}\', exist_ok=True) 28 os.makedirs(f\'test/{path}\', exist_ok=True) 29 30 #原始文件夹当前类别的图片进行匹配 31 files = glob.glob(os.path.join( path, \'*.jpg\')) 32 # print(raw_path, path) 33 34 files += glob.glob(os.path.join( path, \'*.JPG\')) 35 files += glob.glob(os.path.join( path, \'*.png\')) 36 37 random.shuffle(files)#原地shuffle,因为要取出来验证集 38 39 boundary = int(len(files)*test_split_ratio) # 训练集和测试集的边界 40 41 for i, file in enumerate(files): 42 img = Image.open(file).convert(\'RGB\') 43 44 old_size = img.size 45 46 ratio = float(desired_size)/max(old_size) 47 48 new_size = tuple([int(x*ratio) for x in old_size])#等比例缩放 49 50 im = img.resize(new_size, Image.ANTIALIAS)#后面的方法不会造成模糊 51 52 new_im = Image.new(\"RGB\", (desired_size, desired_size)) 53 54 #new_im在某个尺寸上更大,我们将旧图片贴到上面 55 new_im.paste(im, ((desired_size-new_size[0])//2, 56 (desired_size-new_size[1])//2)) 57 58 assert new_im.mode == \'RGB\' 59 60 if i <= boundary: 61 new_im.save(os.path.join(f\'test/{path}\', file.split(\'\\\\\')[-1].split(\'.\')[0]+\'.jpg\')) 62 else: 63 new_im.save(os.path.join(f\'train/{path}\', file.split(\'\\\\\')[-1].split(\'.\')[0]+\'.jpg\')) 64 65 test_files = glob.glob(os.path.join(\'test\', \'*\', \'*.jpg\')) 66 train_files = glob.glob(os.path.join(\'train\', \'*\', \'*.jpg\')) 67 68 print(f\'Totally {len(train_files)} files for training\') 69 print(f\'Totally {len(test_files)} files for test\')
(2)图像预处理
包括随即旋转、随机翻转、裁剪等,并进行归一化。
1 #图像预处理 2 train_dir = \'./train\' 3 val_dir = \'./test\' 4 test_dir = \'./test\' 5 classes0 = os.listdir(train_dir) 6 classes=sorted(classes0) 7 print(classes) 8 train_transform=transforms.Compose([ 9 transforms.RandomRotation(10), # 旋转+/-10度 10 transforms.RandomHorizontalFlip(), # 反转50%的图像 11 transforms.Resize(40), # 调整最短边的大小 12 transforms.CenterCrop(40), # 作物最长边 13 transforms.ToTensor(), 14 transforms.Normalize([0.485, 0.456, 0.406], 15 [0.229, 0.224, 0.225]) 16 ])
1 #显示图像 2 def show_image(img,label): 3 print(\'Label: \', trainset.classes[label], \"(\"+str(label)+\")\") 4 plt.imshow(img.permute(1,2,0)) 5 plt.show() 6 7 show_image(*trainset[10]) 8 show_image(*trainset[20])
(3)读取数据
1 batch_size = 64 2 train_loader = DataLoader(train_ds, batch_size, shuffle=True, num_workers=4, pin_memory=True) 3 val_loader = DataLoader(val_ds, batch_size*2, num_workers=4, pin_memory=True) 4 test_loader = DataLoader(test_ds, batch_size*2, num_workers=4, pin_memory=True)
(4)构建CNN模型
#构建CNN模型
1 #构建CNN模型 2 class CnnModel(ImageClassificationBase): 3 def __init__(self): 4 super().__init__() 5 #cnn提取特征 6 self.network = nn.Sequential( 7 nn.Conv2d(3, 100, kernel_size=3, padding=1),#Conv2D层 8 nn.ReLU(), 9 nn.Conv2d(100, 150, kernel_size=3, stride=1, padding=1), 10 nn.ReLU(), 11 nn.MaxPool2d(2, 2), #池化层 12 13 nn.Conv2d(150, 200, kernel_size=3, stride=1, padding=1), 14 nn.ReLU(), 15 nn.Conv2d(200, 200, kernel_size=3, stride=1, padding=1), 16 nn.ReLU(), 17 nn.MaxPool2d(2, 2), 18 19 nn.Conv2d(200, 250, kernel_size=3, stride=1, padding=1), 20 nn.ReLU(), 21 nn.Conv2d(250, 250, kernel_size=3, stride=1, padding=1), 22 nn.ReLU(), 23 nn.MaxPool2d(2, 2), 24 25 #全连接 26 nn.Flatten(), 27 nn.Linear(6250, 256), 28 nn.ReLU(), 29 nn.Linear(256, 128), 30 nn.ReLU(), 31 nn.Linear(128, 64), 32 nn.ReLU(), 33 nn.Linear(64, 32), 34 nn.ReLU(), 35 nn.Dropout(0.25), 36 nn.Linear(32, len(classes))) 37 38 def forward(self, xb): 39 return self.network(xb)
(5)训练网络
#训练网络
1 #训练网络 2 @torch.no_grad() 3 def evaluate(model, val_loader): 4 model.eval() 5 outputs = [model.validation_step(batch) for batch in val_loader] 6 return model.validation_epoch_end(outputs) 7 8 def fit(epochs, lr, model, train_loader, val_loader, opt_func=torch.optim.SGD): 9 history = [] 10 optimizer = opt_func(model.parameters(), lr) 11 for epoch in range(epochs): 12 # 训练阶段 13 model.train() 14 train_losses = [] 15 for batch in tqdm(train_loader,disable=True): 16 loss = model.training_step(batch) 17 train_losses.append(loss) 18 loss.backward() 19 optimizer.step() 20 optimizer.zero_grad() 21 # 验证阶段 22 result = evaluate(model, val_loader) 23 result[\'train_loss\'] = torch.stack(train_losses).mean().item() 24 model.epoch_end(epoch, result) 25 history.append(result) 26 return history 27 28 model = to_device(CnnModel(), device) 29 30 history=[evaluate(model, val_loader)] 31 32 num_epochs = 100 33 opt_func = torch.optim.Adam 34 lr = 0.001 35 36 history+= fit(num_epochs, lr, model, train_dl, val_dl, opt_func)
(6)绘制损失函数和准确率图
1 def plot_accuracies(history): 2 accuracies = [x[\'val_acc\'] for x in history] 3 plt.plot(accuracies, \'-x\') 4 plt.xlabel(\'epoch\') 5 plt.ylabel(\'accuracy\') 6 plt.title(\'Accuracy vs. No. of epochs\') 7 plt.show() 8 9 def plot_losses(history): 10 train_losses = [x.get(\'train_loss\') for x in history] 11 val_losses = [x[\'val_loss\'] for x in history] 12 plt.plot(train_losses, \'-bx\') 13 plt.plot(val_losses, \'-rx\') 14 plt.xlabel(\'epoch\') 15 plt.ylabel(\'loss\') 16 plt.legend([\'Training\', \'Validation\']) 17 plt.title(\'Loss vs. No. of epochs\') 18 plt.show() 19 20 plot_accuracies(history) 21 plot_losses(history) 22 23 evaluate(model, test_loader)
(7)预测
1 #预测分类 2 y_true=[] 3 y_pred=[] 4 with torch.no_grad(): 5 for test_data in test_loader: 6 test_images, test_labels = test_data[0].to(device), test_data[1].to(device) 7 pred = model(test_images).argmax(dim=1) 8 for i in range(len(pred)): 9 y_true.append(test_labels[i].item()) 10 y_pred.append(pred[i].item()) 11 12 from sklearn.metrics import classification_report 13 print(classification_report(y_true,y_pred,target_names=classes,digits=4))
(8)读取图片测试
1 import numpy as np 2 from PIL import Image 3 import matplotlib.pyplot as plt 4 import torchvision.transforms as transforms 5 6 def predict(img_path): 7 img = Image.open(img_path) 8 plt.imshow(img) 9 plt.show() 10 img = img.resize((32,32)) 11 img = transforms.ToTensor()(img) 12 img = img.unsqueeze(0) 13 img = img.to(device) 14 pred = model(img).argmax(dim=1) 15 print(\'预测结果为:\',classes[pred.item()]) 16 return classes[pred.item()] 17 18 predict(\'./raw/apple/Image_1.jpg\')
四、总结
在本次课程设计中,使用深度学习的方法实现了果蔬的36分类,相对来说分类数量是比较多的,在训练了100个epoch以后,分类的准确率可以达到74.3%。通过对果蔬的分类,我明白了当训练集的图像数量较少时,可以采用数据增强对原始图像进行处理,获得更多的数据来增强网络的泛化能力,避免网络过拟合。数据增强的方法一般有随机翻转、随即旋转、随即裁剪、明暗变化、高斯噪声、椒盐噪声等。除此之外,对整个深度学习中图像分类的流程也有了一定的了解,从收集数据、对数据进行预处理、自己构建网络模型、训练网络到最后的预测结果,加深了对图像分类过程的理解。希望在以后的学习中,可以学习更多深度学习的方法和应用。
五、全部代码
1 import os 2 import glob 3 import random 4 import shutil 5 from PIL import Image 6 #对所有图片进行RGB转化,并且统一调整到一致大小,但不让图片发生变形或扭曲,划分了训练集和测试集 7 8 if __name__ == \'__main__\': 9 test_split_ratio = 0.05 #百分之五的比例作为测试集 10 desired_size = 128 # 图片缩放后的统一大小 11 raw_path = \'./raw\' 12 13 #把多少个类别算出来,包括目录也包括文件 14 dirs = glob.glob(os.path.join(raw_path, \'*\')) 15 #进行过滤,只保留目录,一共36个类别 16 dirs = [d for d in dirs if os.path.isdir(d)] 17 18 print(f\'Totally {len(dirs)} classes: {dirs}\') 19 20 for path in dirs: 21 # 对每个类别单独处理 22 23 #只保留类别名称 24 path = path.split(\'/\')[-1] 25 print(path) 26 #创建文件夹 27 os.makedirs(f\'train/{path}\', exist_ok=True) 28 os.makedirs(f\'test/{path}\', exist_ok=True) 29 30 #原始文件夹当前类别的图片进行匹配 31 files = glob.glob(os.path.join(raw_path, path, \'*.jpg\')) 32 # print(raw_path, path) 33 34 files += glob.glob(os.path.join(raw_path, path, \'*.JPG\')) 35 files += glob.glob(os.path.join(raw_path, path, \'*.png\')) 36 37 random.shuffle(files)#原地shuffle,因为要取出来验证集 38 39 boundary = int(len(files)*test_split_ratio) # 训练集和测试集的边界 40 41 for i, file in enumerate(files): 42 img = Image.open(file).convert(\'RGB\') 43 44 old_size = img.size 45 46 ratio = float(desired_size)/max(old_size) 47 48 new_size = tuple([int(x*ratio) for x in old_size])#等比例缩放 49 50 im = img.resize(new_size, Image.ANTIALIAS)#后面的方法不会造成模糊 51 52 new_im = Image.new(\"RGB\", (desired_size, desired_size)) 53 54 #new_im在某个尺寸上更大,我们将旧图片贴到上面 55 new_im.paste(im, ((desired_size-new_size[0])//2, 56 (desired_size-new_size[1])//2)) 57 58 assert new_im.mode == \'RGB\' 59 60 if i <= boundary: 61 new_im.save(os.path.join(f\'test/{path}\', file.split(\'/\')[-1].split(\'.\')[0]+\'.jpg\')) 62 else: 63 new_im.save(os.path.join(f\'train/{path}\', file.split(\'/\')[-1].split(\'.\')[0]+\'.jpg\')) 64 65 test_files = glob.glob(os.path.join(\'test\', \'*\', \'*.jpg\')) 66 train_files = glob.glob(os.path.join(\'train\', \'*\', \'*.jpg\')) 67 68 69 print(f\'Totally {len(train_files)} files for training\') 70 print(f\'Totally {len(test_files)} files for test\') 71 72 73 import os 74 import random 75 import numpy as np 76 import pandas as pd 77 import torch 78 import torch.nn as nn 79 import torch.nn.functional as F 80 from tqdm.notebook import tqdm 81 from torchvision import datasets, transforms, models 82 from torchvision.datasets import ImageFolder 83 from torchvision.transforms import ToTensor 84 from torchvision.utils import make_grid 85 from torch.utils.data import random_split 86 from torch.utils.data.dataloader import DataLoader 87 import matplotlib.pyplot as plt 88 89 if __name__ == \'__main__\': 90 # 使用第2个GPU 91 os.environ[\"CUDA_VISIBLE_DEVICES\"] = \"1\" 92 93 #图像预处理 94 train_dir = \'./train\' 95 val_dir = \'./test\' 96 test_dir = \'./test\' 97 classes0 = os.listdir(train_dir) 98 classes=sorted(classes0) 99 # print(classes) 100 train_transform=transforms.Compose([ 101 transforms.RandomRotation(10), # 旋转+/-10度 102 transforms.RandomHorizontalFlip(), # 反转50%的图像 103 transforms.Resize(40), # 调整最短边的大小 104 transforms.CenterCrop(40), # 作物最长边 105 transforms.ToTensor(), 106 transforms.Normalize([0.485, 0.456, 0.406], 107 [0.229, 0.224, 0.225]) 108 ]) 109 110 trainset = ImageFolder(train_dir, transform=train_transform) 111 valset = ImageFolder(val_dir, transform=train_transform) 112 testset = ImageFolder(test_dir, transform=train_transform) 113 # print(len(trainset)) 114 115 #查看数据集的一个图像形状 116 img, label = trainset[10] 117 # print(img.shape) 118 119 #显示图像 120 def show_image(img,label): 121 print(\'Label: \', trainset.classes[label], \"(\"+str(label)+\")\") 122 plt.imshow(img.permute(1,2,0)) 123 plt.show() 124 125 # show_image(*trainset[10]) 126 # show_image(*trainset[20]) 127 128 torch.manual_seed(10) 129 train_size = len(trainset) 130 val_size = len(valset) 131 test_size = len(testset) 132 133 train_ds=trainset 134 val_ds=valset 135 test_ds=testset 136 len(train_ds), len(val_ds), len(test_ds) 137 138 #读取数据 139 batch_size = 64 140 train_loader = DataLoader(train_ds, batch_size, shuffle=True, num_workers=4, pin_memory=True) 141 val_loader = DataLoader(val_ds, batch_size*2, num_workers=4, pin_memory=True) 142 test_loader = DataLoader(test_ds, batch_size*2, num_workers=4, pin_memory=True) 143 144 145 if __name__ == \'__main__\': 146 for images, labels in train_loader: 147 fig, ax = plt.subplots(figsize=(18,10)) 148 ax.set_xticks([]) 149 ax.set_yticks([]) 150 ax.imshow(make_grid(images,nrow=16).permute(1,2,0)) 151 break 152 153 154 155 torch.cuda.is_available() 156 157 158 #选择GPU或CPU 159 def get_default_device(): 160 if torch.cuda.is_available(): 161 return torch.device(\'cuda\') 162 else: 163 return torch.device(\'cpu\') 164 165 #移动到所选的设备 166 def to_device(data, device): 167 if isinstance(data, (list,tuple)): 168 return [to_device(x, device) for x in data] 169 return data.to(device, non_blocking=True) 170 171 class DeviceDataLoader(): 172 #包装数据加载器以将数据移动到设备 173 def __init__(self, dl, device): 174 self.dl = dl 175 self.device = device 176 177 def __iter__(self): 178 #将数据移动到设备后生成一批数据 179 for b in self.dl: 180 yield to_device(b, self.device) 181 182 def __len__(self): 183 #分批次 184 return len(self.dl) 185 186 device = get_default_device() 187 188 189 train_loader = DeviceDataLoader(train_loader, device) 190 val_loader = DeviceDataLoader(val_loader, device) 191 test_loader = DeviceDataLoader(test_loader, device) 192 193 input_size = 3*40*40 194 output_size = 3 195 196 197 198 def accuracy(outputs, labels): 199 _, preds = torch.max(outputs, dim=1) 200 return torch.tensor(torch.sum(preds == labels).item() / len(preds)) 201 202 #图像分类 203 class ImageClassificationBase(nn.Module): 204 def training_step(self, batch): 205 images, labels = batch 206 out = self(images) # 生成预测 207 loss = F.cross_entropy(out, labels) # 计算损失 208 return loss 209 210 def validation_step(self, batch): 211 images, labels = batch 212 out = self(images) # 生成预测 213 loss = F.cross_entropy(out, labels) # 计算损失 214 acc = accuracy(out, labels) # 计算精度 215 return {\'val_loss\': loss.detach(), \'val_acc\': acc} 216 217 def validation_epoch_end(self, outputs): 218 batch_losses = [x[\'val_loss\'] for x in outputs] 219 epoch_loss = torch.stack(batch_losses).mean() # 合并损失 220 batch_accs = [x[\'val_acc\'] for x in outputs] 221 epoch_acc = torch.stack(batch_accs).mean() # 结合精度 222 return {\'val_loss\': epoch_loss.item(), \'val_acc\': epoch_acc.item()} 223 224 def epoch_end(self, epoch, result): 225 print(\"Epoch [{}], train_loss: {:.4f}, val_loss: {:.4f}, val_acc: {:.4f}\".format( 226 epoch, result[\'train_loss\'], result[\'val_loss\'], result[\'val_acc\'])) 227 228 #构建CNN模型 229 class CnnModel(ImageClassificationBase): 230 def __init__(self): 231 super().__init__() 232 #cnn提取特征 233 self.network = nn.Sequential( 234 nn.Conv2d(3, 100, kernel_size=3, padding=1),#Conv2D层 235 nn.ReLU(), 236 nn.Conv2d(100, 150, kernel_size=3, stride=1, padding=1), 237 nn.ReLU(), 238 nn.MaxPool2d(2, 2), #池化层 239 240 nn.Conv2d(150, 200, kernel_size=3, stride=1, padding=1), 241 nn.ReLU(), 242 nn.Conv2d(200, 200, kernel_size=3, stride=1, padding=1), 243 nn.ReLU(), 244 nn.MaxPool2d(2, 2), 245 246 nn.Conv2d(200, 250, kernel_size=3, stride=1, padding=1), 247 nn.ReLU(), 248 nn.Conv2d(250, 250, kernel_size=3, stride=1, padding=1), 249 nn.ReLU(), 250 nn.MaxPool2d(2, 2), 251 252 #全连接 253 nn.Flatten(), 254 nn.Linear(6250, 256), 255 nn.ReLU(), 256 nn.Linear(256, 128), 257 nn.ReLU(), 258 nn.Linear(128, 64), 259 nn.ReLU(), 260 nn.Linear(64, 32), 261 nn.ReLU(), 262 nn.Dropout(0.25), 263 nn.Linear(32, len(classes))) 264 265 def forward(self, xb): 266 return self.network(xb) 267 268 # 将模型加载到GPU上去 269 model = CnnModel() 270 271 # model.cuda() 272 273 if __name__ == \'__main__\': 274 for images, labels in train_loader: 275 out = model(images) 276 print(\'images.shape:\', images.shape) 277 print(\'out.shape:\', out.shape) 278 print(\'out[0]:\', out[0]) 279 break 280 281 device = get_default_device() 282 283 train_dl = DeviceDataLoader(train_loader, device) 284 val_dl = DeviceDataLoader(val_loader, device) 285 test_dl = DeviceDataLoader(test_loader, device) 286 to_device(model, device) 287 288 289 #训练网络 290 def evaluate(model, val_loader): 291 model.eval() 292 outputs = [model.validation_step(batch) for batch in val_loader] 293 return model.validation_epoch_end(outputs) 294 295 def fit(epochs, lr, model, train_loader, val_loader, opt_func=torch.optim.SGD): 296 history = [] 297 optimizer = opt_func(model.parameters(), lr) 298 for epoch in range(epochs): 299 # 训练阶段 300 model.train() 301 train_losses = [] 302 for batch in tqdm(train_loader,disable=True): 303 loss = model.training_step(batch) 304 train_losses.append(loss) 305 loss.backward() 306 optimizer.step() 307 optimizer.zero_grad() 308 # 验证阶段 309 result = evaluate(model, val_loader) 310 result[\'train_loss\'] = torch.stack(train_losses).mean().item() 311 model.epoch_end(epoch, result) 312 history.append(result) 313 return history 314 315 model = to_device(CnnModel(), device) 316 317 318 history=[evaluate(model, val_loader)] 319 num_epochs = 5 320 opt_func = torch.optim.Adam 321 lr = 0.001 322 323 history+= fit(num_epochs, lr, model, train_dl, val_dl, opt_func) 324 325 326 # # 绘制损失函数和准确率图 327 328 def plot_accuracies(history): 329 accuracies = [x[\'val_acc\'] for x in history] 330 plt.plot(accuracies, \'-x\') 331 plt.xlabel(\'epoch\') 332 plt.ylabel(\'accuracy\') 333 plt.title(\'Accuracy vs. No. of epochs\') 334 plt.show() 335 336 def plot_losses(history): 337 train_losses = [x.get(\'train_loss\') for x in history] 338 val_losses = [x[\'val_loss\'] for x in history] 339 plt.plot(train_losses, \'-bx\') 340 plt.plot(val_losses, \'-rx\') 341 plt.xlabel(\'epoch\') 342 plt.ylabel(\'loss\') 343 plt.legend([\'Training\', \'Validation\']) 344 plt.title(\'Loss vs. No. of epochs\') 345 plt.show() 346 347 plot_accuracies(history) 348 plot_losses(history) 349 350 evaluate(model, test_loader) 351 352 353 #预测分类 354 y_true=[] 355 y_pred=[] 356 with torch.no_grad(): 357 for test_data in test_loader: 358 test_images, test_labels = test_data[0].to(device), test_data[1].to(device) 359 pred = model(test_images).argmax(dim=1) 360 for i in range(len(pred)): 361 y_true.append(test_labels[i].item()) 362 y_pred.append(pred[i].item()) 363 364 from sklearn.metrics import classification_report 365 print(classification_report(y_true,y_pred,target_names=classes,digits=4)) 366 367 # 读取图片进行预测 368 import numpy as np 369 from PIL import Image 370 import matplotlib.pyplot as plt 371 import torchvision.transforms as transforms 372 373 def predict(img_path): 374 img = Image.open(img_path) 375 plt.imshow(img) 376 plt.show() 377 img = img.resize((32,32)) 378 img = transforms.ToTensor()(img) 379 img = img.unsqueeze(0) 380 img = img.to(device) 381 pred = model(img).argmax(dim=1) 382 print(\'预测结果为:\',classes[pred.item()]) 383 return classes[pred.item()] 384 385 predict(\'./raw/apple/Image_1.jpg\')
来源:https://www.cnblogs.com/jihua056/p/17000109.html
本站部分图文来源于网络,如有侵权请联系删除。