pytorch 图片分类，python 图片分类，resnet18 图片分类
程序员文章站
2022-04-19 09:09:40
pytorh 图片分类，python 图片分类，resnet18 图片分类pytorh版本：1.5.0+cu101全部源码，可以直接运行。下载地址：未上传网络是用resnet18 ，可以修改图片的大小，默认是32 x32图片结构：训练代码：import torch as timport torchvision as tvimport osimport timeimport numpy as npfrom tqdm import tqdm# 一些参数配置.....
pytorch 图片分类，python 图片分类，resnet18 图片分类，深度学习 图片分类
pytorch版本：1.5.0+cu101
全部源码，可以直接运行。
下载地址：https://download.csdn.net/download/TangLingBo/12598435
网络是用 resnet18 ，可以修改图片的大小，默认是32 x32
如果出现需要下载的文件或者问题可以联系：QQ 1095788063
图片结构：
测试结果：
训练代码：
import torch as t
import torchvision as tv
import os
import time
import numpy as np
from tqdm import tqdm
# 一些参数配置
class DefaultConfigs(object):
data_dir = "./imageData/" # 图片目录
data_list = ["train", "test"] # train=训练集，test=测试集
lr = 0.001 # 学习率（默认值：1e-3
epochs = 51 # 训练次，越多就越好
num_classes = 10 # 分类
image_size = 32 # 图片大小 ，可以改，因为用的是 resnet18 的网络，越大就越慢
batch_size = 40 # 批量大小，看自己电脑的配置,需要占用 CPU或者GPU资源
channels = 3 # 通道数
use_gpu = t.cuda.is_available() # 启用gpu，如果电脑不支持，直接设置为 False ，GPU 训练效果最好
config = DefaultConfigs()
config.use_gpu = False # 我的电脑不支持，设置为 False
# 对Tensor进行变换 颜色转换 mean=给定均值：(R,G,B) std=方差：（R，G，B）
normalize = tv.transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
# Train数据需要进行随机裁剪，Test数据不要进行裁剪了
transform = {
# tv.transforms.Resize 用于重设图片大小 train 训练集数据
# tv.transforms.CenterCrop([224,224]) 将给定的PIL.Image进行中心切割
config.data_list[0]: tv.transforms.Compose(
[tv.transforms.Resize([config.image_size, config.image_size]),
tv.transforms.CenterCrop([config.image_size,
config.image_size]),
tv.transforms.ToTensor(), normalize]),
# test 测试数据
config.data_list[1]: tv.transforms.Compose([
tv.transforms.Resize([config.image_size, config.image_size]),
tv.transforms.ToTensor(),
normalize
])
}
# 数据集
datasets = {
x: tv.datasets.ImageFolder(root=os.path.join(config.data_dir, x), transform=transform[x])
for x in config.data_list
}
# 数据加载器
dataloader = {
x: t.utils.data.DataLoader(dataset=datasets[x],
batch_size=config.batch_size,
shuffle=True)
for x in config.data_list
}
# 构建网络模型 resnet18
def get_model(num_classes):
#resnet18 好像要下载什么的，忘记了，可以联系我
model = tv.models.resnet18(pretrained=True)
# 梯度什么的，电脑硬件支持，可以把下述代码屏蔽，则训练整个网络，最终准确率会上升，训练数据会变慢
# for parma in model.parameters():
# parma.requires_grad = False
model.fc = t.nn.Sequential(t.nn.Dropout(p=0.3), t.nn.Linear(512, num_classes))
return model
# 训练模型（支持自动GPU加速）
def train(epochs):
model = get_model(config.num_classes)
loss_f = t.nn.CrossEntropyLoss()
# GPU
if config.use_gpu:
model = model.cuda()
loss_f = loss_f.cuda()
opt = t.optim.Adam(model.fc.parameters(), lr=config.lr)
# 时间
time_start = time.time()
for epoch in range(epochs):
train_loss = []
train_acc = []
test_loss = []
test_acc = []
model.train(True) # 将模块设置为训练模式
print("Epoch {}/{}".format(epoch + 1, epochs))
for batch, datas in tqdm(enumerate(iter(dataloader["train"]))):
x, y = datas
# 开启GPU 加速
if config.use_gpu:
x, y = x.cuda(), y.cuda()
y_ = model(x)
# print(x.shape, y.shape, y_.shape)
_, pre_y_ = t.max(y_, 1)
pre_y = y
# print(y_.shape)
loss = loss_f(y_, pre_y)
# print(y_.shape)
acc = t.sum(pre_y_ == pre_y)
loss.backward()
opt.step()
opt.zero_grad()
if config.use_gpu:
loss = loss.cpu()
acc = acc.cpu()
train_loss.append(loss.data)
train_acc.append(acc)
time_end = time.time()
print("正式 批次 {}, Train 损失:{:.4f}, Train 准确率:{:.4f}, 训练时间: {}".format(batch + 1,
np.mean(train_loss) / config.batch_size,
np.mean(train_acc) / config.batch_size,
(time_end - time_start)))
model.train(False) # 关闭训练模式
for batch, datas in tqdm(enumerate(iter(dataloader["test"]))):
x, y = datas
if config.use_gpu:
x, y = x.cuda(), y.cuda()
y_ = model(x)
# print(x.shape,y.shape,y_.shape)
_, pre_y_ = t.max(y_, 1)
pre_y = y
# print(y_.shape)
loss = loss_f(y_, pre_y)
acc = t.sum(pre_y_ == pre_y)
if config.use_gpu:
loss = loss.cpu()
acc = acc.cpu()
test_loss.append(loss.data)
test_acc.append(acc)
print("测试 批次 {}, 损失:{:.4f}, 准确率:{:.4f}".format(batch + 1, np.mean(test_loss) / config.batch_size,
np.mean(test_acc) / config.batch_size))
t.save(model, 'model/' + str(epoch + 1) + "_ttmodel.pkl") # 保存整个神经网络的结构和模型参数
t.save(model.state_dict(), 'model/' + str(epoch + 1) + "_ttmodel_params.pkl") # 只保存神经网络的模型参数
print('训练结束')
#开始训练
if __name__ == "__main__":
train(config.epochs)
调用代码：
import torch as t
import torchvision as tv
from PIL import Image
import matplotlib.pyplot as plt
from torch.autograd import Variable
import numpy as np
bCuda = t.cuda.is_available() # 是否开启 GPU
bCuda = False # 不启用GPU 我的电脑不支持
device = t.device("cuda:0" if bCuda else "cpu")
img_size = 32 # 图片大小，可以改
# 对Tensor进行变换 颜色转换 mean=给定均值：(R,G,B) std=方差：（R，G，B）
normalize = tv.transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
transform = tv.transforms.Compose(
[tv.transforms.Resize([img_size, img_size]), tv.transforms.CenterCrop([img_size, img_size]),
tv.transforms.ToTensor(), normalize])
# 分类数组
classes = ['凹下标志-0', '凸上标志-1', '打滑标志-2', '左弯标志-3', '右弯标志-4', '连续转弯标志-5', '00020-6', '00021-7', '00022-8', '00023-9']
# 显示图片方法
def imshow(img):
plt.imshow(img)
plt.show()
# 单张图片调用
def prediect(model, img_path, imgType, isShowSoftmax=False, isShowImg=False):
t.no_grad()
image_PIL = Image.open(img_path)
# imshow(image_PIL)
image_tensor = transform(image_PIL)
# 以下语句等效于 img = torch.unsqueeze(image_tensor, 0)
image_tensor.unsqueeze_(0)
# 没有这句话会报错
image_tensor = image_tensor.to(device)
out = model(image_tensor)
# 得到预测结果，并且从大到小排序
_, indices = t.sort(out, descending=True)
# 返回每个预测值的百分数
percentage = t.nn.functional.softmax(out, dim=1)[0] * 100
# 是否显示每个分类的预测值
item = indices[0]
if isShowSoftmax:
for idx in item:
ss = percentage[idx]
value = ss.item();
name = classes[idx]
print('名称：', name, '预测值：', value)
# 预测最大值
_, predicted = t.max(out.data, 1)
maxPredicted = classes[predicted.item()]
maxAccuracy = percentage[item[0]].item()
if imgType == maxPredicted:
print('预测正确，预测结果：', maxPredicted, '预测值：', maxAccuracy)
else:
print('预测错误，正确结果：', imgType, '，预测结果：', maxPredicted, '预测值：', maxAccuracy, '图片：', img_path)
if isShowImg:
plt.imshow(image_PIL)
plt.show()
# 构建网络模型 resnet18
def get_model(num_classes):
# resnet18 好像要下载什么的，忘记了，可以联系我
model = tv.models.resnet18(pretrained=True)
# 梯度什么的，电脑硬件支持，可以把下述代码屏蔽，则训练整个网络，最终准确率会上升，训练数据会变慢
# for parma in model.parameters():
# parma.requires_grad = False
model.fc = t.nn.Sequential(t.nn.Dropout(p=0.3), t.nn.Linear(512, num_classes))
return model
# 测试集
def loadtestdata():
path = "./imageData/test/"
testset = tv.datasets.ImageFolder(path, transform=transform)
testloader = t.utils.data.DataLoader(testset, batch_size=40, shuffle=True, num_workers=6)
return testloader
# 测试全部
def testAll(model):
testloader = loadtestdata()
dataiter = iter(testloader)
images, labels = dataiter.next()
print(labels)
print('真实值: '
, " ".join('%5s' % classes[labels[j]] for j in range(25))) # 打印前25个GT（test集里图片的标签）
outputs = model(Variable(images))
_, predicted = t.max(outputs.data, 1)
print('预测值: ', " ".join('%5s' % classes[predicted[j]] for j in range(25)))
# 打印前25个预测值
imshow2(tv.utils.make_grid(images, nrow=5)) # nrow是每行显示的图片数量，缺省值为8
def imshow2(img):
img = img / 2 + 0.5 # unnormalize
npimg = img.numpy()
plt.imshow(np.transpose(npimg, (1, 2, 0)))
plt.show()
if __name__ == '__main__':
# 直接加载
model = t.load('model/51_ttmodel.pkl')
# 加载2 ，看官方的解释
# model = get_model(classes.__len__()) # 10 分类数量
# load_weights = t.load('model/51_ttmodel_params.pkl', map_location='cpu')
# model.load_state_dict(load_weights)
model = model.to(device) # GPU
model.eval() # 运行模式
# 测试全部图片
testAll(model)
# 测试一张图片
# # 凹下标志-0
# prediect(model,'imageData/test/00000/01160_00000.png', classes[0], False, False)
# prediect(model,'imageData/test/00000/01160_00001.png', classes[0], False, False)
# prediect(model,'imageData/test/00000/01160_00002.png', classes[0], False, False)
# prediect(model,'imageData/test/00000/01798_00000.png', classes[0], False, False)
# prediect(model,'imageData/test/00000/01798_00001.png', classes[0], False, False)
# prediect(model,'imageData/test/00000/01798_00002.png', classes[0], False, False)
#
# # 凸上标志-1
# prediect(model,'imageData/test/00001/00029_00000.png', classes[1], False, False)
# prediect(model,'imageData/test/00001/00029_00001.png', classes[1], False, False)
# prediect(model,'imageData/test/00001/00029_00002.png', classes[1], False, False)
# prediect(model,'imageData/test/00001/00079_00000.png', classes[1], False, False)
# prediect(model,'imageData/test/00001/00079_00002.png', classes[1], False, False)
# prediect(model,'imageData/test/00001/00079_00001.png', classes[1], False, False)
#
# # 打滑标志-2
# prediect(model,'imageData/test/00002/01503_00000.png', classes[2], False, False)
# prediect(model,'imageData/test/00002/01503_00001.png', classes[2], False, False)
# prediect(model,'imageData/test/00002/01503_00002.png', classes[2], False, False)
# prediect(model,'imageData/test/00002/01515_00000.png', classes[2], False, False)
# prediect(model,'imageData/test/00002/01515_00001.png', classes[2], False, False)
# prediect(model,'imageData/test/00002/01515_00002.png', classes[2], False, False)
#
# # 左弯标志-3
# prediect(model,'imageData/test/00003/00207_00000.png', classes[3], False, False)
# prediect(model,'imageData/test/00003/00207_00001.png', classes[3], False, False)
# prediect(model,'imageData/test/00003/00207_00002.png', classes[3], False, False)
# prediect(model,'imageData/test/00003/00211_00000.png', classes[3], False, False)
# prediect(model,'imageData/test/00003/00211_00001.png', classes[3], False, False)
# prediect(model,'imageData/test/00003/00211_00002.png', classes[3], False, False)
# prediect(model,'imageData/test/00003/02664_00000.png', classes[3], False, False)
# prediect(model,'imageData/test/00003/02664_00001.png', classes[3], False, False)
# prediect(model,'imageData/test/00003/02664_00002.png', classes[3], False, False)
#
# # 右弯标志-4
# prediect(model,'imageData/test/00004/00214_00000.png', classes[4], False, False)
# prediect(model,'imageData/test/00004/00214_00001.png', classes[4], False, False)
# prediect(model,'imageData/test/00004/00214_00002.png', classes[4], False, False)
# prediect(model,'imageData/test/00004/00282_00000.png', classes[4], False, False)
# prediect(model,'imageData/test/00004/00282_00001.png', classes[4], False, False)
# prediect(model,'imageData/test/00004/00282_00002.png', classes[4], False, False)
# prediect(model,'imageData/test/00004/02567_00000.png', classes[4], False, False)
# prediect(model,'imageData/test/00004/02567_00001.png', classes[4], False, False)
# prediect(model,'imageData/test/00004/02567_00002.png', classes[4], False, False)
# prediect(model,'imageData/test/00004/02660_00000.png', classes[4], False, False)
# prediect(model,'imageData/test/00004/02660_00001.png', classes[4], False, False)
# prediect(model,'imageData/test/00004/02660_00002.png', classes[4], False, False)
#
# # 连续转弯标志-5
# prediect(model,'imageData/test/00005/00575_00000.png', classes[5], False, False)
# prediect(model,'imageData/test/00005/00575_00001.png', classes[5], False, False)
# prediect(model,'imageData/test/00005/00575_00002.png', classes[5], False, False)
# prediect(model,'imageData/test/00005/01893_00000.png', classes[5], False, False)
# prediect(model,'imageData/test/00005/01893_00001.png', classes[5], False, False)
# prediect(model,'imageData/test/00005/01893_00002.png', classes[5], False, False)
# prediect(model,'imageData/test/00005/02225_00000.png', classes[5], False, False)
# prediect(model,'imageData/test/00005/02225_00001.png', classes[5], False, False)
# prediect(model,'imageData/test/00005/02225_00002.png', classes[5], False, False)
#
#
# # 00020-6
# prediect(model,'imageData/test/00020/00230_00000.png', classes[6], False, False)
# prediect(model, 'imageData/test/00020/00230_00001.png', classes[6], True, True)
# prediect(model,'imageData/test/00020/00230_00002.png', classes[6], False, False)
# prediect(model,'imageData/test/00020/00231_00000.png', classes[6], False, False)
# prediect(model,'imageData/test/00020/00231_00001.png', classes[6], False, False)
# prediect(model,'imageData/test/00020/00231_00002.png', classes[6], False, False)
#
# # 00021-7
# prediect(model, 'imageData/test/00021/00375_00000.png', classes[7], False, False)
# prediect(model, 'imageData/test/00021/00375_00001.png', classes[7], False, False)
# prediect(model, 'imageData/test/00021/00375_00002.png', classes[7], False, False)
# prediect(model, 'imageData/test/00021/00478_00000.png', classes[7], False, False)
# prediect(model, 'imageData/test/00021/00478_00001.png', classes[7], False, False)
# prediect(model, 'imageData/test/00021/00478_00002.png', classes[7], False, False)
#
# # 00022-8
# prediect(model, 'imageData/test/00022/00020_00000.png', classes[8], False, False)
# prediect(model, 'imageData/test/00022/00020_00001.png', classes[8], False, False)
# prediect(model, 'imageData/test/00022/00020_00002.png', classes[8], False, False)
# prediect(model, 'imageData/test/00022/00048_00000.png', classes[8], False, False)
# prediect(model, 'imageData/test/00022/00048_00001.png', classes[8], False, False)
# prediect(model, 'imageData/test/00022/00048_00002.png', classes[8], False, False)
#
# # 00023-9
# prediect(model, 'imageData/test/00023/00465_00000.png', classes[9], False, False)
# prediect(model, 'imageData/test/00023/00465_00001.png', classes[9], False, False)
# prediect(model, 'imageData/test/00023/00465_00002.png', classes[9], False, False)
# prediect(model, 'imageData/test/00023/00535_00000.png', classes[9], False, False)
# prediect(model, 'imageData/test/00023/00535_00001.png', classes[9], False, False)
# prediect(model, 'imageData/test/00023/00535_00002.png', classes[9], False, False)
本文地址：https://blog.csdn.net/TangLingBo/article/details/107258003
推荐阅读
-
PHP 将图片按创建时间进行分类存储的实现代码
-
python数据处理 根据颜色对图片进行分类的方法
-
使用Scrapy自带的ImagesPipeline下载图片，并对其进行分类。
-
pytorch 图片分类，python 图片分类，resnet18 图片分类
-
如何从 900 万张图片中对 600 类照片进行分类？ | 技术头条
-
图片分类-K近邻分类器
-
pytorch + python实现逻辑回归（二分类）
-
蚂蚁分类信息商家发布文章、商品外链及远程图片自动添加nofollow属性
-
Hexo+GitHub基本操作记录 图片无法显示 添加分类
-
基于Tensorflow针对cifar数据集运用卷积神经网络解决100类图片的分类问题。