引入
图像分类数据集最常用的是手写数字识别数据集MNIST (1)，但是大部分模型在其上的分类精度都超过了95%。为了更直观地观察算法之间的差异，将使用一个图像内容更加复杂的数据集[Fashion-MNIST (2)]。
接下来的部分将使用torchvision包，主要用于构建计算机视觉模型，主要由以下4部分组成：
| 组成 | 功能 |
|---|---|
| torchvision.datasets | 加载数据的函数及常用的数据集接口 |
| torchvision.models | 包含常用的模型结构 (含预训练模型) |
| torchvision.transforms | 常用的图片变化，例如裁剪、旋转 |
| torchvision…utils | 其他方法 |
代码已上传至github：
https://github.com/InkiInki/Python/blob/master/Python1/deepLearning/ImageMnist.py
1 获取数据集
需要导入的包如下：
import torch
import torchvision
import torchvision.transforms as transforms
import matplotlib.pyplot as plt
import time
import sys
from IPython import display
下面，将通过torchvision.datasets下载数据集，第一次调用时会自动从网上获取数据 (若出现速度较慢，请向后查看注意)；通过参数train来指定获取训练集或者测试集；通过transform = transforms.Tensor()将数据转化为Tensor，如果不转换，则返回PIL图片。
transforms.Tensor()将尺寸为 (
)且数据位于 (0, 255)的PIL图片或数据类型为np.uint8的Numpy转换为尺寸为 (
)且数据类型为torch.float32且位于 (0.0, 1.0)的Tensor。
使用代码如下：
class ImageMnist():
def __init__(self):
self.mnist_train = torchvision.datasets.FashionMNIST(root='~/DataSets/FashionMNIST',
train=True, download=True, transform=transforms.ToTensor())
self.mnist_test = torchvision.datasets.FashionMNIST(root='~/DataSets/FashionMNIST',
train=False, download=True, transform=transforms.ToTensor())
if __name__ == "__main__":
test = ImageDataSet()
test.__init__()
print(test.mnist_train)
print(len(test.mnist_train), len(test.mnist_test))
运行结果：
Dataset FashionMNIST
Number of datapoints: 60000
Root location: C:\Users\Administrator/DataSets/FashionMNIST
Split: Train
StandardTransform
Transform: ToTensor()
60000 10000
注意：
1）如果用像素值表示图片数据，那么一律将其类型设置成unit8，以避免不必要的bug；
2）第一次下载时速度也许很慢，推荐在cmd中输入以下代码，并复制出现的http链接下载：
import torchvision
import torchvision.transforms as transforms
torchvision.datasets.FashionMNIST(root='~/DataSets/FashionMNIST', train=True, download=True, transform=transforms.ToTensor())
torchvision.datasets.FashionMNIST(root='~/DataSets/FashionMNIST', train=False, download=True, transform=transforms.ToTensor())
2 简单操作
可以通过下标来访问任意一个样本：
if __name__ == "__main__":
test = ImageMnist()
test.__init__()
data, label = test.mnist_train[0]
print(data.shape)
print(label)
运行结果：
torch.Size([1, 28, 28]) # 分别对应通道数、图像高、图像宽
9
Fashion-MNIST共10个类别，分别为t-shirt、trouser、pullover、dress、coat、sandal、shirt、sneaker、bag和ankle boot，以下函数可以将数值标签转换成相应的文本标签：
...
def get_text_labels(self, labels):
text_labels = ['t-shirt', 'trouser', 'pullover', 'dress', 'coat', 'sandal', 'shirt', 'sneaker', 'bag', 'ankle boot']
return [text_labels[int(i)] for i in labels]
if __name__ == "__main__":
test = ImageMnist()
test.__init__()
data, label = test.mnist_train[0]
print(test.get_text_labels([label]))
运行结果：
['ankle boot']
现在定义一个可以在一行里画出多张图像和对应标签的函数：
...
def show_mnist(self, images, labels):
display.set_matplotlib_formats('svg')
_, figs = plt.subplots(1, len(images), figsize=(12, 12))
# zip()接受一系列可迭代对象作为参数，将对象中对应的元素打包成一个个元组，然后返回由这些元组组成的列表
for f, img, lbl in zip(figs, images, labels):
f.imshow(img.view((28, 28)).numpy())
f.set_title(lbl)
f.axis('off')
plt.show()
if __name__ == "__main__":
test = ImageMnist()
test.__init__()
x, y = [], []
for i in range(10):
x.append(test.mnist_train[i][0])
y.append(test.mnist_train[i][1])
test.show_mnist(x, test.get_text_labels(y))
运行结果：
3 读取小批量
torch的DataLoader中一个很方便的功能是运行使用多进程来加速读取数据，这里通过参数num_workers来设置4个进程读取数据。
...
def data_iter(self, batch_size=256):
if sys.platform.startswith('win'):
num_workers = 0 # 0表示不需要额外的进程来加速读取数据
else:
num_workers = 4
train_iter = torch.utils.data.DataLoader(self.mnist_train,
batch_size=batch_size, shuffle=True, num_workers=num_workers)
test_iter = torch.utils.data.DataLoader(self.mnist_test,
batch_size=batch_size, shuffle=False, num_workers=num_workers)
return train_iter, test_iter
if __name__ == "__main__":
start = time.time()
test = ImageMnist()
test.__init__()
train_iter, test_iter = test.data_iter()
for x, y in train_iter:
continue
print("%.2f sec" % (time.time() - start))
运行结果：
6.65 sec
4 完整代码
'''
@(#)test.py
The class of test.
Author: Yu-Xuan Zhang
Email: [email protected]
Created on May 05, 2020
Last Modified on May 05, 2020
@author: inki
'''
import torch
import torchvision
import torchvision.transforms as transforms
import matplotlib.pyplot as plt
import time
import sys
from IPython import display
class ImageMnist():
def __init__(self):
self.mnist_train = torchvision.datasets.FashionMNIST(root='~/DataSets/FashionMNIST',
train=True, download=True, transform=transforms.ToTensor())
self.mnist_test = torchvision.datasets.FashionMNIST(root='~/DataSets/FashionMNIST',
train=False, download=True, transform=transforms.ToTensor())
def get_text_labels(self, labels):
text_labels = ['t-shirt', 'trouser', 'pullover', 'dress', 'coat', 'sandal', 'shirt', 'sneaker', 'bag', 'ankle boot']
return [text_labels[int(i)] for i in labels]
def show_mnist(self, images, labels):
display.set_matplotlib_formats('svg')
_, figs = plt.subplots(1, len(images), figsize=(12, 12))
for f, img, lbl in zip(figs, images, labels):
f.imshow(img.view((28, 28)).numpy())
f.set_title(lbl)
f.axis('off')
plt.show()
def data_iter(self, batch_size=256):
if sys.platform.startswith('win'):
num_workers = 0
else:
num_workers = 4
train_iter = torch.utils.data.DataLoader(self.mnist_train,
batch_size=batch_size, shuffle=True, num_workers=num_workers)
test_iter = torch.utils.data.DataLoader(self.mnist_test,
batch_size=batch_size, shuffle=False, num_workers=num_workers)
return train_iter, test_iter
if __name__ == "__main__":
start = time.time()
test = ImageMnist()
test.__init__()
train_iter, test_iter = test.data_iter()
for x, y in train_iter:
continue
print("%.2f sec" % (time.time() - start))
致谢
特别感谢李沐、Aston Zhang等老师的这本《动手学深度学习》一书~
原创文章 35
获赞 44
访问量 8628