Deep Learning with Pytorch 读书笔记 A pretrained network that recognizes the subject of an image

￡神魔★判官ぃ 2023-02-26 15:22 17阅读 0赞

# 安装pytorch #

由于笔记本CUDA版本过低，如果要装pytorch 1.5.1只能以cpuonly安装：  
![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg_size_16_color_FFFFFF_t_70]

![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg_size_16_color_FFFFFF_t_70 1]  
![在这里插入图片描述][20200718232226849.png]

安装v1.1.0从而利用笔记本上的GPU  
![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg_size_16_color_FFFFFF_t_70 2]  
![在这里插入图片描述][20200718233342829.png]  
![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg_size_16_color_FFFFFF_t_70 3]  
pytorch下载速度过慢，如无法科学上网，可借助清华镜像源来下载相应的包  
![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg_size_16_color_FFFFFF_t_70 4]  
![在这里插入图片描述][20200719140549235.png]  
下载的包放到下面的路径  
![在这里插入图片描述][2020071914073865.png]  
根据下载的文件修改anaconda的下载列表  
![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg_size_16_color_FFFFFF_t_70 5]  
![在这里插入图片描述][20200719145805239.png]  
再次执行conda install命令  
或者直接本地安装：conda install --use-local pytorch-1.1.0-py3.7\_cuda90\_cudnn7\_1.tar.bz2  
![在这里插入图片描述][20200719151429203.png]

# Pytorch的在CV中的用途： #

*  object-recognition
 *  image classification (telling what object categories the image contains) object localization (identifying objects’ position in images)
 *  object detection (identifying and labeling objects in images)
 *  scene classification (classifying a situation in an image)
 *  scene parsing (segmenting an image into regions associated with semantic categories, such as cow, house, cheese, hat)

# 使用预训练的网络进行第一次接触 #

预训练的网络使用的是ImageNet数据集，所有图片都有来自WordNet的标签。  
The pretrained network we’ll explore here was trained on a subset of the ImageNet dataset (http://imagenet.stanford.edu). ImageNet is a very large dataset of over 14 million images maintained by Stanford University. All of the images are labeled with a hierarchy of nouns that come from the WordNet dataset (http://wordnet.princeton.edu), which is in turn a large lexical database of the English language.

输入图像经过处理后存入多维矩阵类torch. Tensor。图像是RGB，有长度和宽度。tensor有三个维度: 三色通道，两个指定大小的空间图像维度。  
The input image will first be preprocessed into an instance of the multidimensional array class torch.Tensor. It is an RGB image with height and width, so this tensor will have three dimensions: the three color channels, and two spatial image dimensions of a specific size.

Our model will take that processed input image and pass it into the pretrained network to obtain scores for each class. The highest score corresponds to the most likely class according to the weights. Each class is then mapped one-to-one onto a class label. That output is contained in a torch.Tensor with 1,000 elements, each representing the score associated with that class.

TorchVision project (https://github.com/pytorch/vision) 包含一些计算机视觉的神经网络架构，比如AlexNet、ResNet和Inception v3。  
which contains a few of the best-performing neural network architectures for computer vision, such as AlexNet (http://mng.bz/lo6z), ResNet (https://arxiv.org/pdf/1512.03385.pdf), and Inception v3 (https://arxiv.org/pdf/1512.00567.pdf).

The predefined models can be found in torchvision.models

In[1]:
    from torchvision import models
     In[2]:
    dir(models)
     Out[2]:
    ['AlexNet',
    'DenseNet',
    'Inception3',
    'ResNet',
    'SqueezeNet',
    'VGG',
    ...
    'alexnet',
    'densenet',
    'densenet121',
    ...
    'resnet',
    'resnet101',
    'resnet152',
    ...
    ]

[Python模块问题][Python]：ImportError: cannot import name ‘PILLOW\_VERSION‘ from ‘PIL‘

TorchVision project中**大写字母开头的是模型类，而小写字母开头的是工厂函数**。比如：  
创建AlexNet类实例

In[3]:
    alexnet = models.AlexNet()

创建的实例是未配置未训练的。the network is uninitialized: its weights, the numbers by which inputs are added and multiplied, have not been trained on anything—the network itself is a blank (or rather, random) slate

创建ResNet类实例  
We’ll pass an argument that will instruct the function to download the weights of resnet101 trained on the ImageNet  
dataset, with 1.2 million images and 1,000 categories: 加载了权重的101层卷积神经网络的resnet实例

In[4]:
    resnet = models.resnet101(pretrained=True)
    In[5]:
    resnet
    Out[5]:
    ResNet(
    (conv1): Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3),
    bias=False)
    (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True,
    track_running_stats=True)
    (relu): ReLU(inplace)
    (maxpool): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1,
    ceil_mode=False)
    (layer1): Sequential(
    (0): Bottleneck(
    ...
    )
    )
    (avgpool): AvgPool2d(kernel_size=7, stride=1, padding=0)
    (fc): Linear(in_features=2048, out_features=1000, bias=True)
    )

## ResNet ##

对图像需要进行预处理，比如大小和值的范围要合法。torchvision模块提供了转换，用于基础预处理的函数。  
Before we can do that, however, we have to preprocess the input images so  
they are the right size and so that their values (colors) sit roughly in the same numerical range. In order to do that, the torchvision module provides transforms, which  
allow us to quickly define pipelines of basic preprocessing functions:

In\[6\]:  
from torchvision import transforms  
preprocess = transforms.Compose(\[  
transforms.Resize(256),  
transforms.CenterCrop(224),  
transforms.ToTensor(),  
transforms.Normalize(  
mean=\[0.485, 0.456, 0.406\],  
std=\[0.229, 0.224, 0.225\]  
)\])

We can start by loading an image from the local filesystem using Pillow (https://pillow.readthedocs.io/en/stable), an image-manipulation module for Python  
![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg_size_16_color_FFFFFF_t_70 6]

The process of running a trained model on new data is called inference in deep learning  
circles. In order to do inference, we need to put the network in eval mode  
![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg_size_16_color_FFFFFF_t_70 7]  
![在这里插入图片描述][20200809182642256.png]

producing a vector of 1,000 scores, one per ImageNet class 产生包含1000个分数向量，每个分数对于于一个ImageNNet类的值。

Let’s load the file containing the 1,000 labels for the ImageNet dataset classes:

with open('../data/p1ch2/imagenet_classes.txt') as f:
    	labels = [line.strip() for line in f.readlines()]
    _, index = torch.max(out, 1)
    percentage = torch.nn.functional.softmax(out, dim=1)[0] * 100
    labels[index[0]], percentage[index[0]].item()

Out\[15\]: (‘golden retriever’, 96.29334259033203)  
Out\[16\]:  
\[(‘golden retriever’, 96.29334259033203),  
(‘Labrador retriever’, 2.80812406539917),  
(‘cocker spaniel, English cocker spaniel, cocker’, 0.28267428278923035),  
(‘redbone’, 0.2086310237646103),  
(‘tennis ball’, 0.11621569097042084)\]

![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg_size_16_color_FFFFFF_t_70 8]

## CycleGAN ##

GAN stands for generative adversarial network, where generative means something is being created (in this case, fake masterpieces), adversarial means the two networks are competing to outsmart the other, and well, network is pretty obvious. The generator network takes the role of the painter in our scenario, tasked with producing realistic-looking images, starting from an arbitrary input. The discriminator network is the amoral art inspector, needing to tell whether a given image was fabricated by the generator or belongs in a set of real images.  
![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg_size_16_color_FFFFFF_t_70 9]  
A CycleGAN can turn images of one domain into images of another domain (and back), without the need for us to explicitly provide matching pairs in the training set.  
![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg_size_16_color_FFFFFF_t_70 10]  
As the figure shows, the first generator learns to produce an image conforming to a target distribution (zebras, in this case) starting from an image belonging to a different distribution (horses), so that the discriminator can’t tell if the image produced from a horse photo is actually a genuine picture of a zebra or not. At the same time—and here’s where the Cycle prefix in the acronym comes in—the resulting fake zebra is sent through a different generator going the other way (zebra to horse, in our case), to be judged by another discriminator on the other side. Creating such a cycle stabilizes the training process considerably, which addresses one of the original issues with GANs.

netG = ResNetGenerator()  # The netG model has been created, but it contains random weights.
    model_path = '../data/p1ch2/horse2zebra_0.4.0.pth' # The weights of the model have been saved in a .pth file
    model_data = torch.load(model_path)
    netG.load_state_dict(model_data)
    netG.eval()

from PIL import Image
    from torchvision import transforms
    preprocess = transforms.Compose([transforms.Resize(256), transforms.ToTensor()])
    img = Image.open("../data/p1ch2/horse.jpg")
    img_t = preprocess(img)
    batch_t = torch.unsqueeze(img_t, 0)
    batch_out = netG(batch_t)   # batch_out is now the output of the generator
    out_t = (batch_out.data.squeeze() + 1.0) / 2.0
    out_img = transforms.ToPILImage()(out_t)

[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg_size_16_color_FFFFFF_t_70]: /images/20230209/c65088c811ae4aefa7b56c30dad2a176.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg_size_16_color_FFFFFF_t_70 1]: https://img-blog.csdnimg.cn/2020071823200496.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg==,size_16,color_FFFFFF,t_70
[20200718232226849.png]: https://img-blog.csdnimg.cn/20200718232226849.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg_size_16_color_FFFFFF_t_70 2]: https://img-blog.csdnimg.cn/20200718233120394.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg==,size_16,color_FFFFFF,t_70
[20200718233342829.png]: https://img-blog.csdnimg.cn/20200718233342829.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg_size_16_color_FFFFFF_t_70 3]: https://img-blog.csdnimg.cn/20200719112726789.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg_size_16_color_FFFFFF_t_70 4]: https://img-blog.csdnimg.cn/20200719140514727.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg==,size_16,color_FFFFFF,t_70
[20200719140549235.png]: https://img-blog.csdnimg.cn/20200719140549235.png
[2020071914073865.png]: https://img-blog.csdnimg.cn/2020071914073865.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg_size_16_color_FFFFFF_t_70 5]: https://img-blog.csdnimg.cn/20200719141259959.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg==,size_16,color_FFFFFF,t_70
[20200719145805239.png]: https://img-blog.csdnimg.cn/20200719145805239.png
[20200719151429203.png]: https://img-blog.csdnimg.cn/20200719151429203.png
[Python]: https://feishujun.blog.csdn.net/article/details/107446240
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg_size_16_color_FFFFFF_t_70 6]: https://img-blog.csdnimg.cn/20200809144818441.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg_size_16_color_FFFFFF_t_70 7]: https://img-blog.csdnimg.cn/20200809145544570.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg==,size_16,color_FFFFFF,t_70
[20200809182642256.png]: https://img-blog.csdnimg.cn/20200809182642256.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg_size_16_color_FFFFFF_t_70 8]: https://img-blog.csdnimg.cn/20200809182751947.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg_size_16_color_FFFFFF_t_70 9]: https://img-blog.csdnimg.cn/20200809205826976.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg==,size_16,color_FFFFFF,t_70
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg_size_16_color_FFFFFF_t_70 10]: https://img-blog.csdnimg.cn/20200809210438810.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2FzbWFydGtpbGxlcg==,size_16,color_FFFFFF,t_70