mnisttorch加载fashion_简单的使用PyTorch和FashionMNIST数据集进行深度学习图像分类...

概述

本文的目的是为那些想要使用PyTorch和Fashion MNIST进行简单深度学习图像分类网络的人提供示例参考代码。

在本文中&＃xff0c;我们将演示深度学习图像分类网络的所有工作部分&＃xff0c;包括加载数据&＃xff0c;定义网络&＃xff0c;优化GPU上的权重以及评估性能。

整理Fashion MNIST数据集

Fashion MNIST是一个包含70,000个灰度图像和10个类的数据集。

1.检查GPU是否可用

import torch

print(torch.cuda.is_available())

device &＃61; torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

print(device)

2.下载并加载Fashion MNIST数据集

import torch

from torchvision import datasets, transforms

import helper

# Instructions from here:

# https://www.kaggle.com/ishvindersethi22/fashion-mnist-using-pytorch/data

# Define a transform to normalize the data

transform &＃61; transforms.Compose([transforms.ToTensor(),

transforms.Normalize([0.], [0.5])])

# Download and load the training data

trainset &＃61; datasets.FashionMNIST(&＃39;~/.pytorch/F_MNIST_data/&＃39;, download&＃61;True, train&＃61;True, transform&＃61;transform)

trainloader &＃61; torch.utils.data.DataLoader(trainset, batch_size&＃61;64, shuffle&＃61;True)

# Download and load the test data

testset &＃61; datasets.FashionMNIST(&＃39;~/.pytorch/F_MNIST_data/&＃39;, download&＃61;True, train&＃61;False, transform&＃61;transform)

testloader &＃61; torch.utils.data.DataLoader(testset, batch_size&＃61;64, shuffle&＃61;True)

3.显示样本图像

import matplotlib.pyplot as plt

import torchvision

import numpy as np

def imshow(img):

img &＃61; img / 2 &＃43; 0. # unnormalize

img &＃61; img.squeeze()

plt.imshow(img, cmap&＃61;&＃39;gray&＃39;)

images, label &＃61; next(iter(trainloader))

imshow(images[0, :])

4.定义和训练网络

from collections import OrderedDict

from torch import optim, nn

hidden_units &＃61; [4, 8, 16]

output_units &＃61; 10

class Flatten(nn.Module):

def forward(self, input):

return input.view(input.size(0), -1)

model_d &＃61; nn.Sequential(OrderedDict([

(&＃39;conv1&＃39;, nn.Conv2d(1, hidden_units[0], 3, stride&＃61;2, padding&＃61;1)),

(&＃39;Relu1&＃39;, nn.ReLU()),

(&＃39;conv2&＃39;, nn.Conv2d(hidden_units[0], hidden_units[1], 3, stride&＃61;2, padding&＃61;1)),

(&＃39;Relu2&＃39;, nn.ReLU()),

(&＃39;conv3&＃39;, nn.Conv2d(hidden_units[1], hidden_units[2], 3, stride&＃61;2, padding&＃61;1)),

(&＃39;Relu3&＃39;, nn.ReLU()),

(&＃39;conv4&＃39;, nn.Conv2d(hidden_units[2], output_units, 4, stride&＃61;4, padding&＃61;0)),

(&＃39;log_softmax&＃39;, nn.LogSoftmax(dim &＃61; 1))

]))

model_d.to(device)

optimizer_d &＃61; optim.Adam(model_d.parameters(), lr &＃61; 0.01)

criterion &＃61; nn.NLLLoss()

epochs &＃61; 10

for i in range(epochs):

running_classification_loss &＃61; 0

running_cycle_consistent_loss &＃61; 0

running_loss &＃61; 0

for images, labels in trainloader:

images, labels &＃61; images.to(device), labels.to(device)

optimizer_d.zero_grad()

# Run classification model

predicted_labels &＃61; model_d(images)

classification_loss &＃61; criterion(Flatten()(predicted_labels), labels)

# Optimize classification weights

classification_loss.backward()

optimizer_d.step()

running_classification_loss &＃43;&＃61; classification_loss.item()

running_loss &＃61; running_classification_loss

else:

print(f"{i} Training loss: {running_loss/len(trainloader)}")

5.评估网络

total_correct &＃61; 0

total_num &＃61; 0

for images, labels in testloader:

images, labels &＃61; images.to(device), labels.to(device)

ps &＃61; Flatten()(torch.exp(model_d(images)))

predictions &＃61; ps.topk(1, 1, True, True)[1].t()

correct &＃61; predictions.eq(labels.view(1, -1))

total_correct &＃43;&＃61; correct.sum().cpu().numpy()

total_num &＃43;&＃61; images.shape[0]

print(&＃39;Accuracy:&＃39;, total_correct / float(total_num))

print(&＃39;Correct Label:&＃39;, labels[0].item())

print(&＃39;Predicted Label:&＃39;, predictions[0, 0].item())

index &＃61; 0

imshow(images[index, :].cpu())

小贴士

本文演示了深度学习图像分类网络的所有工作部分&＃xff0c;我们可以使用最基本的人工智能应用来进行简单的学习。

我们加载了Fashion MNIST数据集

定义一个简单的深度卷积网络

我们使用GPU上的Adam优化器优化网络权重

我们评估网络并达到约85&＃xff05;的准确度