【PyTorch 攻略】（6-7/7）

一、说明

本篇介绍模型模型的参数，模型推理和使用，保存加载。

二、训练参数和模型

在本单元中，我们将了解如何加载模型及其持久参数状态和推理模型预测。为了加载模型，我们将定义模型类，其中包含用于训练模型的神经网络的状态和参数。

%matplotlib inline
import torch
import onnxruntime
from torch import nn
import torch.onnx as onnx
import torchvision.models as models
from torchvision import datasets
from torchvision.transforms import ToTensor

class NeuralNetwork(nn.Module):
    def __init__(self):
        super(NeuralNetwork, self).__init__()
        self.flatten = nn.Flatten()
        self.linear_relu_stack = nn.Sequential(
            nn.Linear(28*28, 512),
            nn.ReLU(),
            nn.Linear(512, 512),
            nn.ReLU(),
            nn.Linear(512, 10),
            nn.ReLU()
        )

    def forward(self, x):
        x = self.flatten(x)
        logits = self.line

加载模型权重时，我们需要首先实例化模型类，因为该类定义了网络的结构。接下来，我们使用 load_state_dict（） 方法加载参数。

model = NeuralNetwork()
model.load_state_dict(torch.load('data/model.pth'))
model.eval()

注意：请务必在推理之前调用 model.eval（） 方法，以将 dropout 和批量归一化层设置为评估模式。如果不这样做，将产生不一致的推理结果。

三、模型推理

优化模型以在各种平台和编程语言上运行是很困难的。在所有不同的框架和硬件组合中最大限度地提高性能是非常耗时的。
开放式神经网络交换 （ONNX） 运行时为您提供了一种解决方案，只需训练一次，即可在任何硬件、云或边缘设备上加速推理。

ONNX 是许多供应商支持的一种通用格式，用于共享神经网络和其他机器学习模型。您可以使用 ONNX 格式在其他编程语言和框架（如 Java、JavaScript、C# 和 ML.NET）上对模型进行推理。

input_image = torch.zeros((1,28,28))
onnx_model = 'data/model.onnx'
onnx.export(model, input_image, onnx_model)

我们将使用测试数据集作为示例数据，以便从 ONNX 模型进行推理以进行预测。

test_data = datasets.FashionMNIST(
    root="data",
    train=False,
    download=True,
    transform=ToTensor()
)

classes = [
    "T-shirt/top",
    "Trouser",
    "Pullover",
    "Dress",
    "Coat",
    "Sandal",
    "Shirt",
    "Sneaker",
    "Bag",
    "Ankle boot",
]
x, y = test_data[0][0], test_data[0][1]

我们需要使用 onnxruntime 创建一个推理会话。推理会话。为了推断 onnx 模型，我们使用 run 和 pass 输入要返回的输出列表（如果需要所有输出，请留空）和输入值映射。结果是一个输出列表：

session = onnxruntime.InferenceSession(onnx_model, None)
input_name = session.get_inputs()[0].name
output_name = session.get_outputs()[0].name

result = session.run([output_name], {input_name: x.numpy()})
predicted, actual = classes[result[0][0].argmax(0)], classes[y]
print(f'Predicted: "{predicted}", Actual: "{actual}"')

**四、**torch.utils.data.DataLoader 和torch.utils.data.Dataset

PyTorch有两个基元来处理数据：torch.utils.data.DataLoader 和torch.utils.data.Dataset 。数据集 存储样本及其相应的标签，DataLoader 围绕数据集包装一个可迭代对象。

%matplotlib inline
import torch
from torch import nn
from torch.utils.data import DataLoader
from torchvision import datasets
from torchvision.transforms import ToTensor, Lambda, Compose
import matplotlib.pyplot as plt

PyTorch提供特定于领域的库，如TorchText，TorchVision和TorchAudio， 所有这些都包括数据集。在本教程中，我们将使用TorchVision数据集。

torchvision.datasets 模块包含许多真实世界视觉数据（如 CIFAR 和 COCO）的数据集 对象。在本教程中，我们将使用 FashionMNIST 数据集。每个TorchVision 数据集都包含两个参数：转换 和target_transform分别修改样本和标签。

# Download training data from open datasets.
training_data = datasets.FashionMNIST(
    root="data",
    train=True,
    download=True,
    transform=ToTensor(),
)

# Download test data from open datasets.
test_data = datasets.FashionMNIST(
    root="data",
    train=False,
    download=True,
    transform=ToTensor(),
)

我们将数据集 作为参数传递给 DataLoader。这将在我们的数据集上包装一个可迭代对象，并支持自动批处理、采样、随机排序和多进程数据加载。这里我们定义一个 64 的批量大小，即 dataloader 迭代中的每个元素将返回一批 64 个特征和标签。

batch_size = 64

# Create data loaders.
train_dataloader = DataLoader(training_data, batch_size=batch_size)
test_dataloader = DataLoader(test_data, batch_size=batch_size)

for X, y in test_dataloader:
    print("Shape of X [N, C, H, W]: ", X.shape)
    print("Shape of y: ", y.shape, y.dtype)
    break
    
# Display sample data
figure = plt.figure(figsize=(10, 8))
cols, rows = 5, 5
for i in range(1, cols * rows + 1):
    idx = torch.randint(len(test_data), size=(1,)).item()
    img, label = test_data[idx]
    figure.add_subplot(rows, cols, i)
    plt.title(label)
    plt.axis("off")
    plt.imshow(img.squeeze(), cmap="gray")
plt.show()

Shape of X [N, C, H, W]:  torch.Size([64, 1, 28, 28])
Shape of y:  torch.Size([64]) torch.int64

五、创建模型

为了在 PyTorch 中定义神经网络，我们创建一个继承自 nn.Module 的类。我们在 init 函数中定义网络层，并在转发函数中指定数据如何通过网络。为了加速神经网络的运算，我们将其转移到 GPU（如果可用）。

# Get cpu or gpu device for training.
device = "cuda" if torch.cuda.is_available() else "cpu"
print("Using {} device".format(device))

# Define model
class NeuralNetwork(nn.Module):
    def __init__(self):
        super(NeuralNetwork, self).__init__()
        self.flatten = nn.Flatten()
        self.linear_relu_stack = nn.Sequential(
            nn.Linear(28*28, 512),
            nn.ReLU(),
            nn.Linear(512, 512),
            nn.ReLU(),
            nn.Linear(512, 10),
            nn.ReLU()
        )

    def forward(self, x):
        x = self.flatten(x)
        logits = self.linear_relu_stack(x)
        return logits

model = NeuralNetwork().to(device)
print(model)

Using cuda device
NeuralNetwork(
  (flatten): Flatten()
  (linear_relu_stack): Sequential(
    (0): Linear(in_features=784, out_features=512, bias=True)
    (1): ReLU()
    (2): Linear(in_features=512, out_features=512, bias=True)
    (3): ReLU()
    (4): Linear(in_features=512, out_features=10, bias=True)
    (5): ReLU()
  )
)

六、优化模型参数

为了训练模型，我们需要一个损失函数和一个优化器。我们将使用 nn。交叉熵损失 用于损失，随机梯度下降用于优化。

loss_fn = nn.CrossEntropyLoss()
learning_rate = 1e-3
optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)

在单个训练循环中，模型对训练数据集进行预测（批量馈送到它），并向后传播预测误差以调整模型的参数。

def train(dataloader, model, loss_fn, optimizer):
    size = len(dataloader.dataset)
    for batch, (X, y) in enumerate(dataloader):
        X, y = X.to(device), y.to(device)
        
        # Compute prediction error
        pred = model(X)
        loss = loss_fn(pred, y)
        
        # Backpropagation
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

        if batch % 100 == 0:
            loss, current = loss.item(), batch * len(X)

我们还可以对照测试数据集检查模型的性能，以确保它正在学习。

def test(dataloader, model):
    size = len(dataloader.dataset)
    model.eval()
    test_loss, correct = 0, 0
    with torch.no_grad():
        for X, y in dataloader:
            X, y = X.to(device), y.to(device)
            pred = model(X)
            test_loss += loss_fn(pred, y).item()
            correct += (pred.argmax(1) == y).type(torch.float).sum().item()
    test_loss /= size
    correct /= size
    print(f"Test Error: \n Accuracy: {(100*correct):>0.1f}%, Avg loss: {test_loss:>8f} \n")

训练过程通过多次迭代（纪元）进行。在每个时期，模型学习参数以做出更好的预测。我们打印模型在每个时期的准确性和损失;我们希望看到精度随着每个时期的增加和损失的减少而减少。

epochs = 15
for t in range(epochs):
    print(f"Epoch {t+1}\n-------------------------------")
    train(train_dataloader, model, loss_fn, optimizer)
    test(test_dataloader, model)
print("Done!")

Epoch 1
-------------------------------
loss: 2.295450  [    0/60000]
loss: 2.293073  [ 6400/60000]
loss: 2.278504  [12800/60000]
loss: 2.282501  [19200/60000]
loss: 2.273211  [25600/60000]
loss: 2.258452  [32000/60000]
loss: 2.248237  [38400/60000]
loss: 2.228594  [44800/60000]
loss: 2.240276  [51200/60000]
loss: 2.221318  [57600/60000]
Test Error: 
 Accuracy: 51.8%, Avg loss: 0.034745 

Epoch 2
-------------------------------
loss: 2.212354  [    0/60000]
loss: 2.207739  [ 6400/60000]
loss: 2.160400  [12800/60000]
loss: 2.176181  [19200/60000]
loss: 2.168270  [25600/60000]
loss: 2.146453  [32000/60000]
loss: 2.119934  [38400/60000]
loss: 2.083791  [44800/60000]
loss: 2.126453  [51200/60000]
loss: 2.077550  [57600/60000]
Test Error: 
 Accuracy: 53.2%, Avg loss: 0.032452 

Epoch 3
-------------------------------
loss: 2.082280  [    0/60000]
loss: 2.068733  [ 6400/60000]
loss: 1.965958  [12800/60000]
loss: 1.997126  [19200/60000]
loss: 2.002057  [25600/60000]
loss: 1.967370  [32000/60000]
loss: 1.910595  [38400/60000]
loss: 1.849006  [44800/60000]
loss: 1.944741  [51200/60000]
loss: 1.861265  [57600/60000]
Test Error: 
 Accuracy: 51.6%, Avg loss: 0.028937 

Epoch 4
-------------------------------
loss: 1.872628  [    0/60000]
loss: 1.844543  [ 6400/60000]
loss: 1.710179  [12800/60000]
loss: 1.779804  [19200/60000]
loss: 1.737971  [25600/60000]
loss: 1.746953  [32000/60000]
loss: 1.624768  [38400/60000]
loss: 1.575720  [44800/60000]
loss: 1.742827  [51200/60000]
loss: 1.653375  [57600/60000]
Test Error: 
 Accuracy: 58.4%, Avg loss: 0.025570 

Epoch 5
-------------------------------
loss: 1.662315  [    0/60000]
loss: 1.636235  [ 6400/60000]
loss: 1.508407  [12800/60000]
loss: 1.606842  [19200/60000]
loss: 1.560728  [25600/60000]
loss: 1.606024  [32000/60000]
loss: 1.426900  [38400/60000]
loss: 1.406240  [44800/60000]
loss: 1.619918  [51200/60000]
loss: 1.521326  [57600/60000]
Test Error: 
 Accuracy: 61.2%, Avg loss: 0.023459 

Epoch 6
-------------------------------
loss: 1.527535  [    0/60000]
loss: 1.511209  [ 6400/60000]
loss: 1.377129  [12800/60000]
loss: 1.494889  [19200/60000]
loss: 1.457990  [25600/60000]
loss: 1.502333  [32000/60000]
loss: 1.291539  [38400/60000]
loss: 1.285098  [44800/60000]
loss: 1.484891  [51200/60000]
loss: 1.414015  [57600/60000]
Test Error: 
 Accuracy: 62.2%, Avg loss: 0.021480 

Epoch 7
-------------------------------
loss: 1.376779  [    0/60000]
loss: 1.384830  [ 6400/60000]
loss: 1.230116  [12800/60000]
loss: 1.382574  [19200/60000]
loss: 1.255630  [25600/60000]
loss: 1.396211  [32000/60000]
loss: 1.157718  [38400/60000]
loss: 1.186382  [44800/60000]
loss: 1.340606  [51200/60000]
loss: 1.321607  [57600/60000]
Test Error: 
 Accuracy: 62.8%, Avg loss: 0.019737 

Epoch 8
-------------------------------
loss: 1.243344  [    0/60000]
loss: 1.279124  [ 6400/60000]
loss: 1.121769  [12800/60000]
loss: 1.293069  [19200/60000]
loss: 1.128232  [25600/60000]
loss: 1.315465  [32000/60000]
loss: 1.069528  [38400/60000]
loss: 1.123324  [44800/60000]
loss: 1.243827  [51200/60000]
loss: 1.255190  [57600/60000]
Test Error: 
 Accuracy: 63.4%, Avg loss: 0.018518 

Epoch 9
-------------------------------
loss: 1.154148  [    0/60000]
loss: 1.205280  [ 6400/60000]
loss: 1.046463  [12800/60000]
loss: 1.229866  [19200/60000]
loss: 1.048813  [25600/60000]
loss: 1.254785  [32000/60000]
loss: 1.010614  [38400/60000]
loss: 1.077114  [44800/60000]
loss: 1.176766  [51200/60000]
loss: 1.206567  [57600/60000]
Test Error: 
 Accuracy: 64.3%, Avg loss: 0.017640 

Epoch 10
-------------------------------
loss: 1.090360  [    0/60000]
loss: 1.149150  [ 6400/60000]
loss: 0.990786  [12800/60000]
loss: 1.183704  [19200/60000]
loss: 0.997114  [25600/60000]
loss: 1.207199  [32000/60000]
loss: 0.967512  [38400/60000]
loss: 1.043431  [44800/60000]
loss: 1.127000  [51200/60000]
loss: 1.169639  [57600/60000]
Test Error: 
 Accuracy: 65.3%, Avg loss: 0.016974 

Epoch 11
-------------------------------
loss: 1.041194  [    0/60000]
loss: 1.104409  [ 6400/60000]
loss: 0.947670  [12800/60000]
loss: 1.149421  [19200/60000]
loss: 0.960403  [25600/60000]
loss: 1.169899  [32000/60000]
loss: 0.935149  [38400/60000]
loss: 1.018250  [44800/60000]
loss: 1.088222  [51200/60000]
loss: 1.139813  [57600/60000]
Test Error: 
 Accuracy: 66.2%, Avg loss: 0.016446 

Epoch 12
-------------------------------
loss: 1.000646  [    0/60000]
loss: 1.067356  [ 6400/60000]
loss: 0.912046  [12800/60000]
loss: 1.122742  [19200/60000]
loss: 0.932827  [25600/60000]
loss: 1.138785  [32000/60000]
loss: 0.910242  [38400/60000]
loss: 0.999010  [44800/60000]
loss: 1.056596  [51200/60000]
loss: 1.114582  [57600/60000]
Test Error: 
 Accuracy: 67.5%, Avg loss: 0.016011 

Epoch 13
-------------------------------
loss: 0.966393  [    0/60000]
loss: 1.035691  [ 6400/60000]
loss: 0.881672  [12800/60000]
loss: 1.100845  [19200/60000]
loss: 0.910265  [25600/60000]
loss: 1.112597  [32000/60000]
loss: 0.889558  [38400/60000]
loss: 0.982751  [44800/60000]
loss: 1.029199  [51200/60000]
loss: 1.092738  [57600/60000]
Test Error: 
 Accuracy: 68.5%, Avg loss: 0.015636 

Epoch 14
-------------------------------
loss: 0.936334  [    0/60000]
loss: 1.007734  [ 6400/60000]
loss: 0.854663  [12800/60000]
loss: 1.081601  [19200/60000]
loss: 0.890581  [25600/60000]
loss: 1.089641  [32000/60000]
loss: 0.872057  [38400/60000]
loss: 0.969192  [44800/60000]
loss: 1.005193  [51200/60000]
loss: 1.073098  [57600/60000]
Test Error: 
 Accuracy: 69.4%, Avg loss: 0.015304 

Epoch 15
-------------------------------
loss: 0.908971  [    0/60000]
loss: 0.982067  [ 6400/60000]
loss: 0.830095  [12800/60000]
loss: 1.064921  [19200/60000]
loss: 0.874204  [25600/60000]
loss: 1.069008  [32000/60000]
loss: 0.856447  [38400/60000]
loss: 0.957340  [44800/60000]
loss: 0.983547  [51200/60000]
loss: 1.055251  [57600/60000]
Test Error: 
 Accuracy: 70.3%, Avg loss: 0.015001 

Done!

准确性最初不会很好（没关系！尝试运行循环以获取更多纪元 或将learning_rate调整为更大的数字。也可能是我们选择的模型配置可能不是此类问题的最佳配置。

七、保存模型

保存模型的常用方法是序列化内部状态字典（包含模型参数）。

torch.save(model.state_dict(), "data/model.pth")
print("Saved PyTorch Model State to model.pth")

八、负载模型

加载模型的过程包括重新创建模型结构并将状态字典加载到其中。

model = NeuralNetwork()
model.load_state_dict(torch.load("data/model.pth"))

此模型现在可用于进行预测。

classes = [
    "T-shirt/top",
    "Trouser",
    "Pullover",
    "Dress",
    "Coat",
    "Sandal",
    "Shirt",
    "Sneaker",
    "Bag",
    "Ankle boot",
]

model.eval()
x, y = test_data[0][0], test_data[0][1]
with torch.no_grad():
    pred = model(x)
    predicted, actual = classes[pred[0].argmax(0)], classes[y]
    print(f'Predicted: "{predicted}", Actual: "{actual}"')

Predicted: "Ankle boot", Actual: "Ankle boot"

祝贺！您已经完成了 PyTorch 初学者教程！我们希望本教程能帮助您在 PyTorch 上开始深度学习。