前面实现了基于多层感知机的MNIST手写体识别,本章将实现以卷积神经网络完成的MNIST手写体识别。
在本例中,依旧使用MNIST数据集,对这个数据集的数据和标签介绍,前面的章节已详细说明过了,相对于前面章节直接对数据进行“折叠”处理,这里需要显式地标注出数据的通道,代码如下:
- import numpy as np
-
- import einops.layers.torch as elt
-
- #载入数据
-
- x_train = np.load("../dataset/mnist/x_train.npy")
-
- y_train_label = np.load("../dataset/mnist/y_train_label.npy")
-
- x_train = np.expand_dims(x_train,axis=1) #在指定维度上进行扩充
-
- print(x_train.shape)
这里是对数据的修正,np.expand_dims的作用是在指定维度上进行扩充,这里在第二维(也就是PyTorch的通道维度)进行扩充,结果如下:
(60000, 1, 28, 28)
下面使用PyTorch 2.0框架对模型进行设计,在本例中将使用卷积层对数据进行处理,完整的模型如下:
- import torch
- import torch.nn as nn
- import numpy as np
- import einops.layers.torch as elt
- class MnistNetword(nn.Module):
- def __init__(self):
- super(MnistNetword, self).__init__()
- #前置的特征提取模块
- self.convs_stack = nn.Sequential(
- nn.Conv2d(1,12,kernel_size=7), #第一个卷积层
- nn.ReLU(),
- nn.Conv2d(12,24,kernel_size=5), #第二个卷积层
- nn.ReLU(),
- nn.Conv2d(24,6,kernel_size=3) #第三个卷积层
- )
- #最终分类器层
- self.logits_layer = nn.Linear(in_features=1536,out_features=10)
- def forward(self,inputs):
- image = inputs
- x = self.convs_stack(image)
- #elt.Rearrange的作用是对输入数据的维度进行调整,读者可以使用torch.nn.Flatten函数完成此工作
- x = elt.Rearrange("b c h w -> b (c h w)")(x)
- logits = self.logits_layer(x)
- return logits
- model = MnistNetword()
- torch.save(model,"model.pth")
这里首先设定了3个卷积层作为前置的特征提取层,最后一个全连接层作为分类器层,需要注意的是,对于分类器的全连接层,输入维度需要手动计算,当然读者可以一步一步尝试打印特征提取层的结果,依次将结果作为下一层的输入维度。最后对模型进行保存。
下面进入本章的最后示例部分,也就是MNIST手写体的分类。完整的训练代码如下:
- import torch
- import torch.nn as nn
- import numpy as np
- import einops.layers.torch as elt
- #载入数据
- x_train = np.load("../dataset/mnist/x_train.npy")
- y_train_label = np.load("../dataset/mnist/y_train_label.npy")
- x_train = np.expand_dims(x_train,axis=1)
- print(x_train.shape)
- class MnistNetword(nn.Module):
- def __init__(self):
- super(MnistNetword, self).__init__()
- self.convs_stack = nn.Sequential(
- nn.Conv2d(1,12,kernel_size=7),
- nn.ReLU(),
- nn.Conv2d(12,24,kernel_size=5),
- nn.ReLU(),
- nn.Conv2d(24,6,kernel_size=3)
- )
- self.logits_layer = nn.Linear(in_features=1536,out_features=10)
- def forward(self,inputs):
- image = inputs
- x = self.convs_stack(image)
- x = elt.Rearrange("b c h w -> b (c h w)")(x)
- logits = self.logits_layer(x)
- return logits
- device = "cuda" if torch.cuda.is_available() else "cpu"
- #注意记得将model发送到GPU计算
- model = MnistNetword().to(device)
- model = torch.compile(model)
- loss_fn = nn.CrossEntropyLoss()
- optimizer = torch.optim.SGD(model.parameters(), lr=1e-4)
- batch_size = 128
- for epoch in range(42):
- train_num = len(x_train)//128
- train_loss = 0.
- for i in range(train_num):
- start = i * batch_size
- end = (i + 1) * batch_size
- x_batch = torch.tensor(x_train[start:end]).to(device)
- y_batch = torch.tensor(y_train_label[start:end]).to(device)
- pred = model(x_batch)
- loss = loss_fn(pred, y_batch)
- optimizer.zero_grad()
- loss.backward()
- optimizer.step()
- train_loss += loss.item() # 记录每个批次的损失值
- # 计算并打印损失值
- train_loss /= train_num
- accuracy = (pred.argmax(1) == y_batch).type(torch.float32).sum().item() / batch_size
- print("epoch:",epoch,"train_loss:", round(train_loss,2),"accuracy:",round(accuracy,2))
在这里,我们使用了本章新定义的卷积神经网络模块作为局部特征抽取,而对于其他的损失函数以及优化函数,只使用了与前期一样的模式进行模型训练。最终结果如下所示,请读者自行验证。
- (60000, 1, 28, 28)
- epoch: 0 train_loss: 2.3 accuracy: 0.11
- epoch: 1 train_loss: 2.3 accuracy: 0.13
- epoch: 2 train_loss: 2.3 accuracy: 0.2
- epoch: 3 train_loss: 2.3 accuracy: 0.18
- …
- epoch: 58 train_loss: 0.5 accuracy: 0.98
- epoch: 59 train_loss: 0.49 accuracy: 0.98
- epoch: 60 train_loss: 0.49 accuracy: 0.98
- epoch: 61 train_loss: 0.48 accuracy: 0.98
- epoch: 62 train_loss: 0.48 accuracy: 0.98
-
- Process finished with exit code 0
本文节选自《PyTorch 2.0深度学习从零开始学》,本书实战案例丰富,可带领读者快速掌握深度学习算法及其常见案例。