![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/07\/20260727162445-6a67864d27447.png\")
<\/p>\n\u524d\u9762\u51e0\u671f\u7684\u5185\u5bb9，\u6211\u4eec\u628a\u795e\u7ecf\u7f51\u7edc\u5927\u90e8\u5206\u77e5\u8bc6\u70b9\u5b66\u4e60\u4e86，\u8fd9\u4e00\u8bb2\u6211\u4eec\u5c06\u7528\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u4f5c\u4e3a\u4f8b\u5b50，\u5e26\u7740\u5927\u5bb6\u5206\u6790\u4ee3\u7801\u90e8\u5206\u8fd8\u6709\u5269\u4e0b\u7684\u539f\u7406\u90e8\u5206，\u6211\u76f8\u4fe1\u8fd9\u4e00\u8bb2\u4f1a\u5f88\u6709\u8da3！\u6211\u4eec\u91c7\u7528\u7684\u6570\u636e\u96c6\u662fMNIST\u624b\u5199\u6570\u636e\u96c6；\u5982\u679c\u524d\u9762\u51e0\u671f\u7684\u5185\u5bb9\u6ca1\u6709\u770b\u7684，\u53ef\u4ee5\u70b9\u51fb\u4e0b\u9762\u7684\u94fe\u63a5！\u89c9\u5f97\u5185\u5bb9\u6709\u5e2e\u52a9\u7684，\u5e0c\u671b\u53ef\u4ee5\u5173\u6ce8\u4e00\u4e0b\u535a\u4e3b！\u535a\u4e3b\u5c06\u6301\u7eed\u66f4\u65b0\u4e13\u680f\u00a0AI\u7b51\u57fa\u5f55\u00a0<\/p>\n
AI\u7b51\u57fa\u5f55\u2014\u2014\u635f\u5931\u51fd\u6570\u7bc7-CSDN\u535a\u5ba2<\/p>\n
AI\u7b51\u57fa\u5f55\u2014\u2014\u6fc0\u6d3b\u51fd\u6570\u7bc7-CSDN\u535a\u5ba2<\/p>\n
AI\u7b51\u57fa\u5f55\u2014\u2014\u68af\u5ea6\u4e0b\u964d\u7bc7-CSDN\u535a\u5ba2<\/p>\n
AI\u7b51\u57fa\u5f55\u2014\u2014\u5377\u79ef\u7bc7-CSDN\u535a\u5ba2<\/p>\n
\u5728\u6b63\u5f0f\u642d\u5efa\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u4e4b\u524d，\u6211\u4eec\u9996\u5148\u8981\u77e5\u9053，\u4e00\u4e2a\u795e\u7ecf\u7f51\u7edc\u7a76\u7adf\u53ef\u4ee5\u7528\u6765\u505a\u4ec0\u4e48？\u7b80\u5355\u7684\u8bf4\u795e\u7ecf\u7f51\u7edc\u7684\u4efb\u52a1\u5c31\u662f\u5b66\u4e60\u8f93\u5165\u548c\u8f93\u51fa\u4e4b\u95f4\u7684\u5173\u7cfb，\u6211\u4eec\u7ed9\u795e\u7ecf\u7f51\u7edc\u5927\u91cf\u5e26\u6709\u7b54\u6848\u7684\u6570\u636e，\u795e\u7ecf\u7f51\u7edc\u901a\u8fc7\u4e0d\u65ad\u8c03\u6574\u5185\u90e8\u53c2\u6570，\u9010\u6e10\u5b66\u4f1a\u5982\u4f55\u6839\u636e\u8f93\u5165\u5f97\u5230\u6b63\u786e\u8f93\u51fa；\u53ef\u4ee5\u7528\u4e8e\u5206\u7c7b\u4efb\u52a1，\u56de\u5f52\u4efb\u52a1，\u751f\u6210\u4efb\u52a1；<\/p>\n
\u7f51\u7edc\u6216\u8005\u6a21\u578b\u90fd\u662f\u901a\u8fc7\u8bad\u7ec3\u624d\u62e5\u6709\u80fd\u529b\u7684，\u6240\u4ee5\u6211\u4eec\u7684\u6574\u4f53\u7684\u6d41\u7a0b\u5206\u4e3a\u8bad\u7ec3\u548c\u6d4b\u8bd5；\u8fd9\u4e9b\u8bad\u7ec3\u6216\u8005\u6d4b\u8bd5\u90fd\u9700\u8981\u6570\u636e，\u90a3\u4e48\u8fd9\u6837\u7684\u201c\u517b\u5206\u201d\u6211\u4eec\u79f0\u4e3a\u6570\u636e\u96c6；\u6240\u4ee5\u6211\u4eec\u53ef\u4ee5\u770b\u51fa，\u8bad\u7ec3\u96c6\u5e94\u8be5\u662f\u5206\u4e3a\u6d4b\u8bd5\u96c6\u548c\u8bad\u7ec3\u96c6，\u5b9e\u9645\u4e0a\u8fd8\u53ef\u4ee5\u5206\u591a\u4e00\u4e2a\u9a8c\u8bc1\u96c6，\u540e\u9762\u6211\u4eec\u4f1a\u4e13\u95e8\u51fa\u4e00\u671f\u5185\u5bb9\u662f\u5173\u4e8e\u6570\u636e\u5982\u4f55\u642d\u5efa，\u5904\u7406\u7684\u6587\u7ae0，\u8fd9\u8fb9\u4e0d\u5c55\u5f00\u63cf\u8ff0<\/p>\n
\u6211\u4eec\u73b0\u5728\u4ecb\u7ecd\u4e00\u4e0b MNIST \u624b\u5199\u8bad\u7ec3\u96c6，7 \u4e07\u5f20 28\u00d728 \u50cf\u7d20\u7684\u624b\u5199\u6570\u5b57\u7070\u5ea6\u56fe，\u6bcf\u5f20\u56fe\u7247\u90fd\u6709\u5bf9\u5e94\u6807\u7b7e（\u6807\u7b7e\u4e5f\u5c31\u662f\u6570\u636e\u5bf9\u5e94\u7684\u6b63\u786e\u7b54\u6848），\u53ef\u8c13\u662f\u662f\u6df1\u5ea6\u5b66\u4e60\u9886\u57df\u7684 "Hello World"<\/p>\n
<\/p>\n
\u56fe\u7247\u5728\u8ba1\u7b97\u673a\u7684\u5f62\u72b6：[1,28,28]，\u5206\u522b\u662f\u901a\u9053\u6570\u548c\u957f\u5bbd，\u56e0\u4e3a\u8bad\u7ec3\u7684\u8fc7\u7a0b\u4e2d\u6211\u4eec\u662f\u6279\u6b21\u8f93\u5165，\u56e0\u6b64\u524d\u9762\u8fd8\u591a\u4e00\u4e2a B \u8fd9\u4e2a\u7ef4\u5ea6；<\/p>\n
\u914d\u7f6e\u73af\u5883\u662f\u4e00\u95e8\u57fa\u672c\u529f，\u4f46\u662f\u6709\u65f6\u5019\u786e\u5b9e\u4e5f\u662f\u975e\u5e38\u75db\u82e6；\u4f46\u662f\u8fd9\u4e2a\u5f88\u7b80\u5355\u914d\u7f6e，\u800c\u4e14\u6709\u4e86 AI \u4e4b\u540e\u57fa\u672c\u90fd\u662f\u5b83\u5e2e\u52a9\u6211\u4eec\u914d\u7f6e\u5373\u53ef；\u8fd9\u8fb9\u5efa\u8bae\u4e0b\u8f7d Anaconda，\u5e93\u7ba1\u7406\u5e73\u53f0；<\/p>\n
\u8fd9\u4e2a\u4efb\u52a1\u6211\u4eec\u53ea\u9700\u8981 PyTorch + torchvision，\u6ca1\u6709\u5176\u4ed6\u7b2c\u4e09\u65b9\u5e93<\/p>\n
\u5b89\u88c5 PyTorch：\u63a8\u8350\u7528 Conda \u521b\u5efa\u72ec\u7acb\u73af\u5883，\u907f\u514d\u5305\u51b2\u7a81<\/p>\n
# \u521b\u5efa\u73af\u5883（Python 3.10+）
\nconda create -n cnn_mnist python=3.11 -y
\nconda activate cnn_mnist<\/p>\n
# \u5b89\u88c5 PyTorch（GPU \u7248，CUDA 12.x）
\npip install torch torchvision –index-url https:\/\/download.pytorch.org\/whl\/cu124<\/p>\n
\u901a\u5e38\u6211\u4eec\u90fd\u662f\u5728 GPU \u4e0a\u8bad\u7ec3\u7684，\u8bad\u7ec3\u901f\u5ea6\u4f1a\u6bd4 CPU \u5feb；<\/p>\n
\u6570\u636e\u96c6\u90e8\u5206\u4e0d\u7528\u4e13\u95e8\u53bb\u5b98\u7f51\u4e0b\u8f7d，\u4e00\u884c\u4ee3\u7801\u5c31\u53ef\u4ee5\u76f4\u63a5\u641e\u5b9a，\u81ea\u52a8\u4e0b\u8f7d\u5230\u672c\u5730\u7f13\u5b58<\/p>\n
torchvision.datasets.MNIST <\/p>\n
\u5377\u79ef\u795e\u7ecf\u7f51\u7edc，\u5377\u79ef\u4e00\u5b9a\u662f\u5bf9\u5e94\u7684\u6838\u5fc3，\u6211\u4eec\u770b\u770b\u4ee3\u7801\u90e8\u5206(\u4ece\u4e0a\u5230\u4e0b\u5206\u6790，\u51e0\u4e4e\u9010\u884c\u5206\u6790)：<\/p>\n
class CNNNet(nn.Module):<\/p>\n
def __init__(self, num_classes=10):
\n super().__init__()<\/p>\n
self.conv1 = nn.Conv2d(1, 32, kernel_size=3, stride=1, padding=1)
\n self.bn1 = nn.BatchNorm2d(32)<\/p>\n
self.conv2 = nn.Conv2d(32, 64, kernel_size=3, stride=1, padding=1)
\n self.bn2 = nn.BatchNorm2d(64)<\/p>\n
self.conv3 = nn.Conv2d(64, 128, kernel_size=3, stride=1, padding=1)
\n self.bn3 = nn.BatchNorm2d(128)<\/p>\n
self.pool = nn.MaxPool2d(2, 2)<\/p>\n
self.dropout = nn.Dropout(0.3)<\/p>\n
# 28 \u2192 pool \u2192 14 \u2192 pool \u2192 7 \u2192 pool \u2192 3 (int)
\n # \u5b9e\u9645: 28\/2=14 \u2192 14\/2=7 \u2192 conv3 \u540e pool: 7\/2=3，\u6240\u4ee5 128 * 3 * 3 = 1152
\n self.fc1 = nn.Linear(128 * 3 * 3, 256)
\n self.fc2 = nn.Linear(256, num_classes)<\/p>\n
def forward(self, x):
\n x = self.pool(F.relu(self.bn1(self.conv1(x)))) # (B,32,14,14)
\n x = self.pool(F.relu(self.bn2(self.conv2(x)))) # (B,64,7,7)
\n x = self.pool(F.relu(self.bn3(self.conv3(x)))) # (B,128,3,3)<\/p>\n
x = x.view(x.size(0), -1) # flatten
\n x = F.relu(self.fc1(self.dropout(x)))
\n x = self.fc2(x)
\n return x <\/p>\n
\u6211\u4eec\u901a\u5e38\u662f\u91c7\u7528\u7c7b\u53bb\u5b9a\u4e49\u6a21\u5757，\u56e0\u4e3a\u5177\u6709\u7ee7\u627f\u7684\u529f\u80fd；\u50cf\u00a0nn.Module \u53ef\u4ee5\u7406\u89e3\u4e3a PyTorch \u63d0\u4f9b\u7684\u201c\u795e\u7ecf\u7f51\u7edc\u57fa\u7840\u6a21\u677f\u201d，\u6211\u4eec\u81ea\u5df1\u8bbe\u8ba1\u7f51\u7edc\u65f6，\u901a\u5e38\u90fd\u9700\u8981\u7ee7\u627f\u5b83；<\/p>\n
\u521d\u59cb\u5316\u51fd\u6570__init__ \u7528\u6765\u63d0\u524d\u5b9a\u4e49\u7f51\u7edc\u4e2d\u9700\u8981\u4f7f\u7528\u53d8\u91cf，\u56e0\u4e3a\u6211\u4eec\u6700\u540e\u662f\u505a\u5206\u7c7b，\u56e0\u6b64\u6211\u4eec\u8981\u6700\u540e\u7684\u7c7b\u522b\u6570，\u4e00\u5171\u6709 10 \u7c7b<\/p>\n
\u5377\u79ef\u5c42，\u53c2\u6570\u4ece\u5de6\u5230\u53f3\u5206\u522b\u8868\u793a\u8f93\u5165\u901a\u9053，\u8f93\u51fa\u901a\u9053，\u5377\u79ef\u6838，\u6b65\u957f，\u586b\u5145；\u6bd4\u5982\u7b2c\u4e00\u5c42\u5377\u79ef\u4e2d\u7684\u8f93\u51fa\u901a\u9053\u662f 32，\u8868\u793a\u7684\u662f\u6bcf\u4e2a\u5377\u79ef\u6838\u90fd\u53ef\u4ee5\u5c1d\u8bd5\u5bfb\u627e\u4e00\u79cd\u4e0d\u540c\u7684\u56fe\u50cf\u7279\u5f81，\u4f8b\u5982\u6a2a\u7ebf\u3001\u7ad6\u7ebf\u3001\u8fb9\u7f18\u3001\u62d0\u89d2\u7b49，32 \u4e2a\u5377\u79ef\u6838\u5904\u7406\u540e，\u5c31\u4f1a\u5f97\u5230 32 \u5f20\u7279\u5f81\u56fe；<\/p>\n
\u5f52\u4e00\u5316，\u6211\u4eec\u524d\u9762\u8bb2\u8fc7，\u8fd9\u4e2a\u5c31\u662f\u5bf9\u8f93\u51fa\u901a\u9053\u8fdb\u884c\u5f52\u4e00\u5316；\u539f\u56e0\u6211\u4eec\u89e3\u91ca\u4e00\u904d：\u5377\u79ef\u5c42\u63d0\u53d6\u51fa\u6765\u7684\u6570\u636e\u6709\u7684\u53ef\u80fd\u7279\u522b\u5927，\u6709\u7684\u53ef\u80fd\u7279\u522b\u5c0f，BatchNorm \u4f1a\u5bf9\u8fd9\u4e9b\u6570\u636e\u505a\u4e00\u6b21\u6574\u7406，\u907f\u514d\u6570\u636e\u5dee\u8ddd\u8fc7\u5927；<\/p>\n
\u6c60\u5316\u5c42，\u4fdd\u7559\u660e\u663e\u7279\u5f81，\u540c\u65f6\u7f29\u5c0f\u7279\u5f81\u56fe，\u51cf\u5c11\u540e\u7eed\u8ba1\u7b97\u91cf；\u8fd9\u8fb9\u91c7\u53d6\u7684\u7a97\u53e3\u5927\u5c0f\u662f 2 ；<\/p>\n
\u8fd9\u8fb9\u6709\u4e00\u4e2a\u70b9\u6ce8\u610f\u4e00\u4e0b，\u540e\u8005\u5377\u79ef\u5c42\u7684\u8f93\u5165\u4e00\u5b9a\u548c\u524d\u8005\u5377\u79ef\u5c42\u7684\u8f93\u51fa\u7684\u901a\u9053\u6570\u5927\u5c0f\u4e00\u81f4，\u5426\u5219\u4f1a\u62a5\u9519；<\/p>\n
\u8fd9\u4e2a\u524d\u9762\u6211\u4eec\u6ca1\u6709\u8bf4\u8fc7，\u8fd9\u4e2a\u662f\u795e\u7ecf\u5143\u968f\u673a\u4e22\u5f03；\u5982\u679c\u6bcf\u6b21\u8bad\u7ec3\u90fd\u4f9d\u8d56\u56fa\u5b9a\u7684\u51e0\u4e2a\u795e\u7ecf\u5143，\u6a21\u578b\u53ef\u80fd\u53ea\u662f\u8bb0\u4f4f\u4e86\u8bad\u7ec3\u6570\u636e，\u800c\u6ca1\u6709\u771f\u6b63\u5b66\u4f1a\u8bc6\u522b\u6570\u5b57\u3002Dropout \u4f1a\u968f\u673a\u5173\u95ed\u4e00\u90e8\u5206\u795e\u7ecf\u5143，\u8feb\u4f7f\u7f51\u7edc\u5b66\u4e60\u66f4\u52a0\u7a33\u5b9a\u3001\u66f4\u52a0\u901a\u7528\u7684\u7279\u5f81；\u50cf\u4ee3\u7801\u91cc\u8868\u793a\u6709 30% \u7684\u795e\u7ecf\u5143\u4f1a\u88ab\u4e22\u5f03；<\/p>\n
\u5168\u8fde\u63a5\u5c42\u8d1f\u8d23\u5c06\u5377\u79ef\u540e\u7684\u7279\u5f81\u7efc\u5408\u8d77\u6765，\u5c55\u5f00\u4e3a\u4e00\u884c，\u4f5c\u51fa\u6700\u7ec8\u5224\u65ad；<\/p>\n
self.fc2 = nn.Linear(256, num_classes) <\/p>\n
\u8fd9\u4e2a\u5c31\u662f\u628a\u5377\u79ef\u5230\u7684\u6700\u540e\u7279\u5f81\u76f4\u63a5\u8f93\u51fa\u6210 10 \u4e2alogit\u5f97\u5206，\u54ea\u4e2a\u6570\u5b57\u5bf9\u5e94\u7684\u5206\u6570\u6700\u9ad8，\u6a21\u578b\u5c31\u66f4\u503e\u5411\u4e8e\u8ba4\u4e3a\u56fe\u7247\u4e2d\u5199\u7684\u662f\u54ea\u4e2a\u6570\u5b57；<\/p>\n
\u524d\u5411\u4f20\u64ad forward\u00a0\u5b9a\u4e49\u4e86\u6570\u636e\u8fdb\u5165\u7f51\u7edc\u4ee5\u540e，\u5177\u4f53\u6309\u7167\u4ec0\u4e48\u987a\u5e8f\u5411\u524d\u4f20\u64ad；\u6bd4\u5982\u8fd9\u8fb9\u5c31\u662f\u6309\u7167\u5377\u79ef，\u5f52\u4e00，\u6c60\u5316\u4f5c\u4e3a\u4e00\u4e2a block \u91cd\u590d 3 \u6b21\u4e3a\u987a\u5e8f；\u6700\u540e\u5c55\u5f00(\u4e3a\u5168\u8fde\u63a5\u505a\u51c6\u5907)，\u6700\u540e\u7ecf\u8fc7\u5168\u8fde\u63a5\u5f97\u5230\u5206\u6570；<\/p>\n
\u8bad\u7ec3\u6a21\u5757\u4efb\u52a1\u5c31\u662f\u8ba9\u6a21\u578b\u628a\u6574\u4e2a\u8bad\u7ec3\u96c6\u5b66\u4e60\u4e00\u904d，\u8fd9\u8fb9\u6709\u4e00\u4e2a\u6982\u5ff5\u53eb\u505a EPOCH，\u5f53\u8fd9\u4e2a\u503c\u4e3a 10 \u7684\u65f6\u5019，\u8868\u793a\u7684\u662f\u8fd9\u4e2a\u8bad\u7ec3\u96c6\u88ab\u7f51\u7edc\u770b\u4e86 10 \u6b21；<\/p>\n
def train_one_epoch(model, loader, optimizer, criterion, device):
\n """
\n \u4f7f\u7528\u6574\u4e2a\u8bad\u7ec3\u96c6\u8bad\u7ec3\u6a21\u578b\u4e00\u8f6e\u3002<\/p>\n
\u8fd4\u56de：
\n \u5e73\u5747\u8bad\u7ec3\u635f\u5931
\n \u8bad\u7ec3\u96c6\u51c6\u786e\u7387
\n """<\/p>\n
# \u5207\u6362\u5230\u8bad\u7ec3\u6a21\u5f0f
\n # Dropout \u548c BatchNorm \u4f1a\u6309\u7167\u8bad\u7ec3\u72b6\u6001\u5de5\u4f5c
\n model.train()<\/p>\n
# \u8bb0\u5f55\u6574\u8f6e\u8bad\u7ec3\u7684\u603b\u635f\u5931\u3001\u6b63\u786e\u6570\u91cf\u548c\u6837\u672c\u603b\u6570
\n total_loss = 0
\n correct = 0
\n total = 0<\/p>\n
# \u6bcf\u6b21\u4ece DataLoader \u4e2d\u53d6\u51fa\u4e00\u6279\u56fe\u7247\u548c\u6807\u7b7e
\n for x, y in loader:<\/p>\n
# \u5c06\u56fe\u7247\u548c\u6807\u7b7e\u79fb\u52a8\u5230 CPU \u6216 GPU
\n x = x.to(device)
\n y = y.to(device)<\/p>\n
# \u6e05\u7a7a\u4e0a\u4e00\u6279\u6570\u636e\u8ba1\u7b97\u51fa\u7684\u68af\u5ea6
\n optimizer.zero_grad()<\/p>\n
# \u524d\u5411\u4f20\u64ad：\u6a21\u578b\u6839\u636e\u56fe\u7247\u8fdb\u884c\u9884\u6d4b
\n out = model(x)<\/p>\n
# \u6bd4\u8f83\u9884\u6d4b\u7ed3\u679c\u548c\u771f\u5b9e\u6807\u7b7e，\u8ba1\u7b97\u635f\u5931
\n loss = criterion(out, y)<\/p>\n
# \u53cd\u5411\u4f20\u64ad：\u8ba1\u7b97\u6bcf\u4e2a\u53c2\u6570\u7684\u68af\u5ea6
\n loss.backward()<\/p>\n
# \u6839\u636e\u68af\u5ea6\u66f4\u65b0\u6a21\u578b\u53c2\u6570
\n optimizer.step()<\/p>\n
# \u5f53\u524d loss \u662f\u8fd9\u4e00\u6279\u56fe\u7247\u7684\u5e73\u5747\u635f\u5931
\n # \u4e58\u4ee5\u6279\u6b21\u5927\u5c0f，\u5f97\u5230\u8fd9\u4e00\u6279\u56fe\u7247\u7684\u635f\u5931\u603b\u548c
\n total_loss += loss.item() * x.size(0)<\/p>\n
# \u627e\u5230\u6bcf\u5f20\u56fe\u7247\u5206\u6570\u6700\u9ad8\u7684\u7c7b\u522b，\u5e76\u7edf\u8ba1\u9884\u6d4b\u6b63\u786e\u6570\u91cf
\n predictions = out.argmax(dim=1)
\n correct += (predictions == y).sum().item()<\/p>\n
# \u7d2f\u52a0\u5df2\u7ecf\u5904\u7406\u8fc7\u7684\u56fe\u7247\u6570\u91cf
\n total += x.size(0)<\/p>\n
# \u8fd4\u56de\u6574\u4e2a\u8bad\u7ec3\u96c6\u7684\u5e73\u5747\u635f\u5931\u548c\u51c6\u786e\u7387
\n average_loss = total_loss \/ total
\n accuracy = 100.0 * correct \/ total<\/p>\n
return average_loss, accuracy <\/p>\n
\u51fd\u6570\u63a5\u53d7\u4e86 5 \u4e2a\u53c2\u6570，\u5206\u522b\u662f\u6a21\u578b，\u52a0\u8f7d\u5668，\u4f18\u5316\u5668，\u635f\u5931\u51fd\u6570，\u8bbe\u5907；<\/p>\n
\u4e4b\u6240\u4ee5\u9700\u8981\u4e13\u95e8\u8bbe\u7f6e\u8bad\u7ec3\u6a21\u5f0f，\u662f\u56e0\u4e3a\u6a21\u578b\u4e2d\u6709\u4e00\u4e9b\u5c42\u5728\u8bad\u7ec3\u548c\u6d4b\u8bd5\u65f6\u8868\u73b0\u4e0d\u540c；\u4ee3\u7801\u7684\u6ce8\u91ca\u5199\u7684\u7b97\u662f\u6bd4\u8f83\u76f8\u4fe1\u7684，\u56e0\u6b64\u6211\u5c31\u7b80\u5355\u7684\u8bf4\u4e00\u4e0b：<\/p>\n
\u5faa\u73af\u7684 x\u3001y \u8868\u793a\u5c31\u662f\u56fe\u7247\u548c\u6807\u7b7e，\u6211\u4eec\u653e\u5230 GPU \u4e2d\u8fdb\u884c\u52a0\u8f7d，\u63a5\u7740\u5c31\u662f\u628a\u4e0a\u4e00\u8f6e\u7684\u68af\u5ea6\u6e05\u96f6，\u63a5\u7740\u5c31\u662f\u8fdb\u884c\u6a21\u578b\u9884\u6d4b，\u635f\u5931\u51fd\u6570\u5bf9\u6bd4\u540e\u8fdb\u884c\u53cd\u5411\u4f20\u64ad，\u6700\u540e\u7528\u4f18\u5316\u5668\u8ddf\u65b0\u53c2\u6570；\u7ed3\u5408\u6211\u4eec\u4e4b\u524d\u7684\u539f\u7406\u90e8\u5206，\u8fd9\u4e2a\u4ee3\u7801\u8fd8\u662f\u5f88\u6e05\u6670\u7406\u89e3\u7684；\u6700\u540e\u5c31\u662f\u5bf9\u7b80\u5355\u7684\u7ed3\u679c\u8fdb\u884c\u8ba1\u7b97；<\/p>\n
\u63a8\u7406\u6a21\u5757\u7684\u4efb\u52a1\u5c31\u662f\u770b\u6a21\u578b\u8bad\u7ec3\u7684\u6548\u679c\u600e\u4e48\u6837，\u4e00\u822c\u662f\u5728\u9a8c\u8bc1\u96c6\u6216\u8005\u662f\u6d4b\u8bd5\u96c6\u4e0a\u63a8\u7406；<\/p>\n
@torch.no_grad()
\ndef evaluate(model, loader, criterion, device):
\n """
\n \u5728\u9a8c\u8bc1\u96c6\u6216\u6d4b\u8bd5\u96c6\u4e0a\u8bc4\u4f30\u6a21\u578b\u3002<\/p>\n
\u8fd4\u56de：
\n \u5e73\u5747\u635f\u5931
\n \u5206\u7c7b\u51c6\u786e\u7387
\n """<\/p>\n
# \u5c06\u6a21\u578b\u5207\u6362\u5230\u8bc4\u4f30\u6a21\u5f0f
\n # Dropout \u505c\u6b62\u968f\u673a\u5173\u95ed\u795e\u7ecf\u5143
\n # BatchNorm \u4f7f\u7528\u8bad\u7ec3\u65f6\u4fdd\u5b58\u7684\u7edf\u8ba1\u4fe1\u606f
\n model.eval()<\/p>\n
# \u8bb0\u5f55\u603b\u635f\u5931\u3001\u9884\u6d4b\u6b63\u786e\u6570\u91cf\u548c\u6837\u672c\u603b\u6570
\n total_loss = 0
\n correct = 0
\n total = 0<\/p>\n
# \u6bcf\u6b21\u53d6\u51fa\u4e00\u6279\u56fe\u7247\u548c\u6807\u7b7e
\n for x, y in loader:<\/p>\n
# \u5c06\u6570\u636e\u79fb\u52a8\u5230 CPU \u6216 GPU
\n x = x.to(device)
\n y = y.to(device)<\/p>\n
# \u524d\u5411\u4f20\u64ad，\u5f97\u5230\u5341\u4e2a\u7c7b\u522b\u5206\u6570
\n out = model(x)<\/p>\n
# \u8ba1\u7b97\u5f53\u524d\u6279\u6b21\u7684\u635f\u5931
\n loss = criterion(out, y)<\/p>\n
# \u7d2f\u8ba1\u5f53\u524d\u6279\u6b21\u7684\u635f\u5931\u603b\u548c
\n total_loss += loss.item() * x.size(0)<\/p>\n
# \u627e\u5230\u6bcf\u5f20\u56fe\u7247\u5206\u6570\u6700\u9ad8\u7684\u7c7b\u522b
\n predictions = out.argmax(dim=1)<\/p>\n
# \u7edf\u8ba1\u9884\u6d4b\u6b63\u786e\u7684\u56fe\u7247\u6570\u91cf
\n correct += (predictions == y).sum().item()<\/p>\n
# \u7d2f\u8ba1\u5df2\u7ecf\u5904\u7406\u7684\u56fe\u7247\u6570\u91cf
\n total += x.size(0)<\/p>\n
# \u8ba1\u7b97\u6574\u4e2a\u6570\u636e\u96c6\u7684\u5e73\u5747\u635f\u5931
\n average_loss = total_loss \/ total<\/p>\n
# \u8ba1\u7b97\u6574\u4e2a\u6570\u636e\u96c6\u7684\u51c6\u786e\u7387
\n accuracy = 100.0 * correct \/ total<\/p>\n
return average_loss, accuracy <\/p>\n
\u51fd\u6570\u63a5\u53d7\u4e86 4\u00a0\u4e2a\u53c2\u6570，\u5206\u522b\u662f\u6a21\u578b，\u52a0\u8f7d\u5668，\u635f\u5931\u51fd\u6570，\u8bbe\u5907<\/p>\n
\u4e0e\u8bad\u7ec3\u8fc7\u7a0b\u4e0d\u540c，\u8bc4\u4f30\u9636\u6bb5\u53ea\u9700\u8981\u8fdb\u884c\u524d\u5411\u4f20\u64ad，\u4e0d\u9700\u8981\u8ba1\u7b97\u68af\u5ea6，\u4e5f\u4e0d\u9700\u8981\u66f4\u65b0\u6a21\u578b\u53c2\u6570\u3002\u56e0\u6b64，\u6211\u4eec\u4f7f\u7528 @torch.no_grad() \u5173\u95ed\u68af\u5ea6\u8bb0\u5f55，\u4ece\u800c\u51cf\u5c11\u8ba1\u7b97\u91cf\u548c\u663e\u5b58\u5360\u7528；\u5f53\u6240\u6709\u56fe\u7247\u6d4b\u8bd5\u5b8c\u6210\u540e，\u51fd\u6570\u4f1a\u8fd4\u56de\u6574\u4e2a\u6570\u636e\u96c6\u7684\u5e73\u5747\u635f\u5931\u548c\u51c6\u786e\u7387<\/p>\n
if __name__ == "__main__":
\n # \u53ea\u6709\u76f4\u63a5\u8fd0\u884c\u5f53\u524d\u6587\u4ef6\u65f6，\u4e0b\u9762\u7684\u4ee3\u7801\u624d\u4f1a\u6267\u884c<\/p>\n
print("-" * 50)
\n print("CNN \u624b\u5199\u6570\u5b57\u8bc6\u522b (MNIST) \u8bad\u7ec3\u4e0e\u6d4b\u8bd5")
\n print("-" * 50)<\/p>\n
# \u81ea\u52a8\u9009\u62e9 GPU \u6216 CPU
\n device = torch.device(
\n "cuda" if torch.cuda.is_available() else "cpu"
\n )
\n print(f"[Device] {device}")<\/p>\n
# ==================== 1. \u51c6\u5907\u6570\u636e ====================<\/p>\n
# \u5b9a\u4e49\u56fe\u7247\u9884\u5904\u7406\u65b9\u5f0f
\n transform = transforms.Compose([
\n # \u5c06\u56fe\u7247\u8f6c\u6362\u6210 PyTorch \u5f20\u91cf
\n transforms.ToTensor(),<\/p>\n
# \u5bf9 MNIST \u56fe\u7247\u8fdb\u884c\u6807\u51c6\u5316
\n transforms.Normalize((0.1307,), (0.3081,))
\n ])<\/p>\n
# \u52a0\u8f7d MNIST \u8bad\u7ec3\u96c6
\n train_ds = datasets.MNIST(
\n root="data\/mnist",
\n train=True,
\n download=True,
\n transform=transform
\n )<\/p>\n
# \u52a0\u8f7d MNIST \u6d4b\u8bd5\u96c6
\n test_ds = datasets.MNIST(
\n root="data\/mnist",
\n train=False,
\n download=True,
\n transform=transform
\n )<\/p>\n
# \u6bcf\u6b21\u53d6\u51fa 128 \u5f20\u8bad\u7ec3\u56fe\u7247
\n train_loader = DataLoader(
\n train_ds,
\n batch_size=128,
\n shuffle=True,
\n num_workers=2
\n )<\/p>\n
# \u6bcf\u6b21\u53d6\u51fa 128 \u5f20\u6d4b\u8bd5\u56fe\u7247
\n test_loader = DataLoader(
\n test_ds,
\n batch_size=128,
\n shuffle=False,
\n num_workers=2
\n )<\/p>\n
print(
\n f"[Data] Train: {len(train_ds)} \u00b7 "
\n f"Test: {len(test_ds)}"
\n )<\/p>\n
# ==================== 2. \u521b\u5efa\u6a21\u578b ====================<\/p>\n
# \u521b\u5efa 10 \u5206\u7c7b\u7684\u5377\u79ef\u795e\u7ecf\u7f51\u7edc，\u5e76\u79fb\u52a8\u5230 CPU \u6216 GPU
\n model = CNNNet(num_classes=10).to(device)<\/p>\n
# \u7edf\u8ba1\u6a21\u578b\u53c2\u6570\u91cf
\n total_params = (
\n sum(p.numel() for p in model.parameters()) \/ 1e6
\n )
\n print(f"[Model] \u53c2\u6570\u91cf: {total_params:.2f}M")<\/p>\n
# \u5b9a\u4e49\u4ea4\u53c9\u71b5\u635f\u5931\u51fd\u6570
\n criterion = nn.CrossEntropyLoss()<\/p>\n
# \u5b9a\u4e49 Adam \u4f18\u5316\u5668，\u5b66\u4e60\u7387\u4e3a 0.001
\n optimizer = torch.optim.Adam(
\n model.parameters(),
\n lr=0.001
\n )<\/p>\n
# ==================== 3. \u8bad\u7ec3\u6a21\u578b ====================<\/p>\n
# \u5c06\u6574\u4e2a\u8bad\u7ec3\u96c6\u5b66\u4e60 5 \u904d
\n epochs = 5<\/p>\n
for epoch in range(1, epochs + 1):
\n # \u4f7f\u7528\u8bad\u7ec3\u96c6\u8bad\u7ec3\u4e00\u8f6e
\n train_loss, train_acc = train_one_epoch(
\n model,
\n train_loader,
\n optimizer,
\n criterion,
\n device
\n )<\/p>\n
# \u4f7f\u7528\u6d4b\u8bd5\u96c6\u68c0\u67e5\u5f53\u524d\u6a21\u578b\u6548\u679c
\n test_loss, test_acc = evaluate(
\n model,
\n test_loader,
\n criterion,
\n device
\n )<\/p>\n
# \u6253\u5370\u5f53\u524d\u8f6e\u6b21\u7684\u8bad\u7ec3\u548c\u6d4b\u8bd5\u7ed3\u679c
\n print(
\n f"Epoch {epoch:2d} | "
\n f"Train Loss: {train_loss:.4f} "
\n f"Acc: {train_acc:.2f}% | "
\n f"Test Loss: {test_loss:.4f} "
\n f"Acc: {test_acc:.2f}%"
\n )<\/p>\n
# \u6253\u5370\u6700\u540e\u4e00\u8f6e\u7684\u6d4b\u8bd5\u51c6\u786e\u7387
\n print("-" * 50)
\n print(f"[Result] \u6700\u7ec8\u6d4b\u8bd5\u51c6\u786e\u7387: {test_acc:.2f}%")
\n print("-" * 50)<\/p>\n
# ==================== 4. \u5355\u5f20\u56fe\u7247\u9884\u6d4b ====================<\/p>\n
# \u5c06\u6a21\u578b\u5207\u6362\u5230\u8bc4\u4f30\u6a21\u5f0f
\n model.eval()<\/p>\n
# \u53d6\u51fa\u6d4b\u8bd5\u96c6\u4e2d\u7684\u7b2c\u4e00\u5f20\u56fe\u7247\u548c\u771f\u5b9e\u6807\u7b7e
\n sample, label = test_ds[0]<\/p>\n
# \u5355\u5f20\u9884\u6d4b\u4e0d\u9700\u8981\u8ba1\u7b97\u68af\u5ea6
\n with torch.no_grad():
\n # \u539f\u5f62\u72b6：[1, 28, 28]
\n # \u589e\u52a0\u6279\u6b21\u7ef4\u5ea6\u540e：[1, 1, 28, 28]
\n sample = sample.unsqueeze(0).to(device)<\/p>\n
# \u5f97\u5230\u6a21\u578b\u8f93\u51fa
\n out = model(sample)<\/p>\n
# \u627e\u5230\u5206\u6570\u6700\u9ad8\u7684\u6570\u5b57\u7c7b\u522b
\n pred = out.argmax(dim=1).item()<\/p>\n
# \u8f93\u51fa\u771f\u5b9e\u7b54\u6848\u548c\u6a21\u578b\u9884\u6d4b\u7ed3\u679c
\n result = "\u2713" if pred == label else "\u2717"<\/p>\n
print(
\n f"[Inference] \u7b2c\u4e00\u5f20\u56fe: "
\n f"\u771f\u5b9e={label}, \u9884\u6d4b={pred} {result}"
\n ) <\/p>\n
\u4ee3\u7801\u7684\u6ce8\u91ca\u8db3\u591f\u8be6\u7ec6，\u56e0\u6b64\u8fd9\u8fb9\u5c31\u8bf4\u51e0\u4e2a\u70b9：<\/p>\n
\u7b80\u5355\u7684\u8fd0\u884c\u4e86\u4e00\u4e0b，\u53ef\u4ee5\u770b\u770b\u6548\u679c<\/p>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/07\/20260727162445-6a67864dbf8d7.png\")
<\/p>\n
\u5b8c\u6574\u4ee3\u7801：<\/p>\n
import torch
\nimport torch.nn as nn
\nimport torch.nn.functional as F
\nfrom torch.utils.data import DataLoader
\nfrom torchvision import datasets, transforms<\/p>\n
class CNNNet(nn.Module):
\n """\u7b80\u5355\u7684 CNN，\u7528\u4e8e MNIST \u624b\u5199\u6570\u5b57\u5206\u7c7b（28×28 \u7070\u5ea6\u56fe \u2192 10 \u7c7b）"""
\n def __init__(self, num_classes=10):
\n super().__init__()<\/p>\n
self.conv1 = nn.Conv2d(1, 32, kernel_size=3, stride=1, padding=1)
\n self.bn1 = nn.BatchNorm2d(32)<\/p>\n
self.conv2 = nn.Conv2d(32, 64, kernel_size=3, stride=1, padding=1)
\n self.bn2 = nn.BatchNorm2d(64)<\/p>\n
self.conv3 = nn.Conv2d(64, 128, kernel_size=3, stride=1, padding=1)
\n self.bn3 = nn.BatchNorm2d(128)<\/p>\n
self.pool = nn.MaxPool2d(2, 2)<\/p>\n
self.dropout = nn.Dropout(0.3)<\/p>\n
# 28 \u2192 pool \u2192 14 \u2192 pool \u2192 7 \u2192 pool \u2192 3 (int)
\n # \u5b9e\u9645: 28\/2=14 \u2192 14\/2=7 \u2192 conv3 \u540e pool: 7\/2=3，\u6240\u4ee5 128 * 3 * 3 = 1152
\n self.fc1 = nn.Linear(128 * 3 * 3, 256)
\n self.fc2 = nn.Linear(256, num_classes)<\/p>\n
def forward(self, x):
\n x = self.pool(F.relu(self.bn1(self.conv1(x)))) # (B,32,14,14)
\n x = self.pool(F.relu(self.bn2(self.conv2(x)))) # (B,64,7,7)
\n x = self.pool(F.relu(self.bn3(self.conv3(x)))) # (B,128,3,3)<\/p>\n
x = x.view(x.size(0), -1) # flatten
\n x = F.relu(self.fc1(self.dropout(x)))
\n x = self.fc2(x)
\n return x<\/p>\n
def train_one_epoch(model, loader, optimizer, criterion, device):
\n model.train()
\n total_loss, correct, total = 0, 0, 0
\n for x, y in loader:
\n x, y = x.to(device), y.to(device)<\/p>\n
optimizer.zero_grad()
\n out = model(x)
\n loss = criterion(out, y)
\n loss.backward()
\n optimizer.step()<\/p>\n
total_loss += loss.item() * x.size(0)
\n correct += (out.argmax(1) == y).sum().item()
\n total += x.size(0)<\/p>\n
return total_loss \/ total, 100.0 * correct \/ total<\/p>\n
@torch.no_grad()
\ndef evaluate(model, loader, criterion, device):
\n model.eval()
\n total_loss, correct, total = 0, 0, 0
\n for x, y in loader:
\n x, y = x.to(device), y.to(device)
\n out = model(x)
\n loss = criterion(out, y)<\/p>\n
total_loss += loss.item() * x.size(0)
\n correct += (out.argmax(1) == y).sum().item()
\n total += x.size(0)<\/p>\n
return total_loss \/ total, 100.0 * correct \/ total<\/p>\n
if __name__ == "__main__":
\n print("-" * 50)
\n print("CNN \u624b\u5199\u6570\u5b57\u8bc6\u522b (MNIST) \u8bad\u7ec3\u4e0e\u6d4b\u8bd5")
\n print("-" * 50)<\/p>\n
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
\n print(f"[Device] {device}")<\/p>\n
# ———- \u6570\u636e ———-
\n transform = transforms.Compose([
\n transforms.ToTensor(),
\n transforms.Normalize((0.1307,), (0.3081,))
\n ])<\/p>\n
train_ds = datasets.MNIST("data\/mnist", train=True, download=True, transform=transform)
\n test_ds = datasets.MNIST("data\/mnist", train=False, download=True, transform=transform)<\/p>\n
train_loader = DataLoader(train_ds, batch_size=128, shuffle=True, num_workers=2)
\n test_loader = DataLoader(test_ds, batch_size=128, shuffle=False, num_workers=2)<\/p>\n
print(f"[Data] Train: {len(train_ds)} \u00b7 Test: {len(test_ds)}")<\/p>\n
# ———- \u6a21\u578b ———-
\n model = CNNNet(num_classes=10).to(device)
\n total_params = sum(p.numel() for p in model.parameters()) \/ 1e6
\n print(f"[Model] \u53c2\u6570\u91cf: {total_params:.2f}M")<\/p>\n
criterion = nn.CrossEntropyLoss()
\n optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)<\/p>\n
# ———- \u8bad\u7ec3 ———-
\n epochs = 5
\n for epoch in range(1, epochs + 1):
\n train_loss, train_acc = train_one_epoch(model, train_loader, optimizer, criterion, device)
\n test_loss, test_acc = evaluate(model, test_loader, criterion, device)
\n print(f"Epoch {epoch:2d} | "
\n f"Train Loss: {train_loss:.4f} Acc: {train_acc:.2f}% | "
\n f"Test Loss: {test_loss:.4f} Acc: {test_acc:.2f}%")<\/p>\n
print("-" * 50)
\n print(f"[Result] \u6700\u7ec8\u6d4b\u8bd5\u51c6\u786e\u7387: {test_acc:.2f}%")
\n print("-" * 50)<\/p>\n
# ———- \u5355\u5f20\u9884\u6d4b\u9a8c\u8bc1 ———-
\n model.eval()
\n sample, label = test_ds[0]
\n with torch.no_grad():
\n pred = model(sample.unsqueeze(0).to(device)).argmax(1).item()
\n print(f"[Inference] \u7b2c\u4e00\u5f20\u56fe: \u771f\u5b9e={label}, \u9884\u6d4b={pred} {'\u2713' if pred == label else '\u2717'}")<\/p>\n
\u56e0\u4e3a\u8fd9\u4e00\u5757\u7684\u4ee3\u7801\u6ce8\u91ca\u786e\u5b9e\u662f\u771f\u7684\u5f88\u8be6\u7ec6，\u5982\u679c\u6211\u518d\u7ee7\u7eed\u8d58\u8ff0\u786e\u5b9e\u663e\u5f97\u6bd4\u8f83\u65e0\u804a；\u6211\u4eec\u53ef\u4ee5\u770b\u5230\u4e00\u4e2a\u7b80\u5355\u7684\u795e\u7ecf\u7f51\u7edc\u7684\u8109\u7edc\u662f\u6bd4\u8f83\u7b80\u5355\u7684；\u5e0c\u671b\u5927\u5bb6\u53ef\u4ee5\u505a\u5230\u624b\u6413\u8fd9\u6837\u7684\u5377\u79ef\u7f51\u7edc，\u96be\u5ea6\u4e0d\u5927\u7684；\u6211\u4eec\u5728\u6700\u540e\u7ed3\u679c\u53ef\u4ee5\u53d1\u73b0\u5b9e\u9645\u4e0a\u8fd9\u6837\u7684\u4e00\u4e2a\u7b80\u5355\u7684\u7f51\u7edc\u53c2\u6570\u91cf\u8fd8\u662f\u5f88\u591a\u7684，\u6709\u5174\u8da3\u7684\u540c\u5b66\u53ef\u4ee5\u770b\u770b\u6211\u4e4b\u524d\u4e3b\u9875\u7684mapping net，\u8fd9\u7bc7\u6587\u7ae0\u5c31\u662f\u6709\u5c11\u91cf\u7684\u53c2\u6570\u53bb\u903c\u8fd1\u5927\u53c2\u6570\u7684\u6548\u679c；<\/p>\n
AI\u7b51\u57fa\u5f55\u00a0\u7684\u4e13\u680f\u4f1a\u6162\u6162\u66f4\u65b0\u5b8c，\u5e26\u5927\u5bb6\u6709\u4f53\u7cfb\u7684\u7406\u89e3\u795e\u7ecf\u7f51\u7edc\u7684\u7531\u6765！\u5236\u4f5c\u4e0d\u6613，\u559c\u6b22\u535a\u4e3b\u6587\u7ae0\u7684\u53ef\u4ee5\u70b9\u8d5e\u6536\u85cf\u5173\u6ce8，\u8fd9\u4e9b\u90fd\u662f\u6211\u6301\u7eed\u66f4\u65b0\u7684\u52a8\u529b！<\/p>\n","protected":false},"excerpt":{"rendered":"
\u524d\u8a00
\n\u524d\u9762\u51e0\u671f\u7684\u5185\u5bb9，\u6211\u4eec\u628a\u795e\u7ecf\u7f51\u7edc\u5927\u90e8\u5206\u77e5\u8bc6\u70b9\u5b66\u4e60\u4e86，\u8fd9\u4e00\u8bb2\u6211\u4eec\u5c06\u7528\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u4f5c\u4e3a\u4f8b\u5b50，\u5e26\u7740\u5927\u5bb6\u5206\u6790\u4ee3\u7801\u90e8\u5206\u8fd8\u6709\u5269\u4e0b\u7684\u539f\u7406\u90e8\u5206，\u6211\u76f8\u4fe1\u8fd9\u4e00\u8bb2\u4f1a\u5f88\u6709\u8da3！\u6211\u4eec\u91c7\u7528\u7684\u6570\u636e\u96c6\u662fMNIST\u624b\u5199\u6570\u636e\u96c6；\u5982\u679c\u524d\u9762\u51e0\u671f\u7684\u5185\u5bb9\u6ca1\u6709\u770b\u7684，\u53ef\u4ee5\u70b9\u51fb\u4e0b\u9762\u7684\u94fe\u63a5！\u89c9\u5f97\u5185\u5bb9\u6709\u5e2e\u52a9\u7684，\u5e0c\u671b\u53ef\u4ee5\u5173\u6ce8\u4e00\u4e0b\u535a\u4e3b！\u535a\u4e3b\u5c06\u6301\u7eed\u66f4\u65b0\u4e13\u680f\u00a0AI\u7b51\u57fa\u5f55\u00a0
\nAI\u7b51\u57fa\u5f55\u2014\u2014\u635f\u5931\u51fd<\/p>\n","protected":false},"author":2,"featured_media":85914,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[81,152,50,2863,86],"topic":[],"class_list":["post-85916","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-python","tag-pytorch","tag-50","tag-2863","tag-86"],"yoast_head":"\n