{"id":631,"date":"2025-12-15T06:52:20","date_gmt":"2025-12-14T22:52:20","guid":{"rendered":"https:\/\/www.idcyz.com\/blog\/?p=631"},"modified":"2025-12-15T06:52:20","modified_gmt":"2025-12-14T22:52:20","slug":"pytorch%e4%b8%ad%e6%9c%89%e6%95%88%e9%98%b2%e6%ad%a2%e8%bf%87%e6%8b%9f%e5%90%88%e7%9a%84%e5%a4%9a%e7%a7%8d%e6%8a%80%e6%9c%af%e4%b8%8e%e5%ae%9e%e8%b7%b5","status":"publish","type":"post","link":"https:\/\/www.idcyz.com\/blog\/631.html","title":{"rendered":"PyTorch\u4e2d\u6709\u6548\u9632\u6b62\u8fc7\u62df\u5408\u7684\u591a\u79cd\u6280\u672f\u4e0e\u5b9e\u8df5"},"content":{"rendered":"

PyTorch \u662f\u4e00\u4e2a\u5f3a\u5927\u7684\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\uff0c\u4f46\u5728\u8bad\u7ec3\u6a21\u578b\u65f6\uff0c\u8fc7\u62df\u5408\u662f\u4e00\u4e2a\u5e38\u89c1\u7684\u95ee\u9898\u3002\u8fc7\u62df\u5408\u7684\u60c5\u51b5\u5f80\u5f80\u51fa\u73b0\u5728\u6a21\u578b\u5728\u8bad\u7ec3\u96c6\u4e0a\u8868\u73b0\u826f\u597d\uff0c\u800c\u5728\u9a8c\u8bc1\u96c6\u6216\u6d4b\u8bd5\u96c6\u4e0a\u8868\u73b0\u4e0d\u4f73\u3002\u4e3a\u4e86\u9632\u6b62\u8fc7\u62df\u5408\uff0c\u6211\u4eec\u53ef\u4ee5\u91c7\u7528\u591a\u79cd\u65b9\u6cd5\u3002\u5728\u8fd9\u7bc7\u6587\u7ae0\u4e2d\uff0c\u6211\u4eec\u5c06\u8ba8\u8bba\u4e00\u4e9b\u5e38\u7528\u7684\u6280\u672f\uff0c\u5177\u4f53\u5305\u62ec\uff1a\u6570\u636e\u589e\u5f3a\u3001\u6b63\u5219\u5316\u3001Dropout\u3001\u65e9\u505c\u6cd5\u548c\u6a21\u578b\u96c6\u6210\u3002<\/p>\n

\"PyTorch\u4e2d\u6709\u6548\u9632\u6b62\u8fc7\u62df\u5408\u7684\u591a\u79cd\u6280\u672f\u4e0e\u5b9e\u8df5\"<\/p>\n

1. \u6570\u636e\u589e\u5f3a<\/h3>\n<\/p>\n

\u6570\u636e\u589e\u5f3a\u662f\u4e00\u79cd\u901a\u8fc7\u5bf9\u8bad\u7ec3\u6570\u636e\u8fdb\u884c\u53d8\u6362\u6765\u589e\u52a0\u6570\u636e\u91cf\u7684\u65b9\u6cd5\u3002\u5e38\u89c1\u7684\u6570\u636e\u589e\u5f3a\u6280\u672f\u5305\u62ec\u65cb\u8f6c\u3001\u7f29\u653e\u3001\u7ffb\u8f6c\u3001\u88c1\u526a\u548c\u989c\u8272\u53d8\u5316\u7b49\u3002\u8fd9\u4e9b\u53d8\u6362\u6709\u52a9\u4e8e\u8ba9\u6a21\u578b\u5b66\u4e60\u5230\u66f4\u4e3a\u666e\u9002\u7684\u7279\u5f81\uff0c\u4ece\u800c\u63d0\u9ad8\u6a21\u578b\u7684\u6cdb\u5316\u80fd\u529b\u3002\u5728 PyTorch \u4e2d\uff0c\u53ef\u4ee5\u4f7f\u7528 torchvision \u5e93\u6765\u5b9e\u73b0\u6570\u636e\u589e\u5f3a\u3002<\/p>\n<\/p>\n

\u4ee3\u7801\u793a\u4f8b\uff1a<\/p>\n<\/p>\n

<\/p>\r\n
from torchvision import transforms<\/p>\r\n\r\n
transform = transforms.Compose([<\/p>\r\n
    transforms.RandomHorizontalFlip(),<\/p>\r\n
    transforms.RandomRotation(20),<\/p>\r\n
    transforms.ColorJitter(brightness=0.2, contrast=0.2),<\/p>\r\n
    transforms.ToTensor()<\/p>\r\n
])<\/p>\r\n
<\/code><\/pre>\n<\/p>\n
2. \u6b63\u5219\u5316<\/h3>\n<\/p>\n
\u6b63\u5219\u5316\u662f\u4e00\u79cd\u5728\u635f\u5931\u51fd\u6570\u4e2d\u589e\u52a0\u989d\u5916\u9879\u7684\u65b9\u6cd5\uff0c\u4ee5\u9632\u6b62\u6a21\u578b\u8fc7\u5ea6\u62df\u5408\u8bad\u7ec3\u6570\u636e\u3002\u5e38\u7528\u7684\u6b63\u5219\u5316\u65b9\u6cd5\u5305\u62ec L1 \u6b63\u5219\u5316\u548c L2 \u6b63\u5219\u5316\u3002L2 \u6b63\u5219\u5316\u901a\u8fc7\u60e9\u7f5a\u53c2\u6570\u7684\u5e73\u65b9\u548c\u6765\u4f7f\u6a21\u578b\u66f4\u52a0\u5e73\u6ed1\uff0c\u800c L1 \u6b63\u5219\u5316\u5219\u901a\u8fc7\u60e9\u7f5a\u53c2\u6570\u7684\u7edd\u5bf9\u503c\u548c\u6765\u63d0\u4f9b\u4e00\u5b9a\u7684\u7279\u5f81\u9009\u62e9\u3002<\/p>\n<\/p>\n
\u4ee3\u7801\u793a\u4f8b\uff1a<\/p>\n<\/p>\n
<\/p>\r\n
import torch<\/p>\r\n\r\n
# L2 \u6b63\u5219\u5316\u7684\u4f7f\u7528<\/p>\r\n
optimizer = torch.optim.Adam(model.parameters(), weight_decay=0.01)  # weight_decay \u4e3a L2 \u6b63\u5219\u5316\u7684\u8d85\u53c2\u6570<\/p>\r\n
<\/code><\/pre>\n<\/p>\n
3. Dropout<\/h3>\n<\/p>\n
Dropout \u662f\u4e00\u79cd\u7b80\u5355\u800c\u6709\u6548\u7684\u9632\u6b62\u8fc7\u62df\u5408\u7684\u65b9\u6cd5\u3002\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u5b83\u4f1a\u968f\u673a\u5c06\u4e00\u90e8\u5206\u795e\u7ecf\u5143\u7684\u8f93\u51fa\u8bbe\u4e3a\u96f6\uff0c\u4ece\u800c\u907f\u514d\u6a21\u578b\u5bf9\u67d0\u4e9b\u7279\u5f81\u8fc7\u5ea6\u4f9d\u8d56\uff0c\u4fc3\u4f7f\u7f51\u7edc\u5b66\u4e60\u66f4\u4e3a\u9c81\u68d2\u7684\u7279\u5f81\u3002\u901a\u5e38\u5728\u5168\u8fde\u63a5\u5c42\u4e4b\u524d\u6dfb\u52a0 Dropout \u5c42\uff0c\u53ef\u4ee5\u5f88\u597d\u5730\u51cf\u8f7b\u8fc7\u62df\u5408\u3002<\/p>\n<\/p>\n
\u4ee3\u7801\u793a\u4f8b\uff1a<\/p>\n<\/p>\n
<\/p>\r\n
import torch.nn as nn<\/p>\r\n\r\n
model = nn.Sequential(<\/p>\r\n
    nn.Linear(128, 64),<\/p>\r\n
    nn.ReLU(),<\/p>\r\n
    nn.Dropout(0.5),  # \u5728\u6b64\u4f7f\u7528 Dropout<\/p>\r\n
    nn.Linear(64, 10)<\/p>\r\n
)<\/p>\r\n
<\/code><\/pre>\n<\/p>\n
4. \u65e9\u505c\u6cd5<\/h3>\n<\/p>\n
\u65e9\u505c\u6cd5\u662f\u901a\u8fc7\u76d1\u63a7\u6a21\u578b\u5728\u9a8c\u8bc1\u96c6\u4e0a\u7684\u6027\u80fd\u6765\u51b3\u5b9a\u4f55\u65f6\u505c\u6b62\u8bad\u7ec3\u3002\u5982\u679c\u5728\u591a\u4e2a\u8bad\u7ec3\u5468\u671f\u5185\u9a8c\u8bc1\u635f\u5931\u6ca1\u6709\u4e0b\u964d\uff0c\u5219\u53ef\u4ee5\u505c\u6b62\u8bad\u7ec3\u3002\u8fd9\u79cd\u65b9\u5f0f\u53ef\u4ee5\u9632\u6b62\u6a21\u578b\u5728\u8bad\u7ec3\u96c6\u4e0a\u8fdb\u884c\u8fc7\u591a\u7684\u8fed\u4ee3\uff0c\u4ece\u800c\u5bfc\u81f4\u8fc7\u62df\u5408\u3002<\/p>\n<\/p>\n
\u4ee3\u7801\u793a\u4f8b\uff1a<\/p>\n<\/p>\n
<\/p>\r\n
best_val_loss = float('inf')<\/p>\r\n
patience = 0<\/p>\r\n\r\n
for epoch in range(num_epochs):<\/p>\r\n
    train_model(model, train_loader)<\/p>\r\n
    val_loss = validate_model(model, val_loader)<\/p>\r\n\r\n
    if val_loss < best_val_loss:<\/p>\r\n
        best_val_loss = val_loss<\/p>\r\n
        patience = 0  # \u91cd\u7f6e\u8010\u5fc3\u8ba1\u6570<\/p>\r\n
    else:<\/p>\r\n
        patience += 1<\/p>\r\n
        if patience > early_stopping_patience:<\/p>\r\n
            print(\"Early stopping\")<\/p>\r\n
            break<\/p>\r\n
<\/code><\/pre>\n<\/p>\n
5. \u6a21\u578b\u96c6\u6210<\/h3>\n<\/p>\n
\u6a21\u578b\u96c6\u6210\u662f\u901a\u8fc7\u7ed3\u5408\u591a\u4e2a\u6a21\u578b\u7684\u8f93\u51fa\uff0c\u6765\u6539\u8fdb\u9884\u6d4b\u6027\u80fd\u3002\u5404\u4e2a\u6a21\u578b\u7684\u9884\u6d4b\u7ed3\u679c\u53ef\u4ee5\u901a\u8fc7\u6295\u7968\u3001\u5e73\u5747\u6216\u5176\u4ed6\u7b56\u7565\u7ed3\u5408\u5728\u4e00\u8d77\uff0c\u4ece\u800c\u83b7\u5f97\u66f4\u5f3a\u7684\u6cdb\u5316\u80fd\u529b\u3002\u4f7f\u7528\u4e0d\u540c\u7684\u521d\u59cb\u5316\u3001\u8d85\u53c2\u6570\u6216\u6a21\u578b\u67b6\u6784\uff0c\u53ef\u4ee5\u589e\u52a0\u96c6\u6210\u6a21\u578b\u7684\u591a\u6837\u6027\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347\u7ed3\u679c\u3002<\/p>\n<\/p>\n
\u4ee3\u7801\u793a\u4f8b\uff1a<\/p>\n<\/p>\n
<\/p>\r\n
# \u5c06\u591a\u4e2a\u6a21\u578b\u7684\u8f93\u51fa\u8fdb\u884c\u6c42\u5e73\u5747<\/p>\r\n
output1 = model1(input_data)<\/p>\r\n
output2 = model2(input_data)<\/p>\r\n
ensemble_output = (output1 + output2) \/ 2<\/p>\r\n
<\/code><\/pre>\n<\/p>\n
\u95ee\u7b54\u73af\u8282<\/h3>\n<\/p>\n
\u4f7f\u7528\u6570\u636e\u589e\u5f3a\u4f1a\u6709\u4ec0\u4e48\u597d\u5904\uff1f<\/b><\/p>\n
\u6570\u636e\u589e\u5f3a\u53ef\u4ee5\u589e\u52a0\u8bad\u7ec3\u6570\u636e\u7684\u591a\u6837\u6027\uff0c\u4f7f\u6a21\u578b\u80fd\u591f\u5b66\u5230\u66f4\u4e3a\u666e\u9002\u7684\u7279\u5f81\uff0c\u964d\u4f4e\u8fc7\u62df\u5408\u7684\u98ce\u9669\u3002\u5728\u6d4b\u8bd5\u65f6\uff0c\u80fd\u591f\u66f4\u597d\u5730\u5e94\u5bf9\u672a\u89c1\u8fc7\u7684\u6837\u672c\uff0c\u63d0\u9ad8\u6a21\u578b\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u7684\u8868\u73b0\u3002<\/p>\n<\/p>\n
\u4e3a\u4f55\u6b63\u5219\u5316\u662f\u9884\u9632\u8fc7\u62df\u5408\u7684\u91cd\u8981\u65b9\u6cd5\uff1f<\/b><\/p>\n
\u6b63\u5219\u5316\u901a\u8fc7\u5728\u635f\u5931\u51fd\u6570\u4e2d\u52a0\u5165\u60e9\u7f5a\u9879\uff0c\u9f13\u52b1\u6a21\u578b\u5b66\u4e60\u8f83\u5c0f\u7684\u6743\u91cd\uff0c\u907f\u514d\u6a21\u578b\u8fc7\u4e8e\u590d\u6742\u3002\u8fd9\u6837\u53ef\u4ee5\u63d0\u9ad8\u6a21\u578b\u7684\u6cdb\u5316\u6027\u80fd\uff0c\u51cf\u5c11\u5bf9\u8bad\u7ec3\u96c6\u7279\u5f81\u7684\u4f9d\u8d56\uff0c\u4ece\u800c\u6709\u6548\u5730\u9632\u6b62\u8fc7\u62df\u5408\u3002<\/p>\n<\/p>\n
Dropout \u7684\u539f\u7406\u662f\u4ec0\u4e48\uff1f<\/b><\/p>\n
Dropout \u901a\u8fc7\u968f\u673a\u4e22\u5f03\u4e00\u90e8\u5206\u795e\u7ecf\u5143\u7684\u8f93\u51fa\uff0c\u4f7f\u5f97\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7f51\u7edc\u7684\u7ed3\u6784\u53d1\u751f\u53d8\u5316\uff0c\u4fc3\u4f7f\u7f51\u7edc\u5b66\u4e60\u5230\u66f4\u6709\u4ee3\u8868\u6027\u7684\u7279\u5f81\u3002\u8fd9\u6837\u53ef\u4ee5\u51cf\u5c11\u6a21\u578b\u5bf9\u67d0\u4e9b\u8282\u70b9\u7684\u4f9d\u8d56\uff0c\u8fdb\u800c\u63d0\u9ad8\u5176\u6cdb\u5316\u80fd\u529b\u3002<\/p><\/p>\n","protected":false},"excerpt":{"rendered":"
PyTorch \u662f\u4e00\u4e2a\u5f3a\u5927\u7684\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\uff0c\u4f46\u5728\u8bad\u7ec3\u6a21\u578b\u65f6\uff0c\u8fc7\u62df\u5408\u662f\u4e00\u4e2a\u5e38\u89c1\u7684\u95ee\u9898\u3002\u8fc7\u62df\u5408\u7684\u60c5\u51b5\u5f80\u5f80\u51fa\u73b0\u5728\u6a21\u578b\u5728\u8bad\u7ec3\u96c6\u4e0a\u8868\u73b0\u826f\u597d\uff0c\u800c\u5728\u9a8c\u8bc1\u96c6\u6216\u6d4b\u8bd5\u96c6\u4e0a\u8868\u73b0\u4e0d\u4f73\u3002\u4e3a\u4e86\u9632\u6b62\u8fc7\u62df\u5408\uff0c\u6211\u4eec\u53ef\u4ee5\u91c7\u7528\u591a\u79cd\u65b9\u6cd5\u3002\u5728\u8fd9\u7bc7\u6587\u7ae0\u4e2d\uff0c\u6211\u4eec\u5c06\u8ba8\u8bba\u4e00\u4e9b\u5e38\u7528\u7684\u6280\u672f\uff0c\u5177\u4f53…<\/p>\n","protected":false},"author":1,"featured_media":632,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[120,324,325],"topic":[],"class_list":["post-631","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-jiushu","tag-pytorch","tag-324","tag-325"],"_links":{"self":[{"href":"https:\/\/www.idcyz.com\/blog\/wp-json\/wp\/v2\/posts\/631","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.idcyz.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.idcyz.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.idcyz.com\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.idcyz.com\/blog\/wp-json\/wp\/v2\/comments?post=631"}],"version-history":[{"count":1,"href":"https:\/\/www.idcyz.com\/blog\/wp-json\/wp\/v2\/posts\/631\/revisions"}],"predecessor-version":[{"id":634,"href":"https:\/\/www.idcyz.com\/blog\/wp-json\/wp\/v2\/posts\/631\/revisions\/634"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.idcyz.com\/blog\/wp-json\/wp\/v2\/media\/632"}],"wp:attachment":[{"href":"https:\/\/www.idcyz.com\/blog\/wp-json\/wp\/v2\/media?parent=631"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.idcyz.com\/blog\/wp-json\/wp\/v2\/categories?post=631"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.idcyz.com\/blog\/wp-json\/wp\/v2\/tags?post=631"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.idcyz.com\/blog\/wp-json\/wp\/v2\/topic?post=631"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}