{"id":209,"date":"2025-06-30T18:03:23","date_gmt":"2025-06-30T10:03:23","guid":{"rendered":"https:\/\/www.idcyz.com\/blog\/?p=209"},"modified":"2025-06-30T18:03:23","modified_gmt":"2025-06-30T10:03:23","slug":"pytorch%e9%87%8a%e6%94%be%e6%98%be%e5%ad%98%e6%9c%89%e5%a4%9a%e5%b0%91%e6%9c%89%e6%95%88%e6%96%b9%e6%b3%95%e5%8f%af%e6%8f%90%e9%ab%98%e8%ae%ad%e7%bb%83%e6%95%88%e7%8e%87","status":"publish","type":"post","link":"https:\/\/www.idcyz.com\/blog\/209.html","title":{"rendered":"PyTorch\u91ca\u653e\u663e\u5b58\u6709\u591a\u5c11\u6709\u6548\u65b9\u6cd5\u53ef\u63d0\u9ad8\u8bad\u7ec3\u6548\u7387"},"content":{"rendered":"

\"PyTorch\u91ca\u653e\u663e\u5b58\u6709\u591a\u5c11\u6709\u6548\u65b9\u6cd5\u53ef\u63d0\u9ad8\u8bad\u7ec3\u6548\u7387\"<\/p>\n

1. PyTorch\u91ca\u653e\u663e\u5b58\u7684\u91cd\u8981\u6027<\/h3>\n<\/p>\n

\u91ca\u653e\u663e\u5b58\u662f\u6df1\u5ea6\u5b66\u4e60\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u4e00\u9879\u81f3\u5173\u91cd\u8981\u7684\u64cd\u4f5c\u3002\u5c24\u5176\u662f\u5728\u4f7f\u7528\u5927\u89c4\u6a21\u6a21\u578b\u548c\u6570\u636e\u96c6\u65f6\uff0c\u663e\u5b58\u7684\u7ba1\u7406\u76f4\u63a5\u5f71\u54cd\u5230\u8bad\u7ec3\u7684\u6548\u7387\u548c\u7a33\u5b9a\u6027\u3002PyTorch \u63d0\u4f9b\u591a\u79cd\u65b9\u6cd5\u6765\u91ca\u653e\u663e\u5b58\uff0c\u4ece\u800c\u5e2e\u52a9\u5f00\u53d1\u8005\u4f18\u5316\u5185\u5b58\u4f7f\u7528\uff0c\u907f\u514d\u663e\u5b58\u6ea2\u51fa\u7684\u95ee\u9898\u3002\u8fd9\u7bc7\u6587\u7ae0\u4f1a\u8be6\u7ec6\u4ecb\u7ecd\u6709\u6548\u7684\u663e\u5b58\u91ca\u653e\u65b9\u6cd5\uff0c\u5e76\u63d0\u4f9b\u76f8\u5e94\u7684\u4ee3\u7801\u793a\u4f8b\u3002<\/p>\n<\/p>\n

\n

2. \u4f7f\u7528 torch.cuda.empty_cache()<\/h3>\n<\/p>\n

\u8fd9\u662f PyTorch \u4e2d\u6700\u5e38\u7528\u7684\u65b9\u6cd5\u4e4b\u4e00\u3002<\/br><\/p>\n

\u8be5\u65b9\u6cd5\u7684\u4f5c\u7528\u662f\u91ca\u653e\u672a\u4f7f\u7528\u7684\u7f13\u5b58\u663e\u5b58\uff0c\u4f46\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u91ca\u653e\u7684\u663e\u5b58\u5e76\u4e0d\u4f1a\u9a6c\u4e0a\u8fd4\u56de\u7ed9\u64cd\u4f5c\u7cfb\u7edf\uff0c\u800c\u662f\u88ab PyTorch \u7528\u4f5c\u540e\u7eed\u7684\u64cd\u4f5c\u3002\u8fd9\u80fd\u5e2e\u52a9\u63d0\u9ad8\u663e\u5b58\u7684\u4f7f\u7528\u6548\u7387\uff0c\u51cf\u5c11\u9891\u7e41\u5206\u914d\u548c\u91ca\u653e\u663e\u5b58\u6240\u5e26\u6765\u7684\u6027\u80fd\u635f\u5931\u3002<\/br><\/p>\n

\u5177\u4f53\u4ee3\u7801\u5982\u4e0b\uff1a<\/p>\n<\/p>\n

import torch<\/p>\r\n
torch.cuda.empty_cache()<\/code><\/pre>\n<\/p>\n
\n
3. \u4f7f\u7528 with torch.no_grad() \u8bed\u53e5<\/h3>\n<\/p>\n
\u5728\u4e0d\u9700\u8981\u8ba1\u7b97\u68af\u5ea6\u7684\u573a\u5408\uff0c\u4f7f\u7528 with torch.no_grad() \u53ef\u4ee5\u663e\u8457\u51cf\u5c11 GPU \u663e\u5b58\u7684\u5360\u7528\u3002\u6bcf\u6b21\u8ba1\u7b97\u56fe\u7684\u751f\u6210\u548c\u5b58\u50a8\u90fd\u4f1a\u6d88\u8017\u663e\u5b58\uff0c\u800c\u901a\u8fc7\u7981\u6b62\u68af\u5ea6\u8ba1\u7b97\uff0c\u53ef\u4ee5\u907f\u514d\u8fd9\u79cd\u6d88\u8017\uff0c\u7279\u522b\u662f\u5728\u8bc4\u4f30\u6a21\u578b\u6216\u8fdb\u884c\u63a8\u7406\u65f6\u3002\u793a\u4f8b\u4ee3\u7801\u5982\u4e0b\uff1a<\/p>\n<\/p>\n
with torch.no_grad():<\/p>\r\n
    output = model(input_data)<\/code><\/pre>\n<\/p>\n
\n
4. \u5220\u9664\u4e0d\u518d\u9700\u8981\u7684\u53d8\u91cf<\/h3>\n<\/p>\n
\u5728 PyTorch \u4e2d\uff0c\u53ea\u6709\u90a3\u4e9b\u88ab\u5f15\u7528\u7684\u53d8\u91cf\u4f1a\u5360\u7528\u663e\u5b58\u3002\u56e0\u6b64\uff0c\u5728\u6a21\u578b\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u67d0\u4e9b\u53d8\u91cf\u53ef\u4ee5\u88ab\u5220\u9664\uff0c\u4ece\u800c\u91ca\u653e\u663e\u5b58\u3002\u4f7f\u7528 Python \u7684 del \u8bed\u53e5\u53ef\u4ee5\u5220\u9664\u53d8\u91cf\uff0c\u4ee5\u51cf\u5c11 GPU \u5185\u5b58\u4f7f\u7528\u3002\u793a\u4f8b\u4ee3\u7801\u5982\u4e0b\uff1a<\/p>\n<\/p>\n
del var<\/p>\r\n
torch.cuda.empty_cache()<\/code><\/pre>\n<\/p>\n
\n
5. \u68af\u5ea6\u7d2f\u79ef<\/h3>\n<\/p>\n
\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u6709\u65f6\u6211\u4eec\u9700\u8981\u5904\u7406\u7684\u6279\u91cf\u6570\u636e\u8fc7\u5927\uff0c\u65e0\u6cd5\u5728\u4e00\u8f6e\u4e2d\u5168\u90e8\u4f20\u9012\u7ed9\u6a21\u578b\u3002\u901a\u8fc7\u68af\u5ea6\u7d2f\u79ef\uff0c\u53ef\u4ee5\u5206\u591a\u4e2a\u5c0f\u6279\u91cf\u6765\u8ba1\u7b97\u68af\u5ea6\uff0c\u6700\u7ec8\u66f4\u65b0\u4e00\u6b21\u53c2\u6570\uff0c\u8fd9\u6837\u53ef\u4ee5\u6781\u5927\u5730\u51cf\u5c11\u663e\u5b58\u7684\u5360\u7528\u3002\u5b9e\u73b0\u65b9\u5f0f\u5982\u4e0b\uff1a<\/p>\n<\/p>\n
for i in range(num_steps):<\/p>\r\n
    outputs = model(input_data[i])<\/p>\r\n
    loss = criterion(outputs, target[i])<\/p>\r\n
    loss.backward()<\/p>\r\n
    if (i+1) % accumulation_steps == 0:<\/p>\r\n
        optimizer.step()<\/p>\r\n
        optimizer.zero_grad()<\/code><\/pre>\n<\/p>\n
\n
6. \u5728\u5faa\u73af\u4e2d\u91cd\u7528\u6a21\u578b<\/h3>\n<\/p>\n
\u8bad\u7ec3\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u7684\u8fc7\u7a0b\u4e2d\uff0c\u5c24\u5176\u662f\u5728\u591a\u6b21\u8fed\u4ee3\u7684\u8bad\u7ec3\u4e2d\uff0c\u5982\u679c\u80fd\u591f\u91cd\u7528\u6a21\u578b\u800c\u4e0d\u91cd\u65b0\u5b9e\u4f8b\u5316\uff0c\u5c06\u6709\u6548\u51cf\u5c11\u663e\u5b58\u5360\u7528\u3002\u91cd\u7528\u6a21\u578b\u7684\u65b9\u6cd5\u662f\u5c06\u6a21\u578b\u4fdd\u5b58\u5728 GPU \u4e2d\uff0c\u5e76\u5728\u6bcf\u6b21\u8fed\u4ee3\u65f6\u5bf9\u5176\u8fdb\u884c\u66f4\u65b0\uff0c\u800c\u4e0d\u662f\u91cd\u65b0\u521d\u59cb\u5316\u3002\u5b9e\u73b0\u4ee3\u7801\u793a\u4f8b\u5982\u4e0b\uff1a<\/p>\n<\/p>\n
for epoch in range(num_epochs):<\/p>\r\n
    model.train()<\/p>\r\n
    model(data)  # \u91cd\u7528\u6a21\u578b\u800c\u4e0d\u662f\u91cd\u65b0\u5b9e\u4f8b\u5316<\/code><\/pre>\n<\/p>\n
\n
7. \u4f7f\u7528 mixed precision training<\/h3>\n<\/p>\n
\u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\u901a\u8fc7\u5728\u8ba1\u7b97\u4e2d\u7ed3\u5408\u4f7f\u7528 16 \u4f4d\u548c 32 \u4f4d\u6d6e\u70b9\u6570\uff0c\u53ef\u4ee5\u663e\u8457\u51cf\u5c11\u663e\u5b58\u4f7f\u7528\u5e76\u63d0\u9ad8\u8ba1\u7b97\u901f\u5ea6\u3002PyTorch \u63d0\u4f9b\u4e86\u539f\u751f\u652f\u6301\u7684\u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\uff0c\u4f7f\u7528 torchvision \u548c Apex \u53ef\u4ee5\u5f88\u65b9\u4fbf\u5730\u5b9e\u73b0\u8fd9\u79cd\u8bad\u7ec3\u65b9\u5f0f\u3002\u793a\u4f8b\u4ee3\u7801\u5982\u4e0b\uff1a<\/p>\n<\/p>\n
from torch.cuda.amp import autocast, GradScaler<\/p>\r\n\r\n
scaler = GradScaler()<\/p>\r\n
for data, target in train_loader:<\/p>\r\n
    optimizer.zero_grad()<\/p>\r\n
    with autocast():<\/p>\r\n
        output = model(data)<\/p>\r\n
        loss = criterion(output, target)<\/p>\r\n
    scaler.scale(loss).backward()<\/p>\r\n
    scaler.step(optimizer)<\/p>\r\n
    scaler.update()<\/code><\/pre>\n<\/p>\n
\n
8. \u52a8\u6001\u8ba1\u7b97\u56fe\u7684\u7279\u70b9<\/h3>\n<\/p>\n
PyTorch \u7684\u52a8\u6001\u8ba1\u7b97\u56fe\u4f7f\u5f97\u6211\u4eec\u53ef\u4ee5\u7075\u6d3b\u5730\u6784\u5efa\u6a21\u578b\u7ed3\u6784\uff0c\u56e0\u6b64\u5728\u4e00\u4e9b\u60c5\u51b5\u4e0b\uff0c\u5e94\u5f53\u6839\u636e\u9700\u8981\u8c03\u6574\u6a21\u578b\u7ed3\u6784\u6216\u8bad\u7ec3\u6d41\u7a0b\uff0c\u4ee5\u786e\u4fdd\u663e\u5b58\u7684\u4f7f\u7528\u662f\u9ad8\u6548\u7684\u3002\u4f8b\u5982\uff0c\u907f\u514d\u5728\u6bcf\u4e2a\u8fed\u4ee3\u4e2d\u521b\u5efa\u65b0\u7684\u5f20\u91cf\uff0c\u800c\u662f\u5c3d\u91cf\u91cd\u590d\u4f7f\u7528\u5404\u7c7b\u53d8\u91cf\u3002<\/p>\n<\/p>\n
\n
9. \u5982\u4f55\u5728 PyTorch \u4e2d\u5224\u65ad\u663e\u5b58\u7684\u4f7f\u7528\u60c5\u51b5\uff1f<\/h3>\n<\/p>\n
\u53ef\u4ee5\u901a\u8fc7\u54ea\u4e9b\u65b9\u6cd5\u67e5\u770b\u5f53\u524dGPU\u7684\u663e\u5b58\u4f7f\u7528\u60c5\u51b5\uff1f<\/b><\/br><\/p>\n
PyTorch \u63d0\u4f9b\u4e86 torch.cuda.memory_allocated() \u548c torch.cuda.memory_reserved() \u65b9\u6cd5\uff0c\u53ef\u4ee5\u5206\u522b\u7528\u4e8e\u67e5\u8be2\u5f53\u524d\u6a21\u578b\u7684\u663e\u5b58\u5206\u914d\u60c5\u51b5\u548c\u4fdd\u7559\u7684\u7f13\u5b58\u663e\u5b58\u3002\u793a\u4f8b\u4ee3\u7801\u5982\u4e0b\uff1a<\/p>\n<\/p>\n
allocated = torch.cuda.memory_allocated()<\/p>\r\n
reserved = torch.cuda.memory_reserved()<\/p>\r\n
print(f'Allocated: {allocated}, Reserved: {reserved}')<\/code><\/pre>\n<\/p>\n
\n
10. \u663e\u5b58\u5360\u7528\u7387\u8f83\u9ad8\u7684\u539f\u56e0\u662f\u4ec0\u4e48\uff1f<\/h3>\n<\/p>\n
\u663e\u5b58\u5360\u7528\u9ad8\u7684\u5e38\u89c1\u539f\u56e0\u6709\u54ea\u4e9b\uff1f<\/b><\/br><\/p>\n
\u5e38\u89c1\u7684\u539f\u56e0\u5305\u62ec\u8fc7\u5927\u7684\u6a21\u578b\u53c2\u6570\u3001\u672a\u91ca\u653e\u7684\u5f20\u91cf\u548c\u4e0d\u4f7f\u7528\u7684\u7f13\u5b58\u3002\u6a21\u578b\u4e2d\u7684\u67d0\u4e9b\u5c42\uff0c\u5982\u5168\u8fde\u63a5\u5c42\uff0c\u901a\u5e38\u6d88\u8017\u66f4\u591a\u663e\u5b58\u3002\u6b64\u5916\uff0c\u4f7f\u7528\u8f83\u5927\u7684\u6279\u91cf\u5927\u5c0f\u4e5f\u4f1a\u5bfc\u81f4\u663e\u5b58\u5feb\u901f\u6d88\u8017\u3002<\/p>\n<\/p>\n
\n
11. \u5982\u679c\u9047\u5230\u663e\u5b58\u4e0d\u8db3\u600e\u4e48\u529e\uff1f<\/h3>\n<\/p>\n
\u5982\u4f55\u5e94\u5bf9\u663e\u5b58\u4e0d\u8db3\u7684\u60c5\u51b5\uff1f<\/b><\/br><\/p>\n
\u5e94\u5bf9\u663e\u5b58\u4e0d\u8db3\u7684\u4e00\u4e9b\u65b9\u6cd5\u5305\u62ec\uff1a\u8c03\u5c0f\u6279\u91cf\u5927\u5c0f\uff0c\u51cf\u5c11\u6a21\u578b\u7684\u590d\u6742\u6027\uff0c\u8fdb\u884c\u6a21\u578b\u526a\u679d\uff0c\u6216\u8005\u4f7f\u7528\u66f4\u9ad8\u6548\u7684\u6570\u636e\u52a0\u8f7d\u65b9\u6cd5\u3002\u8fd8\u53ef\u4ee5\u901a\u8fc7\u5728\u8bad\u7ec3\u65f6\u4f18\u5316\u6570\u636e\u96c6\u6216\u8c03\u6574\u8d85\u53c2\u6570\u6765\u6539\u5584\u663e\u5b58\u5229\u7528\u7387\u3002<\/p><\/p>\n","protected":false},"excerpt":{"rendered":"
1. PyTorch\u91ca\u653e\u663e\u5b58\u7684\u91cd\u8981\u6027 \u91ca\u653e\u663e\u5b58\u662f\u6df1\u5ea6\u5b66\u4e60\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u4e00\u9879\u81f3\u5173\u91cd\u8981\u7684\u64cd\u4f5c\u3002\u5c24\u5176\u662f\u5728\u4f7f\u7528\u5927\u89c4\u6a21\u6a21\u578b\u548c\u6570\u636e\u96c6\u65f6\uff0c\u663e\u5b58\u7684\u7ba1\u7406\u76f4\u63a5\u5f71\u54cd\u5230\u8bad\u7ec3\u7684\u6548\u7387\u548c\u7a33\u5b9a\u6027\u3002PyTorch \u63d0\u4f9b\u591a\u79cd\u65b9\u6cd5\u6765\u91ca\u653e\u663e\u5b58\uff0c\u4ece\u800c\u5e2e\u52a9\u5f00\u53d1\u8005\u4f18\u5316\u5185\u5b58\u4f7f\u7528\uff0c\u907f\u514d\u663e\u5b58\u6ea2\u51fa…<\/p>\n","protected":false},"author":1,"featured_media":210,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[120,119,121],"topic":[],"class_list":["post-209","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-jiushu","tag-pytorch","tag-119","tag-121"],"_links":{"self":[{"href":"https:\/\/www.idcyz.com\/blog\/wp-json\/wp\/v2\/posts\/209","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=209"}],"version-history":[{"count":1,"href":"https:\/\/www.idcyz.com\/blog\/wp-json\/wp\/v2\/posts\/209\/revisions"}],"predecessor-version":[{"id":212,"href":"https:\/\/www.idcyz.com\/blog\/wp-json\/wp\/v2\/posts\/209\/revisions\/212"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.idcyz.com\/blog\/wp-json\/wp\/v2\/media\/210"}],"wp:attachment":[{"href":"https:\/\/www.idcyz.com\/blog\/wp-json\/wp\/v2\/media?parent=209"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.idcyz.com\/blog\/wp-json\/wp\/v2\/categories?post=209"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.idcyz.com\/blog\/wp-json\/wp\/v2\/tags?post=209"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.idcyz.com\/blog\/wp-json\/wp\/v2\/topic?post=209"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}