{"id":184,"date":"2025-06-26T14:17:54","date_gmt":"2025-06-26T06:17:54","guid":{"rendered":"https:\/\/mianka.xyz\/?p=184"},"modified":"2025-06-26T14:20:41","modified_gmt":"2025-06-26T06:20:41","slug":"pytorch","status":"publish","type":"post","link":"https:\/\/www.mianka.xyz\/?p=184","title":{"rendered":"PyTorch\u57fa\u7840\u5165\u95e8\u6559\u7a0b"},"content":{"rendered":"

\u4e00\u3001PyTorch\u7b80\u4ecb<\/h2>\n

PyTorch\u662f\u4e00\u4e2a\u57fa\u4e8ePython\u7684\u79d1\u5b66\u8ba1\u7b97\u5305\uff0c\u4e3b\u8981\u7528\u9014\u5305\u62ecNumPy\u7684\u66ff\u4ee3\u54c1\uff08\u4ee5\u4f7f\u7528GPU\u548c\u5176\u4ed6\u52a0\u901f\u5668\u7684\u5f3a\u5927\u529f\u80fd\uff09\u4ee5\u53ca\u4e00\u4e2a\u7528\u4e8e\u5b9e\u73b0\u795e\u7ecf\u7f51\u7edc\u7684\u81ea\u52a8\u5fae\u5206\u5e93\u3002\u5b83\u5177\u6709\u52a8\u6001\u8ba1\u7b97\u56fe\u7684\u7279\u70b9\uff0c\u53ef\u4ee5\u6839\u636e\u8ba1\u7b97\u9700\u8981\u5b9e\u65f6\u6539\u53d8\u8ba1\u7b97\u56fe\uff0c\u4e0eTensorFlow\u7684\u9759\u6001\u8ba1\u7b97\u56fe\u4e0d\u540c\u3002\u7531\u4e8e\u5176\u7b80\u6d01\u6027\u548c\u7b26\u5408Python\u98ce\u683c\u7684\u7279\u70b9\uff0cPyTorch\u5728\u673a\u5668\u5b66\u4e60\u9886\u57df\u5f97\u5230\u4e86\u5e7f\u6cdb\u5e94\u7528\u3002<\/p>\n

\u4e8c\u3001\u5b89\u88c5PyTorch<\/h2>\n

\u5728\u5f00\u59cb\u4f7f\u7528PyTorch\u4e4b\u524d\uff0c\u9700\u8981\u5148\u8fdb\u884c\u5b89\u88c5\u3002\u4ee5\u4e0b\u662f\u5b89\u88c5PyTorch\u7684\u57fa\u672c\u6b65\u9aa4\uff1a<\/p>\n

    \n
  1. \n

    \u5b89\u88c5Conda<\/span>\uff1a\u8bbf\u95eeConda\u5b98\u7f51\uff0c\u6839\u636e\u4f60\u7684\u64cd\u4f5c\u7cfb\u7edf\u9009\u62e9\u5408\u9002\u7684\u5b89\u88c5\u5305\u8fdb\u884c\u4e0b\u8f7d\u548c\u5b89\u88c5\u3002<\/p>\n<\/li>\n

  2. \n

    \u521b\u5efaConda\u73af\u5883<\/span>\uff1a\u5728Anaconda Prompt\u4e2d\u8f93\u5165\u547d\u4ee4\u521b\u5efa\u4e00\u4e2a\u65b0\u7684Conda\u73af\u5883\uff0c\u4f8b\u5982conda create -n pytorch python=3.9<\/code>\u3002<\/p>\n<\/li>\n

  3. \n

    \u6fc0\u6d3b\u73af\u5883<\/span>\uff1a\u4f7f\u7528\u547d\u4ee4conda activate pytorch<\/code>\u6fc0\u6d3b\u65b0\u521b\u5efa\u7684\u73af\u5883\u3002<\/p>\n<\/li>\n

  4. \n

    \u5b89\u88c5PyTorch<\/span>\uff1a\u5728\u6fc0\u6d3b\u7684Conda\u73af\u5883\u4e2d\uff0c\u6839\u636ePyTorch\u5b98\u7f51\u7684\u6307\u5bfc\uff0c\u9009\u62e9\u5408\u9002\u7684\u5b89\u88c5\u547d\u4ee4\u8fdb\u884c\u5b89\u88c5\uff0c\u4f8b\u5982conda install pytorch torchvision torchaudio pytorch-cuda=12.4 -c pytorch -c nvidia<\/code>\u3002<\/p>\n<\/li>\n

  5. \n

    \u68c0\u9a8c\u5b89\u88c5<\/span>\uff1a\u8fdb\u5165Python\u73af\u5883\uff0c\u5bfc\u5165PyTorch\u5e76\u68c0\u9a8cCUDA\u53ef\u7528\u6027\uff0c\u4f8b\u5982import torch<\/code>\u548ctorch.cuda.is_available()<\/code>\u3002<\/p>\n<\/li>\n<\/ol>\n

    \u4e09\u3001Tensor\u57fa\u7840\u64cd\u4f5c<\/h2>\n

    Tensor\uff08\u5f20\u91cf\uff09\u662fPyTorch\u4e2d\u7684\u57fa\u672c\u64cd\u4f5c\u5bf9\u8c61\uff0c\u53ef\u4ee5\u7406\u89e3\u4e3aNumPy\u4e2d\u7684ndarray\u3002\u4ee5\u4e0b\u662fTensor\u7684\u4e00\u4e9b\u57fa\u7840\u64cd\u4f5c\uff1a<\/p>\n

    1. \u521b\u5efaTensor<\/h3>\n

    Tensor\u53ef\u4ee5\u901a\u8fc7\u591a\u79cd\u65b9\u5f0f\u521b\u5efa\uff0c\u5305\u62ec\u76f4\u63a5\u4ece\u6570\u636e\u521b\u5efa\u3001\u4eceNumPy\u6570\u7ec4\u521b\u5efa\u3001\u4ece\u53e6\u4e00\u4e2aTensor\u521b\u5efa\uff0c\u4ee5\u53ca\u6307\u5b9a\u7279\u5b9a\u503c\u6216\u968f\u673a\u521b\u5efa\u3002<\/p>\n

    import torch\nimport numpy as np\n\n# \u76f4\u63a5\u4ece\u6570\u636e\u521b\u5efa\ndata = [[1, 2], [3, 4]]\nx_data = torch.tensor(data)\n\n# \u4eceNumPy\u6570\u7ec4\u521b\u5efa\nnp_array = np.array(data)\nx_np = torch.from_numpy(np_array)\n\n# \u4ece\u53e6\u4e00\u4e2aTensor\u521b\u5efa\nx_ones = torch.ones_like(x_data)\nx_rand = torch.rand_like(x_data, dtype=torch.float)\n\n# \u6307\u5b9a\u7279\u5b9a\u503c\u6216\u968f\u673a\u521b\u5efa\nshape = (2, 3)\nrand_tensor = torch.rand(shape)\nones_tensor = torch.ones(shape)\nzeros_tensor = torch.zeros(shape)<\/pre>\n
    2. Tensor\u7684\u5c5e\u6027<\/h3>\n
    \u53ef\u4ee5\u901a\u8fc7.shape()<\/code>\u6216.size()<\/code>\u65b9\u6cd5\u6765\u83b7\u53d6Tensor\u7684\u5f62\u72b6\uff1b\u901a\u8fc7.dtype()<\/code>\u6765\u67e5\u770b\u5176\u4e2d\u5b58\u653e\u7684\u6570\u636e\u7c7b\u578b\uff1b\u901a\u8fc7.device()<\/code>\u65b9\u6cd5\u6765\u67e5\u770b\u8fd0\u7b97\u6240\u5904\u4e8e\u7684\u8bbe\u5907\u3002<\/p>\n
    pythontensor = torch.rand(3, 4)print(f"Shape of tensor: {tensor.shape}")print(f"Datatype of tensor: {tensor.dtype}")print(f"Device tensor is stored on: {tensor.device}")<\/pre>\n
    3. Tensor\u7684\u8fd0\u7b97<\/h3>\n
    Tensor\u652f\u6301\u591a\u79cd\u8fd0\u7b97\uff0c\u5305\u62ec\u7d22\u5f15\u548c\u5207\u7247\u3001\u5f20\u91cf\u7684\u8fde\u63a5\u3001\u7b97\u6570\u64cd\u4f5c\uff08\u5982\u52a0\u6cd5\uff09\u3001\u77e9\u9635\u4e58\u6cd5\u4ee5\u53ca\u5143\u7d20\u7ea7\u4e58\u6cd5\u7b49\u3002<\/p>\n
    tensor = torch.rand(3, 4)\nprint(f"Shape of tensor: {tensor.shape}")\nprint(f"Datatype of tensor: {tensor.dtype}")\nprint(f"Device tensor is stored on: {tensor.device}")<\/pre>\n
    \u56db\u3001\u81ea\u52a8\u5fae\u5206\uff08Autograd\uff09<\/h2>\n
    PyTorch\u7684\u81ea\u52a8\u5fae\u5206\u673a\u5236\u4f7f\u5f97\u53cd\u5411\u4f20\u64ad\u6c42\u68af\u5ea6\u53d8\u5f97\u975e\u5e38\u7b80\u5355\u3002\u901a\u8fc7\u8bbe\u7f6erequires_grad=True<\/code>\uff0c\u53ef\u4ee5\u4f7f\u5f97Tensor\u5728\u53c2\u4e0e\u8ba1\u7b97\u65f6\u81ea\u52a8\u8bb0\u5f55\u68af\u5ea6\u4fe1\u606f\u3002<\/p>\n
    pythonx = torch.tensor(2.0, requires_grad=True)y = x ** 2y.backward()  # \u53cd\u5411\u4f20\u64ad\u6c42\u68af\u5ea6print(x.grad)  # \u8f93\u51fa\u68af\u5ea6\u503c<\/pre>\n
    \u4e94\u3001\u6570\u636e\u52a0\u8f7d\uff08DataLoader\uff09<\/h2>\n
    PyTorch\u901a\u8fc7Dataset\u548cDataLoader\u8fdb\u884c\u6784\u5efa\u6570\u636e\u7ba1\u9053\u3002Dataset\u7528\u4e8e\u5c01\u88c5\u6570\u636e\u96c6\uff0c\u800cDataLoader\u5219\u7528\u4e8e\u5c06\u6570\u636e\u96c6\u5206\u6279\u52a0\u8f7d\u5230\u6a21\u578b\u4e2d\u8fdb\u884c\u8bad\u7ec3\u3002<\/p>\n
    # \u7d22\u5f15\u548c\u5207\u7247\ntensor = torch.ones(4, 4)\nprint(f"First row: {tensor[0]}")\nprint(f"First column: {tensor[:, 0]}")\nprint(f"Last column: {tensor[..., -1]}")\ntensor[:, 1] = 0\nprint(tensor)\n\n# \u5f20\u91cf\u7684\u8fde\u63a5\nt1 = torch.cat([tensor, tensor, tensor], dim=1)\nprint(t1)\n\n# \u7b97\u6570\u64cd\u4f5c\uff08\u4ee5\u52a0\u6cd5\u4e3a\u4f8b\uff09\nt = torch.randn(4, 4)\nprint(tensor)\nprint(t)\nprint(tensor + t)  # \u76f4\u63a5\u76f8\u52a0\nprint(torch.add(tensor, t))  # \u4f7f\u7528add\nresult = torch.empty(4, 4)\ntorch.add(tensor, t, out=result)  # \u6307\u5b9a\u8f93\u51fa\u5230result\u4e2d\nprint(result)\nt.add_(tensor)  # inplace\u64cd\u4f5c\nprint(t)\n\n# \u77e9\u9635\u4e58\u6cd5\ny1 = tensor @ tensor.T  # tensor.T \u662f tensor \u7684\u8f6c\u7f6e\ny3 = torch.rand_like(y1)\ntorch.matmul(tensor, tensor.T, out=y3)\n\n# \u5143\u7d20\u7ea7\u4e58\u6cd5\n# \u8ba1\u7b97 tensor \u7684\u5143\u7d20\u7ea7\u4e58\u79ef\uff0c\u5373\u6bcf\u4e2a\u4f4d\u7f6e\u4e0a\u7684\u5143\u7d20\u76f8\u4e58<\/pre>\n
    \u516d\u3001\u6a21\u578b\u6784\u5efa\u4e0e\u8bad\u7ec3<\/h2>\n
    \u5728PyTorch\u4e2d\uff0c\u53ef\u4ee5\u901a\u8fc7\u7ee7\u627fModule<\/code>\u7c7b\u6765\u6784\u9020\u6a21\u578b\u3002\u4ee5\u4e0b\u662f\u4e00\u4e2a\u7b80\u5355\u7684\u6a21\u578b\u6784\u5efa\u548c\u8bad\u7ec3\u793a\u4f8b\uff1a<\/p>\n
    import torch.nn as nn\nimport torch.optim as optim\n\n# \u5b9a\u4e49\u6a21\u578b\nclass SimpleModel(nn.Module):\n    def __init__(self):\n        super(SimpleModel, self).__init__()\n        self.fc = nn.Linear(3, 1)  # \u8f93\u5165\u7279\u5f813\uff0c\u8f93\u51fa\u7279\u5f811\n\n    def forward(self, x):\n        return self.fc(x)\n\n# \u521d\u59cb\u5316\u6a21\u578b\u3001\u635f\u5931\u51fd\u6570\u548c\u4f18\u5316\u5668\nmodel = SimpleModel()\ncriterion = nn.MSELoss()  # \u5747\u65b9\u8bef\u5dee\u635f\u5931\u51fd\u6570\noptimizer = optim.SGD(model.parameters(), lr=0.01)  # \u968f\u673a\u68af\u5ea6\u4e0b\u964d\u4f18\u5316\u5668\n\n# \u8bad\u7ec3\u6a21\u578b\nfor epoch in range(100):\n    for batch_data, batch_labels in dataloader:\n        # \u524d\u5411\u4f20\u64ad\n        outputs = model(batch_data)\n        loss = criterion(outputs, batch_labels.float())  # \u5047\u8bbe\u6807\u7b7e\u662f\u6d6e\u70b9\u6570\n\n        # \u53cd\u5411\u4f20\u64ad\u548c\u4f18\u5316\n        optimizer.zero_grad()\n        loss.backward()\n        optimizer.step()\n\n    print(f'Epoch {epoch+1}, Loss: {loss.item()}')<\/pre>\n
    \u4e03\u3001\u7ed3\u8bed<\/h2>\n
    \u901a\u8fc7\u672c\u6587\u7684\u5b66\u4e60\uff0c\u4f60\u5df2\u7ecf\u638c\u63e1\u4e86PyTorch\u7684\u57fa\u7840\u77e5\u8bc6\uff0c\u5305\u62ecTensor\u7684\u521b\u5efa\u548c\u64cd\u4f5c\u3001\u81ea\u52a8\u5fae\u5206\u673a\u5236\u3001\u6570\u636e\u52a0\u8f7d\u4ee5\u53ca\u6a21\u578b\u6784\u5efa\u548c\u8bad\u7ec3\u7b49\u3002\u5e0c\u671b\u8fd9\u4e9b\u5185\u5bb9\u80fd\u591f\u5e2e\u52a9\u4f60\u5feb\u901f\u5165\u95e8PyTorch\uff0c\u5e76\u5f00\u542f\u4f60\u7684\u6df1\u5ea6\u5b66\u4e60\u4e4b\u65c5\u3002<\/p>\n
    <\/p>\n","protected":false},"excerpt":{"rendered":"
    <\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3],"tags":[],"class_list":["post-184","post","type-post","status-publish","format-standard","hentry","category-pythonbiji"],"_links":{"self":[{"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=\/wp\/v2\/posts\/184","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=184"}],"version-history":[{"count":0,"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=\/wp\/v2\/posts\/184\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=184"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=184"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=184"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}