Merge pull request #6 from Beyondzjl/main

附录A第三部分

appendix-A/03_main-chapter-code/DDP-script.py

@@ -1,10 +1,10 @@
-# Appendix A: Introduction to PyTorch (Part 3)
+# 附录A ：PyTorch的介绍（第三部分）
 
 import torch
 import torch.nn.functional as F
 from torch.utils.data import Dataset, DataLoader
 
-# NEW imports:
+# 导入新的库
 import os
 import torch.multiprocessing as mp
 from torch.utils.data.distributed import DistributedSampler
@@ -12,22 +12,23 @@ from torch.nn.parallel import DistributedDataParallel as DDP
 from torch.distributed import init_process_group, destroy_process_group
 
 
-# NEW: function to initialize a distributed process group (1 process / GPU)
+# 创建一个新的函数用于初始化一个分布式进程（每个GPU一个进程）
-# this allows communication among processes
+# 该函数允许进程之间的通信
 def ddp_setup(rank, world_size):
     """
-    Arguments:
+    提示：
-        rank: a unique process ID
+        rank:特定的进程编号（进程ID)
-        world_size: total number of processes in the group
+        world_size:组内的进程总数
     """
-    # rank of machine running rank:0 process
+    
-    # here, we assume all GPUs are on the same machine
+    # 正在运行的机器编号 ID：进程0
+    # 这里的前提是假设所有的GPU在同一台机器上
     os.environ["MASTER_ADDR"] = "localhost"
-    # any free port on the machine
+    # 机器上任意的空闲端口号
     os.environ["MASTER_PORT"] = "12345"
 
-    # initialize process group
+    # 初始化进程
-    # Windows users may have to use "gloo" instead of "nccl" as backend
+    # Windows 用户使用"gloo"来替代下面代码中的"nccl"
     # nccl: NVIDIA Collective Communication Library
     init_process_group(backend="nccl", rank=rank, world_size=world_size)
     torch.cuda.set_device(rank)
@@ -52,15 +53,15 @@ class NeuralNetwork(torch.nn.Module):
         super().__init__()
 
         self.layers = torch.nn.Sequential(
-            # 1st hidden layer
+            # 第一个隐藏层
             torch.nn.Linear(num_inputs, 30),
             torch.nn.ReLU(),
 
-            # 2nd hidden layer
+            # 第二个隐藏层
             torch.nn.Linear(30, 20),
             torch.nn.ReLU(),
 
-            # output layer
+            # 输出层
             torch.nn.Linear(20, num_outputs),
         )
 
@@ -91,11 +92,11 @@ def prepare_dataset():
     train_loader = DataLoader(
         dataset=train_ds,
         batch_size=2,
-        shuffle=False, # NEW: False because of DistributedSampler below
+        shuffle=False, # 这里设置为False 
         pin_memory=True,
         drop_last=True,
-        # NEW: chunk batches across GPUs without overlapping samples:
+        # 在多个GPU上划分批次，确保批次之间不重叠样本
-        sampler=DistributedSampler(train_ds) # NEW
+        sampler=DistributedSampler(train_ds) 
     )
     test_loader = DataLoader(
         dataset=test_ds,
@@ -105,33 +106,33 @@ def prepare_dataset():
     return train_loader, test_loader
 
 
-# NEW: wrapper
+# 包装器
 def main(rank, world_size, num_epochs):
 
-    ddp_setup(rank, world_size) # NEW: initialize process groups
+    ddp_setup(rank, world_size) # 
 
     train_loader, test_loader = prepare_dataset()
     model = NeuralNetwork(num_inputs=2, num_outputs=2)
     model.to(rank)
     optimizer = torch.optim.SGD(model.parameters(), lr=0.5)
 
-    model = DDP(model, device_ids=[rank]) # NEW: wrap model with DDP
+    model = DDP(model, device_ids=[rank]) # 使用分布式数据并行（DDP）将模型进行包装
-    # the core model is now accessible as model.module
+    # 现在核心模型可以通过 model.module 访问
     
     for epoch in range(num_epochs):
     
         model.train()
         for features, labels in enumerate(train_loader):
     
-            features, labels = features.to(rank), labels.to(rank) # New: use rank
+            features, labels = features.to(rank), labels.to(rank) 
             logits = model(features)
-            loss = F.cross_entropy(logits, labels) # Loss function
+            loss = F.cross_entropy(logits, labels) # 损失函数
     
             optimizer.zero_grad()
             loss.backward()
             optimizer.step()
     
-            ### LOGGING
+            ### 日志
             print(f"[GPU{rank}] Epoch: {epoch+1:03d}/{num_epochs:03d}"
                   f" | Batchsize {labels.shape[0]:03d}"
                   f" | Train/Val Loss: {loss:.2f}")
@@ -142,7 +143,7 @@ def main(rank, world_size, num_epochs):
     test_acc = compute_accuracy(model, test_loader, device=rank)
     print(f"[GPU{rank}] Test accuracy", test_acc)
 
-    destroy_process_group() # NEW: cleanly exit distributed mode
+    destroy_process_group() # 清理退出分布式模式
 
 
 def compute_accuracy(model, dataloader, device):
@@ -169,10 +170,10 @@ if __name__ == "__main__":
 
     torch.manual_seed(123)
 
-    # NEW: spawn new processes
+    # 新建进程
-    # note that spawn will automatically pass the rank
+    # 请注意，spawn会自动传递排名
     num_epochs = 3
     world_size = torch.cuda.device_count()
     mp.spawn(main, args=(world_size, num_epochs), nprocs=world_size)
-    # nprocs=world_size spawns one process per GPU
+    # nprocs=world_size 会为每个GPU生成一个进程
 

appendix-A/03_main-chapter-code/code-part1.ipynb

@@ -5,7 +5,7 @@
    "id": "ca7fc8a0-280c-4979-b0c7-fc3a99b3b785",
    "metadata": {},
    "source": [
-    "# Appendix A: Introduction to PyTorch (Part 1)"
+    "# 附件A：PyTorch的介绍（第一部分）"
    ]
   },
   {
@@ -13,12 +13,12 @@
    "id": "f5bf13d2-8fc2-483e-88cc-6b4310221e68",
    "metadata": {},
    "source": [
-    "## A.1 What is PyTorch"
+    "## A.1 什么是PyTorch"
    ]
   },
   {
    "cell_type": "code",
-   "execution_count": 1,
+   "execution_count": null,
    "id": "96ee5660-5327-48e2-9104-a882b3b2afa4",
    "metadata": {},
    "outputs": [
@@ -32,13 +32,13 @@
    ],
    "source": [
     "import torch\n",
-    "\n",
+    "# 显示PyTorch的版本\n",
     "print(torch.__version__)"
    ]
   },
   {
    "cell_type": "code",
-   "execution_count": 2,
+   "execution_count": null,
    "id": "f73ad4e4-7ec6-4467-a9e9-0cdf6d195264",
    "metadata": {},
    "outputs": [
@@ -51,6 +51,7 @@
     }
    ],
    "source": [
+    "# 显示PyTorch是否是GPU版本，False表示CPU版本，True表示GPU版本\n",
     "print(torch.cuda.is_available())"
    ]
   },
@@ -59,7 +60,7 @@
    "id": "2100cf2e-7459-4ab3-92a8-43e86ab35a9b",
    "metadata": {},
    "source": [
-    "## A.2 Understanding tensors"
+    "## A.2 向量"
    ]
   },
   {
@@ -67,7 +68,7 @@
    "id": "26d7f785-e048-42bc-9182-a556af6bb7f4",
    "metadata": {},
    "source": [
-    "### A.2.1 Scalars, vectors, matrices, and tensors"
+    "### A.2.1 标量、向量、矩阵和张量\n"
    ]
   },
   {
@@ -80,22 +81,22 @@
     "import torch\n",
     "import numpy as np\n",
     "\n",
-    "# create a 0D tensor (scalar) from a Python integer\n",
+    "# 用Python整数创建一个0维张量\n",
     "tensor0d = torch.tensor(1)\n",
     "\n",
-    "# create a 1D tensor (vector) from a Python list\n",
+    "# 用Python列表创建一个1维张量（向量）\n",
     "tensor1d = torch.tensor([1, 2, 3])\n",
     "\n",
-    "# create a 2D tensor from a nested Python list\n",
+    "# 用Python列表创建一个2维张量（向量）\n",
     "tensor2d = torch.tensor([[1, 2], [3, 4]])\n",
     "\n",
-    "# create a 3D tensor from a nested Python list\n",
+    "# 用嵌套的Python列表创建一个3维张量\n",
     "tensor3d_1 = torch.tensor([[[1, 2], [3, 4]], [[5, 6], [7, 8]]])\n",
     "\n",
-    "# create a 3D tensor from NumPy array\n",
+    "# 从NumPy数组创建一个3维张量\n",
     "ary3d = np.array([[[1, 2], [3, 4]], [[5, 6], [7, 8]]])\n",
-    "tensor3d_2 = torch.tensor(ary3d)  # Copies NumPy array\n",
+    "tensor3d_2 = torch.tensor(ary3d)  # 复制NumPy数组\n",
-    "tensor3d_3 = torch.from_numpy(ary3d)  # Shares memory with NumPy array"
+    "tensor3d_3 = torch.from_numpy(ary3d)  # 与NumPy数组共享内存"
    ]
   },
   {
@@ -118,7 +119,7 @@
    ],
    "source": [
     "ary3d[0, 0, 0] = 999\n",
-    "print(tensor3d_2) # remains unchanged"
+    "print(tensor3d_2) # 保持不变"
    ]
   },
   {
@@ -140,7 +141,7 @@
     }
    ],
    "source": [
-    "print(tensor3d_3) # changes because of memory sharing"
+    "print(tensor3d_3) # 由于内存共享需要改变"
    ]
   },
   {
@@ -148,7 +149,7 @@
    "id": "63dec48d-2b60-41a2-ac06-fef7e718605a",
    "metadata": {},
    "source": [
-    "### A.2.2 Tensor data types"
+    "### A.2.2 向量的数据类型"
    ]
   },
   {
@@ -213,7 +214,7 @@
    "id": "2020deb5-aa02-4524-b311-c010f4ad27ff",
    "metadata": {},
    "source": [
-    "### A.2.3 Common PyTorch tensor operations"
+    "### A.2.3 PyTorch中常见的张量操作"
    ]
   },
   {
@@ -257,7 +258,7 @@
     }
    ],
    "source": [
-    "tensor2d.shape"
+    "tensor2d.shape # 张量形状"
    ]
   },
   {
@@ -280,7 +281,7 @@
     }
    ],
    "source": [
-    "tensor2d.reshape(3, 2)"
+    "tensor2d.reshape(3, 2) # 修改形状"
    ]
   },
   {
@@ -303,7 +304,7 @@
     }
    ],
    "source": [
-    "tensor2d.view(3, 2)"
+    "tensor2d.view(3, 2) # 查看张量"
    ]
   },
   {
@@ -326,7 +327,7 @@
     }
    ],
    "source": [
-    "tensor2d.T"
+    "tensor2d.T # 转置张量"
    ]
   },
   {
@@ -348,7 +349,7 @@
     }
    ],
    "source": [
-    "tensor2d.matmul(tensor2d.T)"
+    "tensor2d.matmul(tensor2d.T) # 张量乘法：tensor2d与其转置相乘"
    ]
   },
   {
@@ -370,7 +371,7 @@
     }
    ],
    "source": [
-    "tensor2d @ tensor2d.T"
+    "tensor2d @ tensor2d.T # 张量乘法的另一种实现方式：tensor2d与其转置相乘"
    ]
   },
   {
@@ -378,7 +379,7 @@
    "id": "4c15bdeb-78e2-4870-8a4f-a9f591666f38",
    "metadata": {},
    "source": [
-    "## A.3 Seeing models as computation graphs"
+    "## A.3 把模型作为计算图"
    ]
   },
   {
@@ -398,13 +399,13 @@
    "source": [
     "import torch.nn.functional as F\n",
     "\n",
-    "y = torch.tensor([1.0])  # true label\n",
+    "y = torch.tensor([1.0])  # 真实样本\n",
-    "x1 = torch.tensor([1.1]) # input feature\n",
+    "x1 = torch.tensor([1.1]) # 输入特征\n",
-    "w1 = torch.tensor([2.2]) # weight parameter\n",
+    "w1 = torch.tensor([2.2]) # 权重变量\n",
-    "b = torch.tensor([0.0])  # bias unit\n",
+    "b = torch.tensor([0.0])  # 偏置单元\n",
     "\n",
-    "z = x1 * w1 + b          # net input\n",
+    "z = x1 * w1 + b          # 网络输入\n",
-    "a = torch.sigmoid(z)     # activation & output\n",
+    "a = torch.sigmoid(z)     # 激活函数 & 输出\n",
     "\n",
     "loss = F.binary_cross_entropy(a, y)\n",
     "print(loss)"
@@ -415,7 +416,7 @@
    "id": "f9424f26-2bac-47e7-b834-92ece802247c",
    "metadata": {},
    "source": [
-    "## A.4 Automatic differentiation made easy"
+    "## A.4 自动求导"
    ]
   },
   {
@@ -470,7 +471,7 @@
     }
    ],
    "source": [
-    "loss.backward()\n",
+    "loss.backward()# 反向传播\n",
     "\n",
     "print(w1.grad)\n",
     "print(b.grad)"
@@ -481,7 +482,7 @@
    "id": "f53bdd7d-44e6-40ab-8a5a-4eef74ef35dc",
    "metadata": {},
    "source": [
-    "## A.5 Implementing multilayer neural networks"
+    "## A.5 多层神经网络的实现"
    ]
   },
   {
@@ -497,15 +498,15 @@
     "\n",
     "        self.layers = torch.nn.Sequential(\n",
     "                \n",
-    "            # 1st hidden layer\n",
+    "            # 第一个隐藏层\n",
     "            torch.nn.Linear(num_inputs, 30),\n",
     "            torch.nn.ReLU(),\n",
     "\n",
-    "            # 2nd hidden layer\n",
+    "            # 第二个隐藏层\n",
     "            torch.nn.Linear(30, 20),\n",
     "            torch.nn.ReLU(),\n",
     "\n",
-    "            # output layer\n",
+    "            # 输出层\n",
     "            torch.nn.Linear(20, num_outputs),\n",
     "        )\n",
     "\n",
@@ -566,7 +567,7 @@
    ],
    "source": [
     "num_params = sum(p.numel() for p in model.parameters() if p.requires_grad)\n",
-    "print(\"Total number of trainable model parameters:\", num_params)"
+    "print(\"Total number of trainable model parameters:\", num_params)# 打印训练模型的参数"
    ]
   },
   {
@@ -592,7 +593,7 @@
     }
    ],
    "source": [
-    "print(model.layers[0].weight)"
+    "print(model.layers[0].weight) # 打印神经网络模型的第一层的权重"
    ]
   },
   {
@@ -618,9 +619,13 @@
     }
    ],
    "source": [
+    "# 设置随机数种子，以确保可复现性\n",
     "torch.manual_seed(123)\n",
     "\n",
+    "# 假设 NeuralNetwork 是一个神经网络类，且其构造函数接受两个参数，分别为输入特征的维度和输出特征的维度\n",
     "model = NeuralNetwork(50, 3)\n",
+    "\n",
+    "# 打印神经网络模型的第一层的权重\n",
     "print(model.layers[0].weight)"
    ]
   },
@@ -639,6 +644,7 @@
     }
    ],
    "source": [
+    "# 打印神经网络模型的第一层权重的形状\n",
     "print(model.layers[0].weight.shape)"
    ]
   },
@@ -657,10 +663,16 @@
     }
    ],
    "source": [
+    "# 设置随机数种子，以确保可复现性\n",
     "torch.manual_seed(123)\n",
     "\n",
+    "# 模型输入特征的维度为 50\n",
     "X = torch.rand((1, 50))\n",
+    "\n",
+    "# 使用模型进行前向传播计算输出\n",
     "out = model(X)\n",
+    "\n",
+    "# 打印输出结果\n",
     "print(out)"
    ]
   },
@@ -679,6 +691,7 @@
     }
    ],
    "source": [
+    "# 使用 torch.no_grad() 上下文管理器，以便在推断时不计算梯度\n",
     "with torch.no_grad():\n",
     "    out = model(X)\n",
     "print(out)"
@@ -709,7 +722,7 @@
    "id": "19858180-0f26-43a8-b2c3-7ed40abf9f85",
    "metadata": {},
    "source": [
-    "## A.6 Setting up efficient data loaders"
+    "## A.6 建立高效的数据加载器"
    ]
   },
   {
@@ -719,6 +732,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
+    "# 定义输入特征张量 X_train\n",
     "X_train = torch.tensor([\n",
     "    [-1.2, 3.1],\n",
     "    [-0.9, 2.9],\n",
@@ -727,6 +741,7 @@
     "    [2.7, -1.5]\n",
     "])\n",
     "\n",
+    "# 定义对应的标签张量 y_train\n",
     "y_train = torch.tensor([0, 0, 0, 1, 1])"
    ]
   },
@@ -756,18 +771,19 @@
     "\n",
     "\n",
     "class ToyDataset(Dataset):\n",
+    "    # 初始化 ToyDataset 类\n",
     "    def __init__(self, X, y):\n",
     "        self.features = X\n",
     "        self.labels = y\n",
-    "\n",
+    "    #  获取指定索引的数据\n",
     "    def __getitem__(self, index):\n",
     "        one_x = self.features[index]\n",
     "        one_y = self.labels[index]        \n",
     "        return one_x, one_y\n",
-    "\n",
+    "    # 获取数据集的长度\n",
     "    def __len__(self):\n",
     "        return self.labels.shape[0]\n",
-    "\n",
+    "# 创建训练数据集和测试数据集实例\n",
     "train_ds = ToyDataset(X_train, y_train)\n",
     "test_ds = ToyDataset(X_test, y_test)"
    ]
@@ -804,6 +820,11 @@
     "\n",
     "torch.manual_seed(123)\n",
     "\n",
+    "# 创建训练数据加载器 train_loader\n",
+    "# dataset 参数传入了您定义的 ToyDataset 类的实例 train_ds\n",
+    "# batch_size 参数指定了每个批次包含的样本数量\n",
+    "# shuffle 参数指定是否在每个 epoch 之前对数据进行洗牌\n",
+    "# num_workers 参数指定用于数据加载的子进程数量\n",
     "train_loader = DataLoader(\n",
     "    dataset=train_ds,\n",
     "    batch_size=2,\n",
@@ -821,6 +842,11 @@
    "source": [
     "test_ds = ToyDataset(X_test, y_test)\n",
     "\n",
+    "# 创建测试数据加载器 test_loader\n",
+    "# dataset 参数传入了您定义的 ToyDataset 类的实例 test_ds\n",
+    "# batch_size 参数指定了每个批次包含的样本数量\n",
+    "# shuffle 参数指定是否在每个 epoch 之前对数据进行洗牌，这里设为 False 表示不洗牌\n",
+    "# num_workers 参数指定用于数据加载的子进程数量\n",
     "test_loader = DataLoader(\n",
     "    dataset=test_ds,\n",
     "    batch_size=2,\n",
@@ -848,7 +874,9 @@
     }
    ],
    "source": [
+    "# 迭代训练数据加载器 train_loader\n",
     "for idx, (x, y) in enumerate(train_loader):\n",
+    "    # 打印每个批次的索引、输入特征和对应的标签\n",
     "    print(f\"Batch {idx+1}:\", x, y)"
    ]
   },
@@ -884,7 +912,7 @@
    "id": "d904ca82-e50f-4f3d-a3ac-fc6ca53dd00e",
    "metadata": {},
    "source": [
-    "## A.7 A typical training loop"
+    "## A.7 一个示例训练轮次"
    ]
   },
   {
@@ -923,19 +951,19 @@
     "\n",
     "        logits = model(features)\n",
     "        \n",
-    "        loss = F.cross_entropy(logits, labels) # Loss function\n",
+    "        loss = F.cross_entropy(logits, labels) # 损失函数\n",
     "        \n",
     "        optimizer.zero_grad()\n",
     "        loss.backward()\n",
     "        optimizer.step()\n",
     "    \n",
-    "        ### LOGGING\n",
+    "        ### 日志\n",
     "        print(f\"Epoch: {epoch+1:03d}/{num_epochs:03d}\"\n",
     "              f\" | Batch {batch_idx:03d}/{len(train_loader):03d}\"\n",
     "              f\" | Train/Val Loss: {loss:.2f}\")\n",
     "\n",
     "    model.eval()\n",
-    "    # Optional model evaluation"
+    "    # 可选的模型评估指标"
    ]
   },
   {
@@ -985,10 +1013,16 @@
     }
    ],
    "source": [
+    "# 设置 PyTorch 的打印选项，以关闭科学计数法\n",
     "torch.set_printoptions(sci_mode=False)\n",
+    "\n",
+    "# 假设 outputs 是模型的输出张量\n",
+    "\n",
+    "# 对模型的输出进行 softmax 操作，计算类别概率\n",
     "probas = torch.softmax(outputs, dim=1)\n",
     "print(probas)\n",
     "\n",
+    "# 获取模型的预测结果，即具有最大概率的类别\n",
     "predictions = torch.argmax(outputs, dim=1)\n",
     "print(predictions)"
    ]
@@ -1008,6 +1042,7 @@
     }
    ],
    "source": [
+    "# 使用 torch.argmax() 函数沿着 dim=1 维度获取每个样本最大值的索引，即模型的预测结果\n",
     "predictions = torch.argmax(outputs, dim=1)\n",
     "print(predictions)"
    ]
@@ -1062,21 +1097,36 @@
    "outputs": [],
    "source": [
     "def compute_accuracy(model, dataloader):\n",
+    "    \"\"\"\n",
+    "    计算模型在给定数据加载器上的准确率。\n",
     "\n",
+    "    参数：\n",
+    "        model (torch.nn.Module): 待评估的模型。\n",
+    "        dataloader (torch.utils.data.DataLoader): 包含输入数据的数据加载器。\n",
+    "\n",
+    "    返回：\n",
+    "        float: 准确率值。\n",
+    "    \"\"\"\n",
+    "    # 将模型设为评估模式\n",
     "    model = model.eval()\n",
     "    correct = 0.0\n",
     "    total_examples = 0\n",
     "    \n",
+    "    # 遍历数据加载器\n",
     "    for idx, (features, labels) in enumerate(dataloader):\n",
     "        \n",
+    "        # 使用 no_grad 上下文，以便不跟踪梯度\n",
     "        with torch.no_grad():\n",
+    "            # 使用模型进行前向传播获取预测结果\n",
     "            logits = model(features)\n",
     "        \n",
+    "        # 获取预测结果并计算正确预测的数量\n",
     "        predictions = torch.argmax(logits, dim=1)\n",
     "        compare = labels == predictions\n",
     "        correct += torch.sum(compare)\n",
     "        total_examples += len(compare)\n",
     "\n",
+    "    # 计算并返回准确率\n",
     "    return (correct / total_examples).item()"
    ]
   },
@@ -1127,7 +1177,7 @@
    "id": "4d5cd469-3a45-4394-944b-3ce543f41dac",
    "metadata": {},
    "source": [
-    "## A.8 Saving and loading models"
+    "## A.8 保存并加载模型"
    ]
   },
   {
@@ -1158,7 +1208,7 @@
     }
    ],
    "source": [
-    "model = NeuralNetwork(2, 2) # needs to match the original  model exactly\n",
+    "model = NeuralNetwork(2, 2) # 需要与原始模型完全匹配\n",
     "model.load_state_dict(torch.load(\"model.pth\"))"
    ]
   },
@@ -1167,7 +1217,7 @@
    "id": "f891c013-43da-4a05-973d-997be313d2d8",
    "metadata": {},
    "source": [
-    "## A.9 Optimizing training performance with GPUs"
+    "## A.9 使用GPU来优化训练性能"
    ]
   },
   {
@@ -1175,7 +1225,7 @@
    "id": "e68ae888-cabf-49c9-bad6-ecdce774db57",
    "metadata": {},
    "source": [
-    "### A.9.1 PyTorch computations on GPU devices"
+    "### A.9.1 在GPU上进行 PyTorch 的运算"
    ]
   },
   {
@@ -1183,7 +1233,7 @@
    "id": "141c845f-efe3-4614-b376-b8b7a9a2c887",
    "metadata": {},
    "source": [
-    "See [code-part2.ipynb](code-part2.ipynb)"
+    "See [code-part2.ipynb](code-part2.ipynb)  "
    ]
   },
   {
@@ -1191,7 +1241,7 @@
    "id": "99811829-b817-42ea-b03e-d35374debcc0",
    "metadata": {},
    "source": [
-    "### A.9.2 Single-GPU training"
+    "### A.9.2 单个GPU的训练"
    ]
   },
   {
@@ -1207,7 +1257,7 @@
    "id": "db6eb2d1-a341-4489-b04b-635c26945333",
    "metadata": {},
    "source": [
-    "### A.9.3 Training with multiple GPUs"
+    "### A.9.3 多GPU的训练"
    ]
   },
   {

appendix-A/03_main-chapter-code/code-part2.ipynb

@@ -6,7 +6,7 @@
     "id": "O9i6kzBsZVaZ"
    },
    "source": [
-    "# Appendix A: Introduction to PyTorch (Part 2)"
+    "# 附件A：PyTorch的介绍（第二部分）"
    ]
   },
   {
@@ -15,7 +15,7 @@
     "id": "ppbG5d-NZezH"
    },
    "source": [
-    "## A.9 Optimizing training performance with GPUs"
+    "## A.9 使用GPU优化训练性能"
    ]
   },
   {
@@ -24,7 +24,7 @@
     "id": "6jH0J_DPZhbn"
    },
    "source": [
-    "### A.9.1 PyTorch computations on GPU devices"
+    "### A.9.1 在GPU上进行 PyTorch 计算"
    ]
   },
   {
@@ -48,13 +48,13 @@
    ],
    "source": [
     "import torch\n",
-    "\n",
+    "# 显示PyTorch的版本\n",
     "print(torch.__version__)"
    ]
   },
   {
    "cell_type": "code",
-   "execution_count": 2,
+   "execution_count": null,
    "metadata": {
     "colab": {
      "base_uri": "https://localhost:8080/"
@@ -72,6 +72,7 @@
     }
    ],
    "source": [
+    "# 显示PyTorch是否支持GPU\n",
     "print(torch.cuda.is_available())"
    ]
   },
@@ -124,6 +125,7 @@
     }
    ],
    "source": [
+    "# 将两个张量移动到CUDA设备上\n",
     "tensor_1 = tensor_1.to(\"cuda\")\n",
     "tensor_2 = tensor_2.to(\"cuda\")\n",
     "\n",
@@ -165,7 +167,7 @@
     "id": "c8j1cWDcWAMf"
    },
    "source": [
-    "## A.9.2 Single-GPU training"
+    "## A.9.2 单GPU训练"
    ]
   },
   {
@@ -264,15 +266,15 @@
     "\n",
     "        self.layers = torch.nn.Sequential(\n",
     "\n",
-    "            # 1st hidden layer\n",
+    "            # 第一个隐藏层\n",
     "            torch.nn.Linear(num_inputs, 30),\n",
     "            torch.nn.ReLU(),\n",
     "\n",
-    "            # 2nd hidden layer\n",
+    "            # 第二个隐藏层\n",
     "            torch.nn.Linear(30, 20),\n",
     "            torch.nn.ReLU(),\n",
     "\n",
-    "            # output layer\n",
+    "            # 输出层\n",
     "            torch.nn.Linear(20, num_outputs),\n",
     "        )\n",
     "\n",
@@ -309,14 +311,22 @@
     "import torch.nn.functional as F\n",
     "\n",
     "\n",
+    "# 设置随机数种子，以确保可复现性\n",
     "torch.manual_seed(123)\n",
+    "\n",
+    "# 创建神经网络模型\n",
     "model = NeuralNetwork(num_inputs=2, num_outputs=2)\n",
     "\n",
-    "device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\") # NEW\n",
+    "# 根据设备可用情况选择设备\n",
-    "model = model.to(device) # NEW\n",
+    "device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
     "\n",
+    "# 将模型移动到所选设备上\n",
+    "model = model.to(device)\n",
+    "\n",
+    "# 定义优化器，使用随机梯度下降 (SGD)\n",
     "optimizer = torch.optim.SGD(model.parameters(), lr=0.5)\n",
     "\n",
+    "# 定义训练循环的 epoch 数量\n",
     "num_epochs = 3\n",
     "\n",
     "for epoch in range(num_epochs):\n",
@@ -324,21 +334,21 @@
     "    model.train()\n",
     "    for batch_idx, (features, labels) in enumerate(train_loader):\n",
     "\n",
-    "        features, labels = features.to(device), labels.to(device) # NEW\n",
+    "        features, labels = features.to(device), labels.to(device) \n",
     "        logits = model(features)\n",
-    "        loss = F.cross_entropy(logits, labels) # Loss function\n",
+    "        loss = F.cross_entropy(logits, labels) # 损失函数\n",
     "\n",
     "        optimizer.zero_grad()\n",
     "        loss.backward()\n",
     "        optimizer.step()\n",
     "\n",
-    "        ### LOGGING\n",
+    "        ### 训练日志\n",
     "        print(f\"Epoch: {epoch+1:03d}/{num_epochs:03d}\"\n",
     "              f\" | Batch {batch_idx:03d}/{len(train_loader):03d}\"\n",
     "              f\" | Train/Val Loss: {loss:.2f}\")\n",
     "\n",
     "    model.eval()\n",
-    "    # Optional model evaluation"
+    "    # 可选的模型参数"
    ]
   },
   {
@@ -349,6 +359,7 @@
    },
    "outputs": [],
    "source": [
+    "# 使用accuracy（准确率）作为指标\n",
     "def compute_accuracy(model, dataloader, device):\n",
     "\n",
     "    model = model.eval()\n",
@@ -356,17 +367,17 @@
     "    total_examples = 0\n",
     "\n",
     "    for idx, (features, labels) in enumerate(dataloader):\n",
-    "\n",
+    "        # 将数据移动到指定的设备上\n",
     "        features, labels = features.to(device), labels.to(device) # New\n",
     "\n",
     "        with torch.no_grad():\n",
     "            logits = model(features)\n",
-    "\n",
+    "        # 获取预测结果并计算准确数量\n",
     "        predictions = torch.argmax(logits, dim=1)\n",
     "        compare = labels == predictions\n",
     "        correct += torch.sum(compare)\n",
     "        total_examples += len(compare)\n",
-    "\n",
+    "    # 计算并返回准确率\n",
     "    return (correct / total_examples).item()"
    ]
   },

appendix-A/03_main-chapter-code/exercise-solutions.ipynb

@@ -4,7 +4,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "## Exercise A.3"
+    "## 练习 A.3"
    ]
   },
   {
@@ -21,15 +21,15 @@
     "\n",
     "        self.layers = torch.nn.Sequential(\n",
     "                \n",
-    "            # 1st hidden layer\n",
+    "            # 第一个隐藏层\n",
     "            torch.nn.Linear(num_inputs, 30),\n",
     "            torch.nn.ReLU(),\n",
     "\n",
-    "            # 2nd hidden layer\n",
+    "            # 第二个隐藏层\n",
     "            torch.nn.Linear(30, 20),\n",
     "            torch.nn.ReLU(),\n",
     "\n",
-    "            # output layer\n",
+    "            # 输出层\n",
     "            torch.nn.Linear(20, num_outputs),\n",
     "        )\n",
     "\n",
@@ -62,7 +62,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "## Exercise A.4"
+    "## 练习 A.4"
    ]
   },
   {
@@ -74,7 +74,7 @@
    "outputs": [],
    "source": [
     "import torch\n",
-    "\n",
+    "# 创建随机向量\n",
     "a = torch.rand(100, 200)\n",
     "b = torch.rand(200, 300)"
    ]
@@ -99,6 +99,9 @@
     }
    ],
    "source": [
+    "# 使用 @ 符号进行矩阵相乘，并计算执行时间\n",
+    "# %timeit 是 IPython 提供的魔术命令，用于多次执行代码以获取平均执行时间\n",
+    "# 它会自动选择执行次数以确保结果的准确性\n",
     "%timeit a @ b"
    ]
   },
@@ -110,6 +113,7 @@
    },
    "outputs": [],
    "source": [
+    "# 将 a 和 b 移动到 CUDA 设备上以利用 GPU 加速计算\n",
     "a, b = a.to(\"cuda\"), b.to(\"cuda\")"
    ]
   },
@@ -135,15 +139,6 @@
    "source": [
     "%timeit a @ b"
    ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "id": "Zqqa-To2L749"
-   },
-   "outputs": [],
-   "source": []
   }
  ],
  "metadata": {
@@ -168,7 +163,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.10.6"
+   "version": "3.11.5"
   }
  },
  "nbformat": 4,