mirror of
https://github.com/datawhalechina/llms-from-scratch-cn.git
synced 2026-01-14 01:07:34 +08:00
1302 lines
30 KiB
Plaintext
1302 lines
30 KiB
Plaintext
{
|
||
"cells": [
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "ca7fc8a0-280c-4979-b0c7-fc3a99b3b785",
|
||
"metadata": {},
|
||
"source": [
|
||
"# 附件A：PyTorch的介绍（第一部分）"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "f5bf13d2-8fc2-483e-88cc-6b4310221e68",
|
||
"metadata": {},
|
||
"source": [
|
||
"## A.1 什么是PyTorch"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "96ee5660-5327-48e2-9104-a882b3b2afa4",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"2.0.1\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"import torch\n",
|
||
"# 显示PyTorch的版本\n",
|
||
"print(torch.__version__)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "f73ad4e4-7ec6-4467-a9e9-0cdf6d195264",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"False\n"
|
||
]
|
||
}
|
||
],
|
||
"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())"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "2100cf2e-7459-4ab3-92a8-43e86ab35a9b",
|
||
"metadata": {},
|
||
"source": [
|
||
"## A.2 向量"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "26d7f785-e048-42bc-9182-a556af6bb7f4",
|
||
"metadata": {},
|
||
"source": [
|
||
"### A.2.1 标量、向量、矩阵和张量\n"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 3,
|
||
"id": "a3a464d6-cec8-4363-87bd-ea4f900baced",
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"import torch\n",
|
||
"import numpy as np\n",
|
||
"\n",
|
||
"# 用Python整数创建一个0维张量\n",
|
||
"tensor0d = torch.tensor(1)\n",
|
||
"\n",
|
||
"# 用Python列表创建一个1维张量（向量）\n",
|
||
"tensor1d = torch.tensor([1, 2, 3])\n",
|
||
"\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",
|
||
"# 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",
|
||
"# 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) # 复制NumPy数组\n",
|
||
"tensor3d_3 = torch.from_numpy(ary3d) # 与NumPy数组共享内存"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 4,
|
||
"id": "dbe14c47-499a-4d48-b354-a0e6fd957872",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"tensor([[[1, 2],\n",
|
||
" [3, 4]],\n",
|
||
"\n",
|
||
" [[5, 6],\n",
|
||
" [7, 8]]])\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"ary3d[0, 0, 0] = 999\n",
|
||
"print(tensor3d_2) # 保持不变"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 5,
|
||
"id": "e3e4c23a-cdba-46f5-a2dc-5fb32bf9117b",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"tensor([[[999, 2],\n",
|
||
" [ 3, 4]],\n",
|
||
"\n",
|
||
" [[ 5, 6],\n",
|
||
" [ 7, 8]]])\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"print(tensor3d_3) # 由于内存共享需要改变"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "63dec48d-2b60-41a2-ac06-fef7e718605a",
|
||
"metadata": {},
|
||
"source": [
|
||
"### A.2.2 向量的数据类型"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 6,
|
||
"id": "3f48c014-e1a2-4a53-b5c5-125812d4034c",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"torch.int64\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"tensor1d = torch.tensor([1, 2, 3])\n",
|
||
"print(tensor1d.dtype)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 7,
|
||
"id": "5429a086-9de2-4ac7-9f14-d087a7507394",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"torch.float32\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"floatvec = torch.tensor([1.0, 2.0, 3.0])\n",
|
||
"print(floatvec.dtype)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 8,
|
||
"id": "a9a438d1-49bb-481c-8442-7cc2bb3dd4af",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"torch.float32\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"floatvec = tensor1d.to(torch.float32)\n",
|
||
"print(floatvec.dtype)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "2020deb5-aa02-4524-b311-c010f4ad27ff",
|
||
"metadata": {},
|
||
"source": [
|
||
"### A.2.3 PyTorch中常见的张量操作"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 9,
|
||
"id": "c02095f2-8a48-4953-b3c9-5313d4362ce7",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"data": {
|
||
"text/plain": [
|
||
"tensor([[1, 2, 3],\n",
|
||
" [4, 5, 6]])"
|
||
]
|
||
},
|
||
"execution_count": 9,
|
||
"metadata": {},
|
||
"output_type": "execute_result"
|
||
}
|
||
],
|
||
"source": [
|
||
"tensor2d = torch.tensor([[1, 2, 3], [4, 5, 6]])\n",
|
||
"tensor2d"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 10,
|
||
"id": "f33e1d45-5b2c-4afe-b4b2-66ac4099fd1a",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"data": {
|
||
"text/plain": [
|
||
"torch.Size([2, 3])"
|
||
]
|
||
},
|
||
"execution_count": 10,
|
||
"metadata": {},
|
||
"output_type": "execute_result"
|
||
}
|
||
],
|
||
"source": [
|
||
"tensor2d.shape # 张量形状"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 11,
|
||
"id": "f3a4129d-f870-4e03-9c32-cd8521cb83fe",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"data": {
|
||
"text/plain": [
|
||
"tensor([[1, 2],\n",
|
||
" [3, 4],\n",
|
||
" [5, 6]])"
|
||
]
|
||
},
|
||
"execution_count": 11,
|
||
"metadata": {},
|
||
"output_type": "execute_result"
|
||
}
|
||
],
|
||
"source": [
|
||
"tensor2d.reshape(3, 2) # 修改形状"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 12,
|
||
"id": "589ac0a7-adc7-41f3-b721-155f580e9369",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"data": {
|
||
"text/plain": [
|
||
"tensor([[1, 2],\n",
|
||
" [3, 4],\n",
|
||
" [5, 6]])"
|
||
]
|
||
},
|
||
"execution_count": 12,
|
||
"metadata": {},
|
||
"output_type": "execute_result"
|
||
}
|
||
],
|
||
"source": [
|
||
"tensor2d.view(3, 2) # 查看张量"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 13,
|
||
"id": "344e307f-ba5d-4f9a-a791-2c75a3d1417e",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"data": {
|
||
"text/plain": [
|
||
"tensor([[1, 4],\n",
|
||
" [2, 5],\n",
|
||
" [3, 6]])"
|
||
]
|
||
},
|
||
"execution_count": 13,
|
||
"metadata": {},
|
||
"output_type": "execute_result"
|
||
}
|
||
],
|
||
"source": [
|
||
"tensor2d.T # 转置张量"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 14,
|
||
"id": "19a75030-6a41-4ca8-9aae-c507ae79225c",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"data": {
|
||
"text/plain": [
|
||
"tensor([[14, 32],\n",
|
||
" [32, 77]])"
|
||
]
|
||
},
|
||
"execution_count": 14,
|
||
"metadata": {},
|
||
"output_type": "execute_result"
|
||
}
|
||
],
|
||
"source": [
|
||
"tensor2d.matmul(tensor2d.T) # 张量乘法：tensor2d与其转置相乘"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 15,
|
||
"id": "e7c950bc-d640-4203-b210-3ac8932fe4d4",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"data": {
|
||
"text/plain": [
|
||
"tensor([[14, 32],\n",
|
||
" [32, 77]])"
|
||
]
|
||
},
|
||
"execution_count": 15,
|
||
"metadata": {},
|
||
"output_type": "execute_result"
|
||
}
|
||
],
|
||
"source": [
|
||
"tensor2d @ tensor2d.T # 张量乘法的另一种实现方式：tensor2d与其转置相乘"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "4c15bdeb-78e2-4870-8a4f-a9f591666f38",
|
||
"metadata": {},
|
||
"source": [
|
||
"## A.3 把模型作为计算图"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 16,
|
||
"id": "22af61e9-0443-4705-94d7-24c21add09c7",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"tensor(0.0852)\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"import torch.nn.functional as F\n",
|
||
"\n",
|
||
"y = torch.tensor([1.0]) # 真实样本\n",
|
||
"x1 = torch.tensor([1.1]) # 输入特征\n",
|
||
"w1 = torch.tensor([2.2]) # 权重变量\n",
|
||
"b = torch.tensor([0.0]) # 偏置单元\n",
|
||
"\n",
|
||
"z = x1 * w1 + b # 网络输入\n",
|
||
"a = torch.sigmoid(z) # 激活函数 & 输出\n",
|
||
"\n",
|
||
"loss = F.binary_cross_entropy(a, y)\n",
|
||
"print(loss)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "f9424f26-2bac-47e7-b834-92ece802247c",
|
||
"metadata": {},
|
||
"source": [
|
||
"## A.4 自动求导"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 17,
|
||
"id": "ebf5cef7-48d6-4d2a-8ab0-0fb10bdd7d1a",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"(tensor([-0.0898]),)\n",
|
||
"(tensor([-0.0817]),)\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"import torch.nn.functional as F\n",
|
||
"from torch.autograd import grad\n",
|
||
"\n",
|
||
"y = torch.tensor([1.0])\n",
|
||
"x1 = torch.tensor([1.1])\n",
|
||
"w1 = torch.tensor([2.2], requires_grad=True)\n",
|
||
"b = torch.tensor([0.0], requires_grad=True)\n",
|
||
"\n",
|
||
"z = x1 * w1 + b \n",
|
||
"a = torch.sigmoid(z)\n",
|
||
"\n",
|
||
"loss = F.binary_cross_entropy(a, y)\n",
|
||
"\n",
|
||
"grad_L_w1 = grad(loss, w1, retain_graph=True)\n",
|
||
"grad_L_b = grad(loss, b, retain_graph=True)\n",
|
||
"\n",
|
||
"print(grad_L_w1)\n",
|
||
"print(grad_L_b)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 18,
|
||
"id": "93c5875d-f6b2-492c-b5ef-7e132f93a4e0",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"tensor([-0.0898])\n",
|
||
"tensor([-0.0817])\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"loss.backward()# 反向传播\n",
|
||
"\n",
|
||
"print(w1.grad)\n",
|
||
"print(b.grad)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "f53bdd7d-44e6-40ab-8a5a-4eef74ef35dc",
|
||
"metadata": {},
|
||
"source": [
|
||
"## A.5 多层神经网络的实现"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 19,
|
||
"id": "84b749e1-7768-4cfe-94d6-a08c7feff4a1",
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"class NeuralNetwork(torch.nn.Module):\n",
|
||
" def __init__(self, num_inputs, num_outputs):\n",
|
||
" super().__init__()\n",
|
||
"\n",
|
||
" self.layers = torch.nn.Sequential(\n",
|
||
" \n",
|
||
" # 第一个隐藏层\n",
|
||
" torch.nn.Linear(num_inputs, 30),\n",
|
||
" torch.nn.ReLU(),\n",
|
||
"\n",
|
||
" # 第二个隐藏层\n",
|
||
" torch.nn.Linear(30, 20),\n",
|
||
" torch.nn.ReLU(),\n",
|
||
"\n",
|
||
" # 输出层\n",
|
||
" torch.nn.Linear(20, num_outputs),\n",
|
||
" )\n",
|
||
"\n",
|
||
" def forward(self, x):\n",
|
||
" logits = self.layers(x)\n",
|
||
" return logits"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 20,
|
||
"id": "c5b59e2e-1930-456d-93b9-f69263e3adbe",
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"model = NeuralNetwork(50, 3)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 21,
|
||
"id": "39d02a21-33e7-4879-8fd2-d6309faf2f8d",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"NeuralNetwork(\n",
|
||
" (layers): Sequential(\n",
|
||
" (0): Linear(in_features=50, out_features=30, bias=True)\n",
|
||
" (1): ReLU()\n",
|
||
" (2): Linear(in_features=30, out_features=20, bias=True)\n",
|
||
" (3): ReLU()\n",
|
||
" (4): Linear(in_features=20, out_features=3, bias=True)\n",
|
||
" )\n",
|
||
")\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"print(model)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 22,
|
||
"id": "94535738-de02-4c2a-9b44-1cd186fa990a",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"Total number of trainable model parameters: 2213\n"
|
||
]
|
||
}
|
||
],
|
||
"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)# 打印训练模型的参数"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 23,
|
||
"id": "2c394106-ad71-4ccb-a3c9-9b60af3fa748",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"Parameter containing:\n",
|
||
"tensor([[-0.0064, 0.0004, -0.0903, ..., -0.1316, 0.0910, 0.0363],\n",
|
||
" [ 0.1354, 0.1124, -0.0476, ..., 0.0578, 0.1014, 0.0008],\n",
|
||
" [ 0.0975, -0.0478, 0.0298, ..., 0.0416, 0.0849, 0.1314],\n",
|
||
" ...,\n",
|
||
" [ 0.0118, 0.0240, 0.0420, ..., -0.1305, -0.0517, -0.0826],\n",
|
||
" [-0.0323, 0.1073, 0.0215, ..., -0.1264, -0.1100, 0.1232],\n",
|
||
" [ 0.0861, 0.0403, -0.0545, ..., 0.1352, 0.0817, -0.0938]],\n",
|
||
" requires_grad=True)\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"print(model.layers[0].weight) # 打印神经网络模型的第一层的权重"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 24,
|
||
"id": "b201882b-9285-4db9-bb63-43afe6a2ff9e",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"Parameter containing:\n",
|
||
"tensor([[-0.0577, 0.0047, -0.0702, ..., 0.0222, 0.1260, 0.0865],\n",
|
||
" [ 0.0502, 0.0307, 0.0333, ..., 0.0951, 0.1134, -0.0297],\n",
|
||
" [ 0.1077, -0.1108, 0.0122, ..., 0.0108, -0.1049, -0.1063],\n",
|
||
" ...,\n",
|
||
" [-0.0787, 0.1259, 0.0803, ..., 0.1218, 0.1303, -0.1351],\n",
|
||
" [ 0.1359, 0.0175, -0.0673, ..., 0.0674, 0.0676, 0.1058],\n",
|
||
" [ 0.0790, 0.1343, -0.0293, ..., 0.0344, -0.0971, -0.0509]],\n",
|
||
" requires_grad=True)\n"
|
||
]
|
||
}
|
||
],
|
||
"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)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 25,
|
||
"id": "1da9a35e-44f3-460c-90fe-304519736fd6",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"torch.Size([30, 50])\n"
|
||
]
|
||
}
|
||
],
|
||
"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)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 26,
|
||
"id": "57eadbae-90fe-43a3-a33f-c23a095ba42a",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"tensor([[-0.1262, 0.1080, -0.1792]], grad_fn=<AddmmBackward0>)\n"
|
||
]
|
||
}
|
||
],
|
||
"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)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 27,
|
||
"id": "48d720cb-ef73-4b7b-92e0-8198a072defd",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"tensor([[-0.1262, 0.1080, -0.1792]])\n"
|
||
]
|
||
}
|
||
],
|
||
"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)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 28,
|
||
"id": "10df3640-83c3-4061-a74d-08f07a5cc6ac",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"tensor([[0.3113, 0.3934, 0.2952]])\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"with torch.no_grad():\n",
|
||
" out = torch.softmax(model(X), dim=1)\n",
|
||
"print(out)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "19858180-0f26-43a8-b2c3-7ed40abf9f85",
|
||
"metadata": {},
|
||
"source": [
|
||
"## A.6 建立高效的数据加载器"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 29,
|
||
"id": "b9dc2745-8be8-4344-80ef-325f02cda7b7",
|
||
"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",
|
||
" [-0.5, 2.6],\n",
|
||
" [2.3, -1.1],\n",
|
||
" [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])"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 30,
|
||
"id": "88283948-5fca-461a-98a1-788b6be191d5",
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"X_test = torch.tensor([\n",
|
||
" [-0.8, 2.8],\n",
|
||
" [2.6, -1.6],\n",
|
||
"])\n",
|
||
"\n",
|
||
"y_test = torch.tensor([0, 1])"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 31,
|
||
"id": "edf323e2-1789-41a0-8e44-f3cab16e5f5d",
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"from torch.utils.data import Dataset\n",
|
||
"\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",
|
||
" # 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",
|
||
" # 获取数据集的长度\n",
|
||
" def __len__(self):\n",
|
||
" return self.labels.shape[0]\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)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 32,
|
||
"id": "b7014705-1fdc-4f72-b892-d8db8bebc331",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"data": {
|
||
"text/plain": [
|
||
"5"
|
||
]
|
||
},
|
||
"execution_count": 32,
|
||
"metadata": {},
|
||
"output_type": "execute_result"
|
||
}
|
||
],
|
||
"source": [
|
||
"len(train_ds)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 33,
|
||
"id": "3ec6627a-4c3f-481a-b794-d2131be95eaf",
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"from torch.utils.data import DataLoader\n",
|
||
"\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",
|
||
" shuffle=True,\n",
|
||
" num_workers=0\n",
|
||
")"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 34,
|
||
"id": "8c9446de-5e4b-44fa-bf9a-a63e2661027e",
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"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",
|
||
" shuffle=False,\n",
|
||
" num_workers=0\n",
|
||
")"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 35,
|
||
"id": "99d4404c-9884-419f-979c-f659742d86ef",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"Batch 1: tensor([[ 2.3000, -1.1000],\n",
|
||
" [-0.9000, 2.9000]]) tensor([1, 0])\n",
|
||
"Batch 2: tensor([[-1.2000, 3.1000],\n",
|
||
" [-0.5000, 2.6000]]) tensor([0, 0])\n",
|
||
"Batch 3: tensor([[ 2.7000, -1.5000]]) tensor([1])\n"
|
||
]
|
||
}
|
||
],
|
||
"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)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 36,
|
||
"id": "9d003f7e-7a80-40bf-a7fb-7a0d7dbba9db",
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"train_loader = DataLoader(\n",
|
||
" dataset=train_ds,\n",
|
||
" batch_size=2,\n",
|
||
" shuffle=True,\n",
|
||
" num_workers=0,\n",
|
||
" drop_last=True\n",
|
||
")"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "4db4d7f4-82da-44a4-b94e-ee04665d9c3c",
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"for idx, (x, y) in enumerate(train_loader):\n",
|
||
" print(f\"Batch {idx+1}:\", x, y)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "d904ca82-e50f-4f3d-a3ac-fc6ca53dd00e",
|
||
"metadata": {},
|
||
"source": [
|
||
"## A.7 一个示例训练轮次"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 38,
|
||
"id": "93f1791a-d887-4fc5-a307-5e5bde9e06f6",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"Epoch: 001/003 | Batch 000/002 | Train/Val Loss: 0.75\n",
|
||
"Epoch: 001/003 | Batch 001/002 | Train/Val Loss: 0.65\n",
|
||
"Epoch: 002/003 | Batch 000/002 | Train/Val Loss: 0.44\n",
|
||
"Epoch: 002/003 | Batch 001/002 | Train/Val Loss: 0.13\n",
|
||
"Epoch: 003/003 | Batch 000/002 | Train/Val Loss: 0.03\n",
|
||
"Epoch: 003/003 | Batch 001/002 | Train/Val Loss: 0.00\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"import torch.nn.functional as F\n",
|
||
"\n",
|
||
"\n",
|
||
"torch.manual_seed(123)\n",
|
||
"model = NeuralNetwork(num_inputs=2, num_outputs=2)\n",
|
||
"optimizer = torch.optim.SGD(model.parameters(), lr=0.5)\n",
|
||
"\n",
|
||
"num_epochs = 3\n",
|
||
"\n",
|
||
"for epoch in range(num_epochs):\n",
|
||
" \n",
|
||
" model.train()\n",
|
||
" for batch_idx, (features, labels) in enumerate(train_loader):\n",
|
||
"\n",
|
||
" logits = model(features)\n",
|
||
" \n",
|
||
" loss = F.cross_entropy(logits, labels) # 损失函数\n",
|
||
" \n",
|
||
" optimizer.zero_grad()\n",
|
||
" loss.backward()\n",
|
||
" optimizer.step()\n",
|
||
" \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",
|
||
" # 可选的模型评估指标"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 39,
|
||
"id": "00dcf57f-6a7e-4af7-aa5a-df2cb0866fa5",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"tensor([[ 2.8569, -4.1618],\n",
|
||
" [ 2.5382, -3.7548],\n",
|
||
" [ 2.0944, -3.1820],\n",
|
||
" [-1.4814, 1.4816],\n",
|
||
" [-1.7176, 1.7342]])\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"model.eval()\n",
|
||
"\n",
|
||
"with torch.no_grad():\n",
|
||
" outputs = model(X_train)\n",
|
||
"\n",
|
||
"print(outputs)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 40,
|
||
"id": "19be7390-18b8-43f9-9841-d7fb1919f6fd",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"tensor([[ 0.9991, 0.0009],\n",
|
||
" [ 0.9982, 0.0018],\n",
|
||
" [ 0.9949, 0.0051],\n",
|
||
" [ 0.0491, 0.9509],\n",
|
||
" [ 0.0307, 0.9693]])\n",
|
||
"tensor([0, 0, 0, 1, 1])\n"
|
||
]
|
||
}
|
||
],
|
||
"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)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 41,
|
||
"id": "07e7e530-f8d3-429c-9f5e-cf8078078c0e",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"tensor([0, 0, 0, 1, 1])\n"
|
||
]
|
||
}
|
||
],
|
||
"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)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 42,
|
||
"id": "5f756f0d-63c8-41b5-a5d8-01baa847e026",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"data": {
|
||
"text/plain": [
|
||
"tensor([True, True, True, True, True])"
|
||
]
|
||
},
|
||
"execution_count": 42,
|
||
"metadata": {},
|
||
"output_type": "execute_result"
|
||
}
|
||
],
|
||
"source": [
|
||
"predictions == y_train"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 43,
|
||
"id": "da274bb0-f11c-4c81-a880-7a031fbf2943",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"data": {
|
||
"text/plain": [
|
||
"tensor(5)"
|
||
]
|
||
},
|
||
"execution_count": 43,
|
||
"metadata": {},
|
||
"output_type": "execute_result"
|
||
}
|
||
],
|
||
"source": [
|
||
"torch.sum(predictions == y_train)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 44,
|
||
"id": "16d62314-8dee-45b0-8f55-9e5aae2b24f4",
|
||
"metadata": {},
|
||
"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()"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 45,
|
||
"id": "4f6c9c17-2a5f-46c0-804b-873f169b729a",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"data": {
|
||
"text/plain": [
|
||
"1.0"
|
||
]
|
||
},
|
||
"execution_count": 45,
|
||
"metadata": {},
|
||
"output_type": "execute_result"
|
||
}
|
||
],
|
||
"source": [
|
||
"compute_accuracy(model, train_loader)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 46,
|
||
"id": "311ed864-e21e-4aac-97c7-c6086caef27a",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"data": {
|
||
"text/plain": [
|
||
"1.0"
|
||
]
|
||
},
|
||
"execution_count": 46,
|
||
"metadata": {},
|
||
"output_type": "execute_result"
|
||
}
|
||
],
|
||
"source": [
|
||
"compute_accuracy(model, test_loader)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "4d5cd469-3a45-4394-944b-3ce543f41dac",
|
||
"metadata": {},
|
||
"source": [
|
||
"## A.8 保存并加载模型"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 47,
|
||
"id": "b013127d-a2c3-4b04-9fb3-a6a7c88d83c5",
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": [
|
||
"torch.save(model.state_dict(), \"model.pth\")"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 48,
|
||
"id": "b2b428c2-3a44-4d91-97c4-8298cf2b51eb",
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"data": {
|
||
"text/plain": [
|
||
"<All keys matched successfully>"
|
||
]
|
||
},
|
||
"execution_count": 48,
|
||
"metadata": {},
|
||
"output_type": "execute_result"
|
||
}
|
||
],
|
||
"source": [
|
||
"model = NeuralNetwork(2, 2) # 需要与原始模型完全匹配\n",
|
||
"model.load_state_dict(torch.load(\"model.pth\"))"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "f891c013-43da-4a05-973d-997be313d2d8",
|
||
"metadata": {},
|
||
"source": [
|
||
"## A.9 使用GPU来优化训练性能"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "e68ae888-cabf-49c9-bad6-ecdce774db57",
|
||
"metadata": {},
|
||
"source": [
|
||
"### A.9.1 在GPU上进行 PyTorch 的运算"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "141c845f-efe3-4614-b376-b8b7a9a2c887",
|
||
"metadata": {},
|
||
"source": [
|
||
"See [code-part2.ipynb](code-part2.ipynb) "
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "99811829-b817-42ea-b03e-d35374debcc0",
|
||
"metadata": {},
|
||
"source": [
|
||
"### A.9.2 单个GPU的训练"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "0b21456c-4af7-440f-9e78-37770277b5bc",
|
||
"metadata": {},
|
||
"source": [
|
||
"See [code-part2.ipynb](code-part2.ipynb)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "db6eb2d1-a341-4489-b04b-635c26945333",
|
||
"metadata": {},
|
||
"source": [
|
||
"### A.9.3 多GPU的训练"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"id": "9d049a81-5fb0-49b5-9d6a-17a9976d8520",
|
||
"metadata": {},
|
||
"source": [
|
||
"See [DDP-script.py](DDP-script.py)"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"id": "b32db94f-f159-4aa3-91cc-5b937eef7fc7",
|
||
"metadata": {},
|
||
"outputs": [],
|
||
"source": []
|
||
}
|
||
],
|
||
"metadata": {
|
||
"kernelspec": {
|
||
"display_name": "Python 3 (ipykernel)",
|
||
"language": "python",
|
||
"name": "python3"
|
||
},
|
||
"language_info": {
|
||
"codemirror_mode": {
|
||
"name": "ipython",
|
||
"version": 3
|
||
},
|
||
"file_extension": ".py",
|
||
"mimetype": "text/x-python",
|
||
"name": "python",
|
||
"nbconvert_exporter": "python",
|
||
"pygments_lexer": "ipython3",
|
||
"version": "3.10.6"
|
||
}
|
||
},
|
||
"nbformat": 4,
|
||
"nbformat_minor": 5
|
||
}
|