🚀 PyTorch与视觉检测原理分析

import torch

# 创建张量
x = torch.Tensor([[1, 2, 3], [4, 5, 6], [7, 8, 9]])

# GPU加速
x = x.cuda()  # 将张量移到GPU

# 自动求导
x = torch.tensor(2.0, requires_grad=True)
y = x**2
y.backward()  # 自动计算梯度 dy/dx = 4

# PyTorch卷积层定义
import torch.nn as nn

# 输入通道=3(RGB), 输出通道=64, 卷积核=3x3
conv = nn.Conv2d(in_channels=3, 
                 out_channels=64, 
                 kernel_size=3, 
                 stride=1, 
                 padding=1)

# 参数说明：
# - in_channels: 输入图像通道数（RGB=3, 灰度=1）
# - out_channels: 输出特征图数量（学习64种不同特征）
# - kernel_size: 卷积核大小（3表示3x3）
# - stride: 步长（每次移动的像素数）
# - padding: 填充（保持输出尺寸）

# YOLO输出张量计算
S = 7  # 网格大小 7x7
B = 2  # 每个网格预测2个边界框
C = 20 # PASCAL VOC数据集有20个类别

# 每个边界框：(x, y, w, h, confidence) = 5个值
# 每个网格：B个边界框 + C个类别概率
output_size = S * S * (B * 5 + C)
# = 7 * 7 * (2 * 5 + 20) = 7 * 7 * 30 = 1470

print(f"YOLO-v1输出张量大小: {S}×{S}×{B*5+C} = {output_size}")

import torch
import torch.nn as nn
import torch.optim as optim

# Step 1: 定义网络结构
class SimpleNet(nn.Module):
    def __init__(self):
        super(SimpleNet, self).__init__()
        self.fc1 = nn.Linear(10, 50)
        self.fc2 = nn.Linear(50, 1)
    
    def forward(self, x):
        x = torch.relu(self.fc1(x))
        x = self.fc2(x)
        return x

# Step 2: 初始化模型
model = SimpleNet()
criterion = nn.MSELoss()
optimizer = optim.Adam(model.parameters(), lr=0.001)

# Step 3: 训练循环
for epoch in range(100):
    # 3.1 前向传播
    outputs = model(inputs)
    
    # 3.2 计算损失
    loss = criterion(outputs, targets)
    
    # 3.3 反向传播
    optimizer.zero_grad()  # 清空梯度
    loss.backward()        # 计算梯度
    
    # 3.4 更新权重
    optimizer.step()

# DataParallel - 简单但效率较低
model = nn.DataParallel(model, device_ids=[0, 1, 2, 3])

# DistributedDataParallel - 高效分布式训练
import torch.distributed as dist
dist.init_process_group(backend='nccl')
model = nn.parallel.DistributedDataParallel(model, 
                                            device_ids=[local_rank])

# PyTorch Lightning - 自动化分布式训练
from pytorch_lightning import Trainer
trainer = Trainer(accelerator="gpu", 
                  devices=4, 
                  strategy="ddp")
trainer.fit(model)