pytorch中tensorboard数据显示

ゝ一纸荒年。 2023-02-18 11:53 17阅读 0赞

## **pytorch** 利用**tensorboard**显示**loss，acc**曲线等 ##

--------------------

## 运行环境： ##

python3.6.9  
pytorch1.13.1  
cuda10.0  
cudnn7.5.1

--------------------

## tensorboard显示 ##

运行**PointRCNN**算法进行**training**，得出`events.out.tfevents.1592297776.hkd-Precision-7920-Tower`

打开终端输入：`tensorboard --logdir path/to/tensorboard_logs/`  
会有输出：`TensorBoard 1.6.0 at http://iccd:6006 (Press CTRL+C to quit)`  
将上述链接复制到**浏览器**中打开便可以显示该训练参数（tensorboard）

![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L1JOR191emlf_size_16_color_FFFFFF_t_70]

--------------------

## tensorboard记录 ##

from tensorboard_logger import Logger
    
    logger = Logger(logdir="./tensorboard_logs", flush_secs=10)
    ...
    def train(net, optimizer):
        for epoch in range(epoch_nums):
            net.train()
            for batch_idx, (inputs, targets) in enumerate(trainloader):           
                inputs = Variable(inputs, requires_grad=True).cuda()
                targets = targets.cuda()
                optimizer.zero_grad()
                outputs = net(inputs)
                loss = criterion(outputs, targets)
                loss.backward()  
                optimizer.step()
                train_loss += loss.item()
                ...
                # 记录所需的变量
                logger.log_value('avg_loss', train_loss/(batch_idx+1), epoch*len(trainloader) + batch_idx)
                logger.log_value('loss', loss.item(), epoch*len(trainloader) + batch_idx)
                logger.log_value('acc', 100. * correct / total, epoch*len(trainloader) + batch_idx)

[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L1JOR191emlf_size_16_color_FFFFFF_t_70]: https://img-blog.csdnimg.cn/20200617170613885.png?x-oss-process=image/watermark,type_ZmFuZ3poZW5naGVpdGk,shadow_10,text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L1JOR191emlf,size_16,color_FFFFFF,t_70