pytorchtensorboard使用的简单教程

1.编写

先初始化

from torch.utils.tensorboard import SummaryWriter
writer = SummaryWriter(comment=f&＃39;LR_{lr}_BS_{batch_size}_SCALE_{img_scale}&＃39;)

散点图

writer.add_scalar(&＃39;Loss/train&＃39;, loss.item(), global_step) #loss

直方图

writer.add_histogram(&＃39;weights/&＃39; + tag, value.data.cpu().numpy(), global_step) #weight

绘图

writer.add_images(&＃39;masks/true&＃39;, true_masks, global_step)
writer.add_images(&＃39;masks/pred&＃39;, torch.sigmoid(masks_pred) > 0.5, global_step)

2.查看

tensorboard --logdir=runs

这里的runs为tensorboard保存log的目录，默认为当前文件夹下的runs文件夹

然后进入http://localhost:6006/ 即可

3.删除log

ctrl c 关闭命令行的tensorboard，然后删除log文件再重新启动tensorboard。

4.详细样例如下,见pytorch-unet：

from torch.utils.tensorboard import SummaryWriter
writer = SummaryWriter(comment=f&＃39;LR_{lr}_BS_{batch_size}_SCALE_{img_scale}&＃39;)for epoch in range(epochs):net.train()epoch_loss = 0with tqdm(total=n_train, desc=f&＃39;Epoch {epoch + 1}/{epochs}&＃39;, unit=&＃39;img&＃39;) as pbar:for batch in train_loader:imgs = batch[&＃39;image&＃39;]true_masks = batch[&＃39;mask&＃39;]assert imgs.shape[1] == net.n_channels, \f&＃39;Network has been defined with {net.n_channels} input channels, &＃39; \f&＃39;but loaded images have {imgs.shape[1]} channels. Please check that &＃39; \&＃39;the images are loaded correctly.&＃39;imgs = imgs.to(device=device, dtype=torch.float32)mask_type = torch.float32 if net.n_classes == 1 else torch.longtrue_masks = true_masks.to(device=device, dtype=mask_type)masks_pred = net(imgs)loss = criterion(masks_pred, true_masks[:, 0, :, :].unsqueeze(1))epoch_loss += loss.item()writer.add_scalar(&＃39;Loss/train&＃39;, loss.item(), global_step)pbar.set_postfix(**{&＃39;loss (batch)&＃39;: loss.item()})optimizer.zero_grad()loss.backward()nn.utils.clip_grad_value_(net.parameters(), 0.1)optimizer.step()pbar.update(imgs.shape[0])global_step += 1if global_step % (n_train // (3 * batch_size)) == 0:for tag, value in net.named_parameters():tag = tag.replace(&＃39;.&＃39;, &＃39;/&＃39;)writer.add_histogram(&＃39;weights/&＃39; + tag, value.data.cpu().numpy(), global_step)writer.add_histogram(&＃39;grads/&＃39; + tag, value.grad.data.cpu().numpy(), global_step)val_score = eval_net(net, val_loader, device)if global_step > 500:scheduler.step(val_score)writer.add_scalar(&＃39;learning_rate&＃39;, optimizer.param_groups[0][&＃39;lr&＃39;], global_step)if net.n_classes > 1:logging.info(&＃39;Validation cross entropy: {}&＃39;.format(val_score))writer.add_scalar(&＃39;Loss/test&＃39;, val_score, global_step)else:logging.info(&＃39;Validation Dice Coeff: {}&＃39;.format(val_score))writer.add_scalar(&＃39;Dice/test&＃39;, val_score, global_step)writer.add_images(&＃39;images&＃39;, imgs, global_step)if net.n_classes == 1:writer.add_images(&＃39;masks/true&＃39;, true_masks, global_step)writer.add_images(&＃39;masks/pred&＃39;, torch.sigmoid(masks_pred) > 0.5, global_step)writer.close()