How to log PyTorch training metrics to TensorBoard

Scalar histories reveal whether an optimizer is converging, stalling, or diverging across epochs. PyTorch can write those values as TensorBoard event data, replacing transient console lines with run data that TensorBoard can read and compare.

The SummaryWriter creates event files asynchronously under its log_dir. A separate directory for each experiment keeps runs distinct, while hierarchical tags such as Loss/train and MAE/train group related curves in the Scalars dashboard.

A small deterministic regression model makes every recorded value come from a real optimization step rather than fixed sample data. Closing the writer commits queued events, and a successful TensorBoard inspection lists both scalar tags with five recorded steps.

1. Install TensorBoard in the active PyTorch environment.

   $ python -m pip install tensorboard

   Related: How to install PyTorch with pip

2. Create the foundation of tensorboard_metrics.py through its SummaryWriter construction.

   tensorboard_metrics.py

   import torch
   from torch import nn
   from torch.utils.tensorboard import SummaryWriter
    
    
   torch.manual_seed(7)
   features = torch.arange(-5, 5, 0.1).view(-1, 1)
   targets = -5 * features + 0.1 * torch.randn(features.size())
   model = nn.Linear(1, 1)
   loss_fn = nn.MSELoss()
   optimizer = torch.optim.SGD(model.parameters(), lr=0.01)
   writer = SummaryWriter("runs/linear-regression")

3. Append the training loop below the SummaryWriter construction in tensorboard_metrics.py.

   for epoch in range(5):
       predictions = model(features)
       loss = loss_fn(predictions, targets)
       mean_absolute_error = (predictions - targets).abs().mean()
       optimizer.zero_grad()
       loss.backward()
       optimizer.step()
       writer.add_scalar("Loss/train", loss.item(), epoch)
       writer.add_scalar("MAE/train", mean_absolute_error.item(), epoch)
       print(
           f"epoch={epoch} loss={loss.item():.4f} "
           f"mae={mean_absolute_error.item():.4f}"
       )
    
   writer.close()

   The epoch number is the global_step for both scalar series, so their points align on the same horizontal scale.

4. Run tensorboard_metrics.py from its project directory.

   $ python tensorboard_metrics.py
   epoch=0 loss=166.7248 mae=11.1799
   epoch=1 loss=115.8200 mae=9.3178
   epoch=2 loss=80.4710 mae=7.7664
   epoch=3 loss=55.9236 mae=6.4736
   epoch=4 loss=38.8768 mae=5.3963

5. Run TensorBoard's inspection on runs/linear-regression to confirm both scalar series contain five steps.

   $ tensorboard --inspect --logdir runs/linear-regression
   TensorFlow installation not found - running with reduced feature set.
   ======================================================================
   Processing event files... (this can take a few minutes)
   ======================================================================
   
   Found event files in:
   runs/linear-regression
   
   These tags are in runs/linear-regression:
   audio -
   histograms -
   images -
   scalars
      Loss/train
      MAE/train
   tensor -
   ======================================================================
   
   Event statistics for runs/linear-regression:
   audio -
   graph -
   histograms -
   images -
   scalars
      first_step           0
      last_step            4
      max_step             4
      min_step             0
      num_steps            5
      outoforder_steps     []
   sessionlog:checkpoint -
   sessionlog:start -
   sessionlog:stop -
   tensor -
   ======================================================================

   The reduced-feature message only indicates that TensorFlow is absent; the scalars tags and num_steps value come from the PyTorch event file.