How to log PyTorch training metrics to TensorBoard

Training loops are easier to compare when loss and accuracy are written as event data instead of disappearing into terminal output. PyTorch sends scalar metrics to TensorBoard through SummaryWriter, so a run directory can show both the metric names and the values recorded at each training step.

SummaryWriter writes TensorBoard event files under its log_dir and flushes entries asynchronously while training continues. Giving each run its own directory, such as runs/tensorboard-demo or runs/experiment-name, keeps the Scalars dashboard organized and prevents unrelated runs from sharing one event file.

Close or flush the writer before inspecting a short smoke run so pending metric points reach disk. The same logging calls can then move into the real training loop, where global_step should come from the epoch, batch index, or optimizer step that the metric represents.

Steps to log PyTorch training metrics to TensorBoard:

Install TensorBoard in the active PyTorch environment.
```
$ python -m pip install tensorboard
```
Run this inside the same virtual environment or Conda environment that imports torch.
Related: How to install PyTorch with pip

Create a SummaryWriter smoke script.

tensorboard_train_metrics.py

from pathlib import Path
import shutil
 
import torch
from torch import nn
from torch.utils.tensorboard import SummaryWriter
from tensorboard.backend.event_processing import event_accumulator
 
 
torch.manual_seed(19)
log_dir = Path("runs/tensorboard-demo")
if log_dir.exists():
    shutil.rmtree(log_dir)
 
features = torch.tensor(
    [
        [-1.0, 0.2, 0.1],
        [-0.5, 0.5, 0.4],
        [0.0, -0.3, 0.9],
        [0.5, 0.7, -0.4],
        [1.0, -0.6, -0.1],
        [1.5, 0.4, 0.2],
    ],
    dtype=torch.float32,
)
targets = torch.tensor([0, 0, 1, 1, 1, 0])
 
model = nn.Sequential(
    nn.Linear(3, 6),
    nn.ReLU(),
    nn.Linear(6, 2),
)
loss_fn = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr=0.15)
writer = SummaryWriter(log_dir=log_dir)
 
for epoch in range(5):
    optimizer.zero_grad(set_to_none=True)
    logits = model(features)
    loss = loss_fn(logits, targets)
    loss.backward()
    optimizer.step()
 
    with torch.no_grad():
        predictions = model(features).argmax(dim=1)
        accuracy = (predictions == targets).float().mean().item()
 
    writer.add_scalar("Loss/train", loss.item(), epoch)
    writer.add_scalar("Accuracy/train", accuracy, epoch)
    print(f"epoch={epoch} loss={loss.item():.4f} accuracy={accuracy:.4f}")
 
writer.flush()
writer.close()
 
events = sorted(log_dir.glob("events.out.tfevents.*"))
reader = event_accumulator.EventAccumulator(str(log_dir))
reader.Reload()
scalar_tags = sorted(reader.Tags()["scalars"])
loss_points = reader.Scalars("Loss/train")
accuracy_points = reader.Scalars("Accuracy/train")
 
print(f"torch_version={torch.__version__}")
print(f"log_dir={log_dir}")
print(f"event_files={len(events)}")
for path in events:
    print(f"event_file={path.name}")
print(f"scalar_tags={','.join(scalar_tags)}")
print(f"loss_points={len(loss_points)}")
print(f"accuracy_points={len(accuracy_points)}")
print(f"last_loss={loss_points[-1].value:.4f}")
print(f"last_accuracy={accuracy_points[-1].value:.4f}")

EventAccumulator is used only for the terminal readback. Training code only needs SummaryWriter and the add_scalar() calls.

Run the smoke script.

$ python tensorboard_train_metrics.py
epoch=0 loss=0.7052 accuracy=0.5000
epoch=1 loss=0.6980 accuracy=0.5000
epoch=2 loss=0.6919 accuracy=0.5000
epoch=3 loss=0.6870 accuracy=0.5000
epoch=4 loss=0.6823 accuracy=0.5000
torch_version=2.12.1+cpu
log_dir=runs/tensorboard-demo
event_files=1
event_file=events.out.tfevents.1783379917.training-host.2508.0
scalar_tags=Accuracy/train,Loss/train
loss_points=5
accuracy_points=5
last_loss=0.6823
last_accuracy=0.5000

event_files=1 confirms that SummaryWriter created an event file. scalar_tags confirms that both metric names were written.

Verify the event directory with TensorBoard.

$ tensorboard --inspect --logdir runs/tensorboard-demo
TensorFlow installation not found - running with reduced feature set.
======================================================================
Processing event files... (this can take a few minutes)
======================================================================

Found event files in:
runs/tensorboard-demo

These tags are in runs/tensorboard-demo:
audio -
histograms -
images -
scalars
   Accuracy/train
   Loss/train
tensor -
======================================================================

Event statistics for runs/tensorboard-demo:
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 is normal when TensorBoard is installed without TensorFlow. The scalars section and num_steps count are the event readback signals.

Add the same writer pattern to the project training loop.

writer = SummaryWriter(log_dir=f"runs/{run_name}")
 
for global_step, (inputs, targets) in enumerate(train_loader):
    loss, accuracy = train_one_batch(model, inputs, targets)
    writer.add_scalar("Loss/train", loss, global_step)
    writer.add_scalar("Accuracy/train", accuracy, global_step)
 
writer.close()

Log Python floats or scalar tensors after each value is computed. Keep tag names stable across runs so TensorBoard can overlay comparable curves.

Start TensorBoard with the parent run directory.
```
$ tensorboard --logdir runs
```
Open the local URL printed by TensorBoard and choose the tensorboard-demo run in the Scalars dashboard. Point --logdir at the parent directory when several run subdirectories should be compared.
Remove the smoke script and demo run directory after moving the logging calls into the project.
```
$ rm -r tensorboard_train_metrics.py runs/tensorboard-demo
```