tensorboard

Create a SummaryWriter from torch.utils.tensorboard with a log directory, then call writer.add_scalar with a tag, value, and step inside your training loop. Launch the dashboard with tensorboard --logdir=runs and open http://localhost:6006.

Write each run to a separate subdirectory under a common log directory, such as runs/exp1 and runs/exp2. Launch TensorBoard on the parent directory and all runs appear together, where you can toggle, filter with regex, and overlay charts.

Yes, set report_to='tensorboard' and logging_dir in TrainingArguments, and the Trainer automatically logs metrics. You can also attach a custom TrainerCallback with a SummaryWriter for manual logging of additional values.

Yes, use torch.profiler with the tensorboard_trace_handler for PyTorch, or set profile_batch in the Keras TensorBoard callback for TensorFlow. The Profile tab shows GPU utilization, kernel stats, memory usage, and operation timelines.

Logging every batch creates excessive event files. Log epoch-level metrics every epoch and batch-level metrics only periodically, such as every 100 batches, and log histograms every few epochs instead of every step.

Call writer.add_embedding with your embedding tensor, optional metadata labels, and label images. Open the Projector tab to explore the data with PCA, t-SNE, or UMAP, and color or filter points by their metadata labels.