TensorBoard

TensorBoard is a visualization toolkit from TensorFlow for inspecting model training metrics, graph structures, and weight histograms. Operators encounter it when training or fine-tuning models locally—it logs scalar values (loss, accuracy), images, and computational graphs to a directory, then serves a web UI at localhost:6006. While TensorFlow-centric, it works with PyTorch via torch.utils.tensorboard and with Hugging Face Transformers through Trainer callbacks. For local-AI operators, TensorBoard helps debug training runs, spot overfitting, and compare experiments without cloud dependencies.

TensorBoard reads event files written by a SummaryWriter. During training, the writer logs scalars (e.g., loss per step), histograms of weight distributions, embeddings, and even audio/video. The TensorBoard server scans the log directory and renders interactive plots. Key panels: Scalars (time-series of metrics), Graphs (model architecture), Distributions/Histograms (weight evolution), and Projector (PCA/t-SNE of embeddings). Operators fine-tuning models like Llama or Mistral with Hugging Face's Trainer can enable TensorBoard via report_to="tensorboard". The logs consume disk space—a typical fine-tuning run might write 100–500 MB of events. TensorBoard's main alternative is Weights & Biases (cloud), but TensorBoard is fully offline, which suits local-AI workflows.

An operator fine-tunes Llama 3.1 8B on a custom dataset using Hugging Face Transformers. They set TrainingArguments(report_to="tensorboard", logging_dir="./logs"). During training, loss and learning rate are logged every 10 steps. After training, they run tensorboard --logdir ./logs and open http://localhost:6006 to view the loss curve. If the loss plateaus early, they adjust the learning rate and restart. The log directory grows to ~200 MB over 1000 steps.

In a local fine-tuning workflow with transformers and accelerate, the operator adds from torch.utils.tensorboard import SummaryWriter and instantiates writer = SummaryWriter(log_dir="runs/experiment1"). Inside the training loop, they call writer.add_scalar("Loss/train", loss, step). After training, they launch tensorboard --logdir runs --port 6006 in a terminal. The browser shows real-time plots. If using Trainer, they simply pass report_to="tensorboard" in TrainingArguments and the logging happens automatically.