Embodied AI Glossary中文TensorBoard
CommonA local, browser-based tool for viewing training curves, images, and other logged data.
TensorBoard is an open-source training-visualization tool released by Google alongside TensorFlow, though PyTorch can also write logs to it directly through torch.utils.tensorboard. During training, a program writes scalars like loss, reward, and learning rate, along with images and histograms, to a log directory; running tensorboard --logdir then shows live curves in a browser and lets you compare multiple runs. It runs entirely locally and needs no account, which is why many robot reinforcement-learning frameworks — such as rsl_rl and Isaac Lab's training scripts — log to it by default; teams that need cloud collaboration and experiment management often switch to Weights & Biases or SwanLab instead.
ExampleWhile training a quadruped walking policy in Isaac Lab, open TensorBoard to watch whether the average reward and episode-length curves are trending up, as a sign that training is converging.
- Also called
- tensorboard
- Related
- Weights & Biases (W&B) · SwanLab · Loss Function · rsl_rl · TensorFlow · PyTorch
- Sources
- TensorBoard 官方页面 (Chinese)
tensorflow/tensorboard (GitHub)