TensorBoard

TensorBoard

Stand-alone tool
Managed service via tensorboard.dev

Open-source
Managed service via tensorboard.dev

Deploy the open-source version wherever you want

N/A (open source only)

None

None

None

None

TensorBoard parses the .logs folder and displays it in the serverless browser UI
You need a filesystem for the .logs folder and a browser for the UI
You may need to configure ports and addresses to make it work in your setup

TensorBoard UI can get slow when you display/compare many experiments
You can organize metadata inside an experiment in a hierarchical structure with the '/' syntax
There are no utilities to easily access model training metadata from other tools

It doesn't do this nativly
Use other tools from the Tensorflow Extended (TFX) ecosystem for this

Supports:
Live monitoring
Metrics, losses, hyperparameters and text (renders markdown)
Images and image series. They can be converted from other image formats (.png, matplotlib) and have descriptions
Graphs for model architecture and flow
Interactive histograms and distribution visualizations of gradients, weights and activations in each layer

TensorBoard includes additional debugging functionalities:
Profiling tool for hardware consumption optimization
Embedding projector for visualizing feature representations
What if a tool for debugging prediction sensitivity
Fairness Indicators for auditing models for fairness

Compare experiments with:
Overlayed learning curves + table with all metrics
Parallel coordinates plot

You can:
Search experiments by tag with a regexp and add them to compare
Organize experiments in subfolders of the .logs folder to make the UI experience better

You cannot:
Filter by parameters or results
Customize the UI to display experiment metadata in any way

Open Source version of TensorBoard not possible
Can share your TensorBoard graphs with TensorBoard.dev
Use other tools from the Tensorflow Extended (TFX) ecosystem for that

Use other tools from the Tensorflow Extended (TFX) ecosystem for that

TensorBoard is probably the most integrated experiment monitoring tool out there
Pretty much any model training library will have an integration with it