tensorboardX

You can install tensorboardX via pip or build it from the source repository.

To improve performance, it is recommended to install crc32c. Additionally, for the add_audio() function, you must install soundfile (which provides a 200x speedup).

You can run the project's test suite using the provided helper script or manually via pytest. If using uv, ensure you have pinned setuptools as described in the environment setup.

# Quick Start using the helper script
./run_pytest.sh

# Manual testing with uv
uv pip install "setuptools==81.0.0"
uv run pytest

To use tensorboardX, install it via pip. To visualize the logs, you must also install the tensorboard web server.

# Install tensorboardX
pip install tensorboardX

# Install the tensorboard web server
pip install tensorboard

The SummaryWriter class is the primary interface for writing TensorBoard events. It supports various summary types including scalar, image, figure, histogram, audio, text, graph, onnx_graph, embedding, pr_curve, mesh, hyper-parameters, and video.

Key usage patterns:

• Data Grouping: Use forward slashes (/) in tag names to group data in the TensorBoard UI (e.g., 'data/scalar1').
• Multiple Scalars: Use add_scalars to plot multiple series in the same chart.
• Audio Requirements: When using add_audio(), ensure the audio amplitude is within the range [-1, 1] and install soundfile for optimal performance.
• PR Curves: add_pr_curve requires TensorBoard version 0.4RC or later.

TensorBoard reads the event files written by tensorboardX. You must have tensorboard installed to view the results.

1. Install TensorBoard: pip install tensorboard (the non-GPU version is recommended for faster startup if you don't need GPU support).
2. Launch the server: Run the following command in your terminal, pointing to your log directory:

tensorboard --logdir=runs

3. Open Browser: Follow the terminal instructions to open the web interface in your browser.

To avoid ModuleNotFoundError caused by the removal of pkg_resources in newer versions of setuptools, you must pin setuptools to version 81.0.0 or earlier during local development and testing. The project uses setuptools_scm to derive its version dynamically from the most recent git tag.

uv pip install "setuptools==81.0.0"

Install tensorboardX using pip via the command line.

pip install tensorboardX

To record data, you must first create a SummaryWriter instance. The SummaryWriter class contains all the functionality of the package.

• With a specific directory: SummaryWriter('runs/exp-1') saves data to runs/exp-1.
• With default settings: SummaryWriter() saves data to a directory named after the current time and machine name (e.g., runs/Aug20-obov01).
• With a comment: SummaryWriter(comment='3xLR') appends a comment to the default directory name (e.g., runs/Aug20-obov01-3xLR).

Each subdirectory within the runs folder is treated as a separate experiment. It is recommended to use unique names (like timestamps or parameter names) for each experiment to facilitate comparison.

Use writer.add_scalar(tag_name, value, iteration) to log scalar data like loss or accuracy.

Note: If you are using PyTorch, do not pass a tensor directly; extract the scalar value using .item() to avoid errors.