TensorBoard Visualizer

A Streamlit-based interactive app to visualize multiple TensorBoard runs grouped by user-defined tags extracted via regex. Supports aggregation (mean, median), individual run plotting, and interactive Plotly charts.

Features

• Scan directories recursively for TensorBoard event files.
• Extract tags from file paths using regex with named groups.
• Group runs by tags and aggregate metrics.
• Plot metrics with mean/median and standard deviation/percentile bands.
• Optionally overlay individual runs.
• Interactive, fully customizable Plotly charts in Streamlit.

Project Structure

tensorboard_visualizer/
│
└── src/
    ├── app.py                 # Main Streamlit app
    └── utils/
        │── __init__.py
        │── file_utils.py      # File scanning and path parsing
        │── event_utils.py     # TensorBoard event parsing and aggregation

Installation

1. Clone the repository:

git clone https://github.com/Geophyte/tensorboard-visualizer
cd tensorboard_visualizer

2. Create a virtual environment (optional but recommended):

python -m venv venv
source venv/bin/activate  # Linux/macOS
venv\Scripts\activate     # Windows

3. Install dependencies:

pip install -r requirenments

Usage

Run the Streamlit app:

streamlit run src/app.py

Open the displayed Local URL (usually http://localhost:8501) in your browser.

Sidebar Options

• Data

  • Base directories: directories to scan for TensorBoard events.
  • Include regex / Exclude regex: filter runs by path.

• Tag extraction

  • Regex patterns to extract tags from file paths (use named groups, e.g. lr=(?P<lr>\d+\.\d+)).

• Plot

  • Aggregation method: mean or median.
  • Show individual runs: overlay each run.
  • Order groups by: none, best_final, best_max.
  • Legend and margins customization.

Example Usage

Suppose your TensorBoard logs are organized like this:

runs/
├── experiment1_lr=0.001_bs=32/
│   └── events.out.tfevents.123
├── experiment1_lr=0.001_bs=64/
│   └── events.out.tfevents.124
├── experiment2_lr=0.01_bs=32/
│   └── events.out.tfevents.125
└── experiment2_lr=0.01_bs=64/
    └── events.out.tfevents.126

You want to group runs by learning rate (lr) and batch size (bs) and visualize the loss metric.

Regex Patterns

Enter the following in the Tag extraction textarea:

lr=(?P<lr>[\d.]+)_bs=(?P<bs>\d+)

This will extract:

Path	lr	bs
runs/experiment1_lr=0.001_bs=32/	0.001	32
runs/experiment2_lr=0.01_bs=64/	0.01	64

Grouping

• In Group runs by these keys, select lr and bs.
• Each combination of lr and bs becomes a separate group in the plot.

Metric

• Choose loss in the Metric to plot dropdown.

Result

• The app will plot the mean (or median) loss across all runs in each group.
• Standard deviation bands will be shaded.
• You can also overlay individual runs if Show individual runs is checked.

License

MIT License