tensorvis 1.3.10

Visualisation tool to support my PhD automating the process of gathering data and plotting it

TensorVis

Overview • Features • Installation • Contribute

Overview

A command line tool to automate the process of grabbing tensorboard events data and visualising them. This allows for faster result analysis and separation of the experiment logic from the visualisation aspect of the metrics logged in tensorboard.

Features

• Uploads experiment metrics logged to tensorboard to tensorboard.dev and creates a log of uploaded experiments.
• Downloads experiments from tensorboard.dev to a local csv file.
• Plots experiment metrics.

Benefits

1. Faster result analysis
2. Less code writting
3. Separate experiments from analysis
4. Allows for more research time

Installation

tensorvis can be installed using pip with the command:

pip install tensorvis

This will install tensorvis in the current python environment and will be available through the terminal.

tensorvis supports autocompletion of commands and experiment names using Click's shell completion. To initialise autocompletion run the following command if using Ubuntu and bash:

mv /path/to/virtualenvs/your-virtualenv/site-packages/.tensorvis-complete.bash ~/

The reason for this path is due to an issue with poetry not packaging data within the .whl file. You can find more about this here.

Follow Click' documentation linked above for different shell support.

Assumptions

There can be many different directory structures when running and logging experiments with tensorboard. This tool makes several assumptions to make it easier to handle dataframes resulting from downloading experiments.

tensorvis assumes the following directory structure of tensorboard logs within the top level directory logs, where each run subdirectory contains the events file:

logs
├── exp_name_1
│   ├── run_1
│   └── run_2
├── exp_name_2
│   ├── run_1
│   ├── run_2
│   └── run_3
└── exp_name_3
    └── run_1

For a description of how the directory structure is represented in a dataframe follow this link.

By default tensorvis assumes a single experiment directory is provided which corresponds to a single experiment having multiple runs. All runs from a single experiment will be aggregate and averaged to plot the mean values along with the standard deviation.

Contribute

Any feedback on tensorvis is welcomed in order to improve its usage and versatility. If you have something specific in mind please don't hesitate to create an issue or better yet open a PR!

Current Contributors

• npitsillos