Embedding Visualizer Using Tensorboard


Install
pip install tbev==0.0.4

Documentation

TensorBoard Embedding Visualizer

DEMO

As the name suggests, this tool is made for visualizing high dimensional embeddings into 3D space using TensorBoard Projector tool. It performs PCA and also has an option for T-SNE for doing dimensionality reduction. All you have to do is provide you embeddings data and its corresponding labels and optionally images(paths) in pickle format as a dictionary. The structure of this dictionary is explained below.

Getting Started

Requirement

  • Python3
  • Tensorflow > 1.4

Install

Install this tool using pip with the following command.
pip install tbev
Upgrade to the latest version by using the --upgrade flag
This will be installed as a command line tool and you can simply run tbev in you command line to get started.

Usage

Usage:
    tbev demo 
    tbev <pickle_file> [--logdir=<path>]

Options:
    -h --h    Show help screen
    --logdir=<path>  Location to store log files [default: ./logs/]

tbev <pickle_file> [--logdir=] You will need to pass your embeddings in form of a pickle file. The pickle file should contain a dictionary in the following format.

{
    "embedding":2D matrix of shape [m embeddings, n dimensions],
    "labels":{
        "label1": List of shape (m,), A label for each embedding,
        "label2": List of shape (m,), A label for each embedding,
        You can put as many labels as you want.
    },
    "sprite_paths":(Optional) List of shape (m,), A path to image to be shown in tensorboard.
}

Save this dictionary into a pickle file and use the command tbev <pickle_file> [--logdir=<path>]
This will create a ./logs folder by default where it will store the checkpoints file. You can optionally mention your own name for storing the logs using the --logdir option. If everything works out well. It will start the tensorboard server for you.

Tensorboard

After the Tensorboard is started, it will show you the local URL to view the tensorboard. By default it is localhost:6006.