absorbguo / tsne-tensorboard-visualisation

This repository provides a starter code for using tensorboard via tensorflow for visualising embeddings

tf-tb-visualisation

How to run

This repository provides a starter code for using tensorboard via tensorflow for visualising embeddings

The following is the folder structure expected by the code:

• sample_data/
  • embeddings/
    • filename_embedding
  • images/
    • data/
    • metadata.txt
  • text/
    • metadata.txt

The filename_embedding consists of the n_dimensional embeddings The data folder consists of all the images The metadata.txt for images consists of the following format: image_filename\tlabel (one to one mapping with embedding vector) The metadata.txt for text consists of the following format: label (one to one mapping with embedding vector)

For visualising embeddings run the following from the command line: For word embeddings:

python visualise_embeddings.py -b $baseDir -f $filename_embedding -m "text" -l $filename_label

Example usage:

python visualise_embeddings.py -b /Users/ayushi/Work/tf-tb-visualisation/sample_data/ -f feature_vectors_400_samples.txt -m text -l metadata_text.txt

For image embeddings:

python visualise_embeddings.py -b $baseDir -f $filename_embedding -m "image" -l $filename_label

Example usage:

python visualise_embeddings.py -b /Users/ayushi/Work/tf-tb-visualisation/sample_data/ -f feature_vectors_400_samples.txt -m "image" -l metadata_images.txt

Then finally run:

tensorboard --logdir=$baseDir

Example usage:

tensorboard --logdir=/Users/ayushi/Work/tf-tb-visualisation/sample_data/

Note: Giving the complete path is important.

Reading:

• Visualizing Data using t-SNE
• How to use t-SNE efficiently

Acknowledgement

I would like to sincerely thank Anuj Shah for his data and code for sprite image.