Embedding Visualization

Visualizing embeddings generated by a model can be very helpful to see it’s performance, or to perform analysis of the effectiveness of the model.

To visualize embeddings, it’s only necessary to specify the following arguments:

  • results-file specifies the path to the results.pkl file generated by appying a model(see Applying a Model to New Data)
  • output specifies the path to an output image or output path_to_output_folder
  • embedding specifies the type of embedding that can be used. (Some of the available ones are TSNE, PCA and IsoMap)
  • tensorboard specifies that the embeddings should be output in a Tensorboard compatible format

For example,

python util/visualization/embedding.py --results-file ./output/test/test/classify=True/02-07-18-15h-46m-32s/results.pkl --output-file ./output.png --embedding tsne

would generate the image embeddings

TensorBoard

This script can be used to output the embeddings in a Tensorboard compatible format.

python util/visualization/embedding.py --results-file <path_to_results.pkl> --output-file <path_to_output_folder> --tensorboard

These embeddings can be then visualized interactively using Tensorboard. For more information, refer to Visualizing the Results