How to see T-SNE?

[pgeusau]

First of all great demo!
Is it possible to have a T-SNE plot instead of a PCA plot?
Also, do you have the ('squad.bin', 'hotpot_distract.bin', and 'babi.bin') models available?

[bvanaken]

Hi, thanks for your interest in our demo.

In order to switch to T-SNE, you can simply change the reduction method in app.py, line 86, to:
# dimensionality reduction
layer_reduced = visualize.reduce(token_vectors, "tsne", 2)

However, we saw that T-SNE does not create meaningful clusters when only fitted on the data of one sample, but PCA does. So in order to get good T-SNE clusters you must probably fit an estimator on a set of samples, e.g. your training data and then apply it to new unseen samples. The estimator could then be loaded and applied in visualize.py.

[bvanaken]

Regarding your question about the models, we decided to provide the models for download here: https://drive.google.com/drive/folders/1RZvlZmhxiJKiAagwvE5vjmevmiqqjR0r

[pgeusau]

Thanks a lot!