t-SNE should expose convergence criteria #5165
Comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
8000
{{ message }}
In sklearn.manifold.TSNE, the parameter "n_iter_without_progress" is in the _gradient_descent method https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/manifold/t_sne.py#L114 is defaulted to 30, and there's no way through the public interface to change it.
On my data sets, this often causes it to stop too early, resulting in poor embeddings. For example, I often observe identical input vectors not getting projected to the same embedding. Reducing the learning rate can help, but then I bump into the min_error_diff parameter which is also not configurable, and causes it to believe it has converged when it hasn't.
The text was updated successfully, but these errors were encountered: