Closed huddlej closed 5 months ago
Based on our results from applying t-SNE to simulated populations in https://github.com/blab/cartography, a more reasonable default learning rate should scale with the input sample size instead of being fixed at 200.0. The default learning rate in the scikit-learn implementation we use is already the string "auto" which allows the learning rate to scale with the sample size. We should change the default value for the tsne subcommand's --learning-rate argument to "auto". This change will require us to define a custom argparse type function, since valid values can be either a float or a string with the value of auto.
tsne
--learning-rate
"auto"
float
auto
issue is fixed with default-learning-rate branch
Based on our results from applying t-SNE to simulated populations in https://github.com/blab/cartography, a more reasonable default learning rate should scale with the input sample size instead of being fixed at 200.0. The default learning rate in the scikit-learn implementation we use is already the string "auto" which allows the learning rate to scale with the sample size. We should change the default value for the
tsne
subcommand's--learning-rate
argument to"auto"
. This change will require us to define a custom argparse type function, since valid values can be either afloat
or a string with the value ofauto
.