tSNE on Colab => DR => tSNE in browser

reconstrue / single_cell

Single cell analysis tools built to run on Jupyter, especially Colab

http://reconstrue.com

Apache License 2.0

1 stars 0 forks source link

tSNE on Colab => DR => tSNE in browser #1

Open JohnTigue opened 4 years ago

JohnTigue commented 4 years ago

How to reduce the Cell X Features grid down to what a laptop GPU can handle. Somewhere around 5K -- 10K. Not sure how many principle components. tSNE seems to like 10 -- 100.

The developer of apricot lives in Seattle. Qouting:

These exemplars can be used for a variety of tasks, including selecting subsets for training machine learning models, visualization in the place of large data sets"

[ ] When working, update https://github.com/reconstrue/neuro_colab/issues/9
See also #69

JohnTigue commented 4 years ago

Tech to try for sub-sampling

[ ] (BBI/FredHutch ?)
[ ] projected support points (PSPs)
[ ] ? Hamiltonian Monte Carlo with Energy Conserving Subsampling

JohnTigue commented 4 years ago

Perhaps a very rough first pass tsne on the server. Just enough to get the data to spread out somewhat evenly. Then something like apricot to downsample. That small sample size (with full data per cell but not so many cells) is then downloaded to client where full tnse with quick exploration of what a good value for perplexity is.

Finally, when know what is a good perplexity is return to the server to do full tSNE.