Function Preserving Projection (FPP)

A linear projection technique for finding a 2D view that capture interpretable pattern of the given function in a high-dimensional domain. The function can be univariate or multivariate, continuous (regression) or discrete (classification). The details of the method can be found in the corresponding paper: https://arxiv.org/pdf/1909.11804.pdf

Dependency:

tensorflow (<=1.15.0), numpy

File Description:

fpp.py - function preserving projection class

fpp_example.ipynb - fpp usage examples

Test Dataset Description:

Circle_in_5D_cube.npy - synthetic dataset where a 2D circle pattern of the function exists in a 5D space.

Circle_in_30D.npy - synthetic dataset where a 2D circle pattern of the function exists in a 30D space.

Fpp Class Usage:

# X - function domain
# f - function range
# epoches - training epoches
# batchSize - traiing batch size
# proj_mat - projection matrix
# embedding - the 2D embedding coordinate
# loss - the loss on the entire training dataset

###### regression task #####
model = fpp()
model.setup(X, f)
model.train()
proj_mat, embedding, loss, _ = model.eval()

###### classification task #####
model = fpp()
model.setupMultiClass(X, f) #f should be a one-hot encoding of the class
model.train(epoches, batchSize)
proj_mat, embedding, loss, _ = model.eval()

Reviewed and released under LLNL-CODE-791217

Author(s): Shusen Liu (liu42@llnl.gov), Rushil Anirudh (anirudh1@llnl.gov)