dustincys/saucie

SAUCIE

An implementation of SAUCIE (Sparse Autoencoder for Clustering, Imputing, and Embedding) in Tensorflow.

Requirements

All tests performed with:

tensorflow 1.4.0
numpy 1.13.3

Usage

SAUCIE is a python object that loads data from a numpy matrix and produces numpy matrix output for the reconstruction, visualization, and/or clusters. Standard usage is to train a model from a numpy matrix and get the embedding, reconstruction, or clusters for that data. This can be done with:

data = ...

from model import SAUCIE
from loader import Loader

saucie = SAUCIE(data.shape[1])
loadtrain = Loader(data, shuffle=True)
saucie.train(loadtrain, steps=1000)

loadeval = Loader(data, shuffle=False)
embedding = saucie.get_embedding(loadeval)
number_of_clusters, clusters = saucie.get_clusters(loadeval)
reconstruction = saucie.get_reconstruction(loadeval)

... work with numpy results as desired ...

Running

SAUCIE also comes with the option of running a full cohort of samples if the data is prepared in a specific way. Namely, for a folder of CSV (or FCS files if the flag --fcs is provided), an example of how to use SAUCIE for both batch correction and clustering is:

python SAUCIE.py --input_dir path/to/input/files
                 --output_dir path/for/output/files
                 --batch_correct
                 --cluster
                 [--lambda_b .1]
                 [--lambda_c .1]
                 [--lambda_d .2];

The input directory must contain the CSV (or FCS if you specify --fcs) you wish to run SAUCIE on. If you do not want to run SAUCIE on all columns in the input file, a file named cols_to_use.txt with the 0-indexed column numbers, one per line can be provided.

In the output directory, if batch correction was done, there will be a folder batch_corrected with a batch-corrected CSV (or FCS) file corresponding to each original file. If clustering was done, there will also be a folder clustered with a clustered file corresponding to each original file. In each clustered file, there is either the original or batch-corrected data with additional columns giving the cluster number and the X and Y coordinate for the visualization.