https://github.com/haschka/pca

A tool to perform fast principal component analyses with millions of samples and thousands of features
https://github.com/haschka/pca

Last synced: about 2 months ago
JSON representation

A tool to perform fast principal component analyses with millions of samples and thousands of features

• Host: GitHub
• URL: https://github.com/haschka/pca
• Owner: haschka
• License: zlib
• Created: 2022-08-11T22:41:22.000Z (over 2 years ago)
• Default Branch: main
• Last Pushed: 2022-08-11T23:59:53.000Z (over 2 years ago)
• Last Synced: 2023-02-28T04:01:16.717Z (almost 2 years ago)
• Language: C
• Size: 15.6 KB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# PCA

(c) Thomas Haschka 2022 

for a license please read the LICENSE (zlib license) file.

A tool to perform fast principal component analyses with millions of samples and thousands of features:

This tool works best (in termes of runtime) if samples >> features. But is not limited to such cases. It was tested with 2.5 millions of samples and up to around 16 000 features. 

A perfect setting is 30 000 samples 4000 features, where on recent machine results are obtained extremly fast. 

### Requirements

In order to compile and run this tools you need to have the lapack libraries,

installed. Further you need a suitable compilier. On a Debian based system you may

install all requirements in the following way:

```

sudo apt-get install build-essential liblapack-dev

```

### compilation

You may compile the tool (pca.c) in the following way, here gcc is assumed:

```

gcc -O2 -march=native pca.c -lm -llapack -pthread -o pca -s

```

### Usage

The tool can be run with the following self explainatory arguments:

```

 [FILE-in ] data to apply PCA on 

 [FILE-out] data projected ontoprincipal components

 [FILE-out] Eigenvalue Spectrum of the PCA 

 [FILE-out] Eigenvectors from the PCA 

 [int] dimensions - how many principal components to projectonto 

 [int] number of threads to use for this computation

```

where number of threads can be chosen by you. The number of cores of the machine

is a good guess.

The **input file** has to be formatted as a tabulator seperated file (tsv). 

The first column of the file is omitted and can hold sample names.

Each succeding column holds a feature ( floating point value ) describing the sample

Each row holds the information about one sample. *The input file requires at least

4 samples to yield correct results!* 

Here's an example:

```

sample1 1.5     10.2    .4      1.

sample2 2.4     21.9    1.0     2.

sample3 3.5     30.9    1.8     3.

sample4 4.5     40.9    2.5     4.

sample5 5.5     50.9    3.0     5.

```

(Beware that you cannot just copy paste this as it needs real tabulators not spaces. If you 

want to test this tool use the supplied test.tsv file. 

The **"data projected onto principal components"** output file contains 

the projections of each sample onto the principal components per row. 

Hence each line represents a sample and holds as much values as dimensions 

you might have chosen. The samples are in the same order as they were in the 

input file. 

The **"Eigenvalue Spectrum"** file shows the eigenvalues corresponding to all 

princpal components

The **"Eigenvectors** file holds number of dimensions eigenvectors corresponding

to the largest eigenvalues.

A sample usage with the supplied test file:

``` 

./pca test.tsv test.pca test.eval test.evec 2 1

```