pylearn-mulm's Introduction

Massive univariate linear model

Provide basic features similar to "statmodels" (like OLS) where Y is a matrix of many responses where many independant fit are requested.

Important links

Installation

Unless you already have Numpy and Scipy installed, you need to install them:

sudo apt-get install python-numpy python-scipy

Clone the repository from github

git clone https://github.com/neurospin/pylearn-mulm.git

Add pylearn-mulm in your $PYTHONPATH

Dataset

import numpy as np
import mulm
import statsmodels.api as sm

n = 100
X = np.random.randn(n, 5)
Y = np.random.randn(n, 10)
beta = np.random.randn(5, 1)
# Causal model: add X on the 2 first variables
Y[:, :2] += np.dot(X, beta)

T-tests with MULM

t-test of all the regressors (by default MULM and statsmodel do two-tailed tests). Use Ordinary least squares (OLS).

contrasts = np.identity(X.shape[1])

mod = mulm.MUOLS(Y, X)
mulm_tvals, mulm_pvals, mulm_df = mod.fit().t_test(contrasts, pval=True, two_tailed=True)

Use statmodels, need to iterate over Y columns

sm_tvals = list()
sm_pvals = list()
for j in range(Y.shape[1]):
    mod = sm.OLS(Y[:, j], X)
    sm_ttest = mod.fit().t_test(contrasts)
    sm_tvals.append(sm_ttest.tvalue)
    sm_pvals.append(sm_ttest.pvalue)

sm_tvals = np.asarray(sm_tvals).T
sm_pvals = np.asarray(sm_pvals).T


# Check that results ar similar
np.allclose(mulm_tvals, sm_tvals)
np.allclose(mulm_pvals, sm_pvals)

Multiple comparison: maxT

import numpy as np
import mulm
import pylab as plt

n = 100
px = 5
py_info = 2
py_noize = 100

beta = np.array([1, 0, .5] + [0] * (px - 4) + [2]).reshape((px, 1))
X = np.hstack([np.random.randn(n, px-1), np.ones((n, 1))]) # X with intercept
Y = np.random.randn(n, py_info + py_noize)
# Causal model: add X on the first py_info variable
Y[:, :py_info] += np.dot(X, beta)

# t-test all the regressors (by default mulm and sm do two-tailed tests)
contrasts = np.identity(X.shape[1])

mod = mulm.MUOLS(Y, X)
tvals, rawp, df = mod.fit().t_test(contrasts, pval=True, two_tailed=True)
tvals, maxT, df2 = mod.t_test_maxT(contrasts, two_tailed=True)


n, bins, patches = plt.hist([rawp[0,:], maxT[0,:]],
                            color=['blue', 'red'], label=['rawp','maxT'])
plt.legend()
plt.show()

pylearn-mulm's People

Contributors

Stargazers

Watchers

pylearn-mulm's Issues

Excessive dependencies

I have some concern about the number of mandatory dependencies of MULM.

Having to install MULM on several computers, I often enter a recursive installation problem:

in order to use MULM, you need EPAC
in order to use EPAC, you need soma-workflow
in order to use soma-workflow, you need dill
Moreover you need fairly recent version of those packages which means:
finding the webpage
cloning git repositories
installing if necessary
updating PYTHONPATH (twice if you use spyder)
All of this has to be done whether or not you use the parallel features.

I know that this is documented and that those dependencies are needed for some tasks. I was just wondering if we could try to import the packages only when needed.

Problem with SomaWorkflowEngine

Hello,

When I run the main() in mulm/models.py all the jobs failed with:

Traceback (most recent call last):
  File "./brainomics/pylearn-epac/bin/epac_mapper", line 78, in <module>
    f = open(relative_filepath, 'r')
IOError: [Errno 2] No such file or directory: './0.job'

Nothing was written in the working directory.

Recommend Projects