[ENH] update ICA to sklearn from mdp

README.md

@@ -30,7 +30,6 @@ More information and documentation can be found at https://tedana.readthedocs.io
 You'll need to set up a working development environment to use `tedana`.
 To set up a local environment, you will need Python >=3.6 and the following packages will need to be installed:
 
-[mdp](https://pypi.org/project/MDP/)  
 [numpy](http://www.numpy.org/)   
 [scikit-learn](http://scikit-learn.org/stable/)   
 [scipy](https://www.scipy.org/)    

docs/installation.rst

@@ -2,10 +2,9 @@ Installation
 ------------
 
 You'll need to set up a working development environment to use ``tedana``.
-To set up a local environment, you will need Python >=3.6 and the following
+To set up a local environment, you will need Python >=3.5 and the following
 packages will need to be installed:
 
-- mdp
 - nilearn
 - nibabel>=2.1.0
 - numpy

requirements.txt

@@ -1,7 +1,6 @@
 numpy>=1.14
 scikit-learn
 scipy
-mdp
 nilearn
 nibabel>=2.1.0
 pywavelets

tedana/decomposition/eigendecomp.py

@@ -333,8 +333,7 @@ def tedpca(catd, OCcatd, combmode, mask, t2s, t2sG, stabilize,
     return n_components, kept_data
 
 
-def tedica(n_components, dd, conv, fixed_seed, cost, final_cost,
-           verbose=False):
+def tedica(n_components, dd, conv, fixed_seed, cost='logcosh'):
     """
     Performs ICA on `dd` and returns mixing matrix
 
@@ -347,16 +346,10 @@ def tedica(n_components, dd, conv, fixed_seed, cost, final_cost,
         samples and `T` is time
     conv : :obj:`float`
         Convergence limit for ICA
-    fixed_seed : :obj:`int`
-        Value passed to ``mdp.numx_rand.seed()``.
-        Set to an integer value for reproducible ICA results;
-        otherwise, set to -1 for varying results across calls.
-    cost : {'tanh', 'pow3', 'gaus', 'skew'}
-        Initial cost function for ICA.
-    final_cost : {'tanh', 'pow3', 'gaus', 'skew'}
-        Final cost function for ICA.
-    verbose : :obj:`bool`, optional
-        Whether to print messages regarding convergence process. Default: False
+    cost : {'logcosh', 'exp', 'cube'} str, optional
+        Cost function for ICA
+    fixed_seed : int
+        Seed for ensuring reproducibility of ICA results
 
     Returns
     -------
@@ -366,20 +359,22 @@ def tedica(n_components, dd, conv, fixed_seed, cost, final_cost,
 
     Notes
     -----
-    Uses `mdp` implementation of FastICA for decomposition
+    Uses `sklearn` implementation of FastICA for decomposition
     """
 
-    import mdp
+    from sklearn.decomposition import FastICA
+
+    if cost not in ('logcosh', 'cube', 'exp'):
+        LGR.error('ICA cost function not understood')
+        raise
+
     climit = float(conv)
     if fixed_seed == -1:
         fixed_seed = np.random.randint(low=1, high=1000)
-    mdp.numx_rand.seed(fixed_seed)
-    icanode = mdp.nodes.FastICANode(white_comp=n_components, approach='symm',
-                                    g=cost, fine_g=final_cost,
-                                    coarse_limit=climit*100, limit=climit,
-                                    verbose=verbose)
-    icanode.train(dd)
-    smaps = icanode.execute(dd)  # noqa
-    mmix = icanode.get_recmatrix().T
+    rand_state = np.random.RandomState(seed=fixed_seed)
+    ica = FastICA(n_components=n_components, algorithm='parallel',
+                  fun=cost, tol=climit, random_state=rand_state)
+    ica.fit(dd)
+    mmix = ica.mixing_
     mmix = stats.zscore(mmix, axis=0)
     return mmix, fixed_seed

tedana/info.py

@@ -30,7 +30,6 @@
 REQUIRES = [
     'numpy',
     'scikit-learn',
-    'mdp',
     'pywavelets',
     'nilearn',
     'nibabel>=2.1.0',

tedana/workflows/tedana.py

@@ -103,19 +103,12 @@ def _get_parser():
                         help=('Combination scheme for TEs: '
                               't2s (Posse 1999, default), ste (Poser)'),
                         default='t2s')
-    parser.add_argument('--initcost',
-                        dest='initcost',
-                        action='store',
-                        choices=['tanh', 'pow3', 'gaus', 'skew'],
-                        help=('Initial cost function for ICA.'),
-                        default='tanh')
-    parser.add_argument('--finalcost',
-                        dest='finalcost',
-                        action='store',
-                        choices=['tanh', 'pow3', 'gaus', 'skew'],
-                        help=('Final cost function for ICA. Same options as '
-                              'initcost.'),
-                        default='tanh')
+    parser.add_argument('--cost',
+                        dest='cost',
+                        help=('Cost func. for ICA: '
+                              'logcosh (default), cube, exp'),
+                        choices=['logcosh', 'cube', 'exp'],
+                        default='logcosh')
     parser.add_argument('--denoiseTEs',
                         dest='dne',
                         action='store_true',
@@ -174,8 +167,7 @@ def _get_parser():
 
 def tedana_workflow(data, tes, mask=None, mixm=None, ctab=None, manacc=None,
                     strict=False, gscontrol=True, kdaw=10., rdaw=1., conv=2.5e-5,
-                    ste=-1, combmode='t2s', dne=False,
-                    initcost='tanh', finalcost='tanh',
+                    ste=-1, combmode='t2s', dne=False, cost='logcosh',
                     stabilize=False, filecsdata=False, wvpca=False,
                     label=None, fixed_seed=42, debug=False, quiet=False):
     """
@@ -219,10 +211,8 @@ def tedana_workflow(data, tes, mask=None, mixm=None, ctab=None, manacc=None,
         Combination scheme for TEs: 't2s' (Posse 1999, default), 'ste' (Poser).
     dne : :obj:`bool`, optional
         Denoise each TE dataset separately. Default is False.
-    initcost : {'tanh', 'pow3', 'gaus', 'skew'}, optional
-        Initial cost function for ICA. Default is 'tanh'.
-    finalcost : {'tanh', 'pow3', 'gaus', 'skew'}, optional
-        Final cost function. Default is 'tanh'.
+    cost : {'logcosh', 'exp', 'cube'} str, optional
+        Cost function for ICA
     stabilize : :obj:`bool`, optional
         Stabilize convergence by reducing dimensionality, for low quality data.
         Default is False.
@@ -337,11 +327,10 @@ def tedana_workflow(data, tes, mask=None, mixm=None, ctab=None, manacc=None,
                                                 t2s, t2sG, stabilize, ref_img,
                                                 tes=tes, kdaw=kdaw, rdaw=rdaw,
                                                 ste=ste, wvpca=wvpca)
-        # Perform ICA on dimensionally reduced data (*without* thermal noise)
-        mmix_orig, fixed_seed = decomposition.tedica(n_components, dd, conv, fixed_seed,
-                                                     cost=initcost, final_cost=finalcost,
-                                                     verbose=debug)
+        mmix_orig, fixed_seed = decomposition.tedica(n_components, dd, conv,
+                                                     fixed_seed, cost=cost)
         np.savetxt(op.join(out_dir, '__meica_mix.1D'), mmix_orig)
+
         LGR.info('Making second component selection guess from ICA results')
         # Estimate betas and compute selection metrics for mixing matrix
         # generated from dimensionally reduced data using full data (i.e., data