Support for memmap as input type

[iarroyof]

Hi, I have a problem while trying to use Regressor. Im using X_train=numpy.loadtxt() methods for loading dense data from file. This file contains a dense vector by each row. There are 525 rows with 300 dimensions each. I configured a randomizedSearch called rs.

X_train: (525, 300)
y_train: (525,)
X_test: (225, 300)
Traceback (most recent call last):
  File "nn.py", line 56, in <module>
    rs.fit(X_train, y_train)
  File "/usr/local/lib/python2.7/dist-packages/sklearn/grid_search.py", line 996, in fit
    return self._fit(X, y, sampled_params)
  File "/usr/local/lib/python2.7/dist-packages/sklearn/grid_search.py", line 553, in _fit
    for parameters in parameter_iterable
  File "/usr/local/lib/python2.7/dist-packages/sklearn/externals/joblib/parallel.py", line 812, in __call__
    self.retrieve()
  File "/usr/local/lib/python2.7/dist-packages/sklearn/externals/joblib/parallel.py", line 762, in retrieve
    raise exception
sklearn.externals.joblib.my_exceptions.JoblibAttributeError: JoblibAttributeError
___________________________________________________________________________
Multiprocessing exception:
...........................................................................
/almac/ignacio/nlp-pipeline/nn.py in <module>()
     51     exit()
     52 
     53 crash = True
     54 while(crash):
     55     try:
---> 56         rs.fit(X_train, y_train)
     57         crash = False
     58     except RuntimeError:
     59         sys.stderr.write("----------- [Crashed by RunTimeERROR] --------------------- \n")
     60         crash = True

...........................................................................
/usr/local/lib/python2.7/dist-packages/sklearn/grid_search.py in fit(self=RandomizedSearchCV(cv=None, error_score='raise',...e, refit=True,
          scoring=None, verbose=0), X=array([[ -7.19239362e+05,  -2.57225021e+05,  -5....758269e+00,   6.37787139e-02,  -6.47838211e-02]]), y=array([ 2.6 ,  2.6 ,  3.8 ,  4.2 ,  0.8 ,  1.4 ,... 4.2 ,  0.  ,  2.25,
        2.6 ,  2.  ,  3.  ]))
    991 
    992         """
    993         sampled_params = ParameterSampler(self.param_distributions,
    994                                           self.n_iter,
    995                                           random_state=self.random_state)
--> 996         return self._fit(X, y, sampled_params)
        self._fit = <bound method RandomizedSearchCV._fit of Randomi..., refit=True,
          scoring=None, verbose=0)>
        X = array([[ -7.19239362e+05,  -2.57225021e+05,  -5....758269e+00,   6.37787139e-02,  -6.47838211e-02]])
        y = array([ 2.6 ,  2.6 ,  3.8 ,  4.2 ,  0.8 ,  1.4 ,... 4.2 ,  0.  ,  2.25,
        2.6 ,  2.  ,  3.  ])
        sampled_params = <sklearn.grid_search.ParameterSampler object>
    997 
    998 
    999 
   1000 

...........................................................................
/usr/local/lib/python2.7/dist-packages/sklearn/grid_search.py in _fit(self=RandomizedSearchCV(cv=None, error_score='raise',...e, refit=True,
          scoring=None, verbose=0), X=array([[ -7.19239362e+05,  -2.57225021e+05,  -5....758269e+00,   6.37787139e-02,  -6.47838211e-02]]), y=array([ 2.6 ,  2.6 ,  3.8 ,  4.2 ,  0.8 ,  1.4 ,... 4.2 ,  0.  ,  2.25,
        2.6 ,  2.  ,  3.  ]), parameter_iterable=<sklearn.grid_search.ParameterSampler object>)
    548         )(
    549             delayed(_fit_and_score)(clone(base_estimator), X, y, self.scorer_,
    550                                     train, test, self.verbose, parameters,
    551                                     self.fit_params, return_parameters=True,
    552                                     error_score=self.error_score)
--> 553                 for parameters in parameter_iterable
        parameters = undefined
        parameter_iterable = <sklearn.grid_search.ParameterSampler object>
    554                 for train, test in cv)
    555 
    556         # Out is a list of triplet: score, estimator, n_test_samples
    557         n_fits = len(out)

...........................................................................
/usr/local/lib/python2.7/dist-packages/sklearn/externals/joblib/parallel.py in __call__(self=Parallel(n_jobs=16), iterable=<generator object <genexpr>>)
    807             if pre_dispatch == "all" or n_jobs == 1:
    808                 # The iterable was consumed all at once by the above for loop.
    809                 # No need to wait for async callbacks to trigger to
    810                 # consumption.
    811                 self._iterating = False
--> 812             self.retrieve()
        self.retrieve = <bound method Parallel.retrieve of Parallel(n_jobs=16)>
    813             # Make sure that we get a last message telling us we are done
    814             elapsed_time = time.time() - self._start_time
    815             self._print('Done %3i out of %3i | elapsed: %s finished',
    816                         (len(self._output), len(self._output),

---------------------------------------------------------------------------
Sub-process traceback:
---------------------------------------------------------------------------
AttributeError                                     Sun Feb 21 19:22:11 2016
PID: 8977                                     Python 2.7.6: /usr/bin/python
...........................................................................
/usr/local/lib/python2.7/dist-packages/sklearn/externals/joblib/parallel.pyc in __call__(self=<sklearn.externals.joblib.parallel.BatchedCalls object>)
     67     def __init__(self, iterator_slice):
     68         self.items = list(iterator_slice)
     69         self._size = len(self.items)
     70 
     71     def __call__(self):
---> 72         return [func(*args, **kwargs) for func, args, kwargs in self.items]
     73 
     74     def __len__(self):
     75         return self._size
     76 

...........................................................................
/usr/local/lib/python2.7/dist-packages/sklearn/cross_validation.pyc in _fit_and_score(estimator=Regressor(batch_size=1, callback=None, debug=Fal...e, warning=None, weight_decay=None, weights=None), X=memmap([[ -7.19239362e+05,  -2.57225021e+05,  -5...758269e+00,   6.37787139e-02,  -6.47838211e-02]]), y=array([ 2.6 ,  2.6 ,  3.8 ,  4.2 ,  0.8 ,  1.4 ,... 4.2 ,  0.  ,  2.25,
        2.6 ,  2.  ,  3.  ]), scorer=<function _passthrough_scorer>, train=array([175, 176, 177, 178, 179, 180, 181, 182, 1...15, 516, 517, 518, 519, 520, 521, 522, 523, 524]), test=array([  0,   1,   2,   3,   4,   5,   6,   7,  ..., 167, 168,
       169, 170, 171, 172, 173, 174]), verbose=0, parameters={'hidden0__type': 'Rectifier', 'hidden0__units': 5, 'hidden1__type': 'Rectifier', 'hidden1__units': 37, 'learning_momentum': 1.2204617397379565, 'learning_rate': 0.09480671449399372, 'learning_rule': 'rmsprop', 'output__type': 'Linear', 'regularize': None}, fit_params={}, return_train_score=False, return_parameters=True, error_score='raise')
   1526 
   1527     try:
   1528         if y_train is None:
   1529             estimator.fit(X_train, **fit_params)
   1530         else:
-> 1531             estimator.fit(X_train, y_train, **fit_params)
   1532 
   1533     except Exception as e:
   1534         if error_score == 'raise':
   1535             raise

...........................................................................
/usr/local/lib/python2.7/dist-packages/sknn/mlp.pyc in fit(self=Regressor(batch_size=1, callback=None, debug=Fal...e, warning=None, weight_decay=None, weights=None), X=memmap([[ -1.71586427e+07,   2.45456060e+06,   6...758269e+00,   6.37787139e-02,  -6.47838211e-02]]), y=array([ 4.4 ,  2.2 ,  3.4 ,  4.  ,  2.6 ,  4.2 ,... 4.8 ,  4.2 ,  0.  ,  2.25,  2.6 ,  2.  ,  3.  ]), w=None)
    288         """
    289 
    290         if self.valid_set is not None:
    291             self.valid_set = self._reshape(*self.valid_set)
    292 
--> 293         return super(Regressor, self)._fit(X, y, w)
    294 
    295     def predict(self, X):
    296         """Calculate predictions for specified inputs.
    297 

...........................................................................
/usr/local/lib/python2.7/dist-packages/sknn/mlp.pyc in _fit(self=Regressor(batch_size=1, callback=None, debug=Fal...e, warning=None, weight_decay=None, weights=None), X=memmap([[ -1.71586427e+07,   2.45456060e+06,   6...758269e+00,   6.37787139e-02,  -6.47838211e-02]]), y=array([[ 4.4 ],
       [ 2.2 ],
       [ 3.4 ],
...
       [ 2.6 ],
       [ 2.  ],
       [ 3.  ]]), w=None)
    227         if self.verbose:
    228             log.debug("\nEpoch       Training Error       Validation Error       Time"
    229                       "\n------------------------------------------------------------")
    230 
    231         try:
--> 232             self._train(X, y, w)
    233         except RuntimeError as e:
    234             log.error("\n{}{}{}\n\n{}\n".format(
    235                 ansi.RED,
    236                 "A runtime exception was caught during training. This likely occurred due to\n"

...........................................................................
/usr/local/lib/python2.7/dist-packages/sknn/mlp.pyc in _train(self=Regressor(batch_size=1, callback=None, debug=Fal...e, warning=None, weight_decay=None, weights=None), X=memmap([[ -1.71586427e+07,   2.45456060e+06,   6...758269e+00,   6.37787139e-02,  -6.47838211e-02]]), y=array([[ 4.4 ],
       [ 2.2 ],
       [ 3.4 ],
...
       [ 2.6 ],
       [ 2.  ],
       [ 3.  ]]), w=None)
    146         for i in itertools.count(1):
    147             start_time = time.time()
    148             self._do_callback('on_epoch_start', locals())
    149 
    150             is_best_train = False
--> 151             avg_train_error = self._backend._train_impl(X, y, w)
    152             if avg_train_error is not None:
    153                 if math.isnan(avg_train_error):
    154                     raise RuntimeError("Training diverged and returned NaN.")
    155                 

...........................................................................
/usr/local/lib/python2.7/dist-packages/sknn/backend/lasagne/mlp.pyc in _train_impl(self=<sknn.backend.lasagne.mlp.MultiLayerPerceptronBackend object>, X=memmap([[ -1.71586427e+07,   2.45456060e+06,   6...758269e+00,   6.37787139e-02,  -6.47838211e-02]]), y=array([[ 4.4 ],
       [ 2.2 ],
       [ 3.4 ],
...
       [ 2.6 ],
       [ 2.  ],
       [ 3.  ]]), w=None)
    300 
    301         self._print('\r')
    302         return loss / count
    303 
    304     def _train_impl(self, X, y, w=None):
--> 305         return self._batch_impl(X, y, w, self.trainer, mode='train', output='.', shuffle=True)
    306 
    307     def _valid_impl(self, X, y, w=None):
    308         return self._batch_impl(X, y, w, self.validator, mode='valid', output=' ', shuffle=False)
    309 

...........................................................................
/usr/local/lib/python2.7/dist-packages/sknn/backend/lasagne/mlp.pyc in _batch_impl(self=<sknn.backend.lasagne.mlp.MultiLayerPerceptronBackend object>, X=memmap([[ -1.71586427e+07,   2.45456060e+06,   6...758269e+00,   6.37787139e-02,  -6.47838211e-02]]), y=array([[ 4.4 ],
       [ 2.2 ],
       [ 3.4 ],
...
       [ 2.6 ],
       [ 2.  ],
       [ 3.  ]]), w=None, processor=<theano.compile.function_module.Function object>, mode=u'train', output=u'.', shuffle=True)
    281             sys.stdout.flush()
    282 
    283     def _batch_impl(self, X, y, w, processor, mode, output, shuffle):
    284         progress, batches = 0, X.shape[0] / self.batch_size
    285         loss, count = 0.0, 0
--> 286         for Xb, yb, wb, _ in self._iterate_data(self.batch_size, X, y, w, shuffle):
    287             self._do_callback('on_batch_start', locals())
    288 
    289             if mode == 'train':
    290                 loss += processor(Xb, yb, wb if wb is not None else 1.0)

...........................................................................
/usr/local/lib/python2.7/dist-packages/sknn/backend/lasagne/mlp.pyc in _iterate_data(self=<sknn.backend.lasagne.mlp.MultiLayerPerceptronBackend object>, batch_size=1, X=memmap([[ -1.71586427e+07,   2.45456060e+06,   6...758269e+00,   6.37787139e-02,  -6.47838211e-02]]), y=array([[ 4.4 ],
       [ 2.2 ],
       [ 3.4 ],
...
       [ 2.6 ],
       [ 2.  ],
       [ 3.  ]]), w=None, shuffle=True)
    270         if shuffle:
    271             numpy.random.shuffle(indices)
    272 
    273         for index in range(0, total_size, batch_size):
    274             excerpt = indices[index:index + batch_size]
--> 275             Xb, yb, wb = cast(X, excerpt), cast(y, excerpt), cast(w, excerpt)
    276             yield Xb, yb, wb, excerpt
    277 
    278     def _print(self, text):
    279         if self.verbose:

...........................................................................
/usr/local/lib/python2.7/dist-packages/sknn/backend/lasagne/mlp.pyc in cast(array=memmap([[  6.39665702e+06,  -1.92380979e+05,   6...693613e+00,   5.43417984e+00,  -1.42057322e+00]]), indices=array([262]))
    258             if array is None:
    259                 return None
    260 
    261             array = array[indices]
    262             if type(array) != numpy.ndarray:
--> 263                 array = array.todense()
    264             if array.dtype != theano.config.floatX:
    265                 array = array.astype(theano.config.floatX)
    266             return array
    267 
AttributeError: 'memmap' object has no attribute 'todense'

Please help..

[alexjc]

For now you'll need to convert your data to a ndarray manually. It should work, then let me know how do you did it so I can include that in the converter!

[iarroyof]

Thank you for reply @alexjc

I acquired the data as follows:

>>> X_train = loadtxt(filename)
>>> type(X_train)
<type 'numpy.ndarray'>

which is the required data type in sknn documentation. I also tried ndarray(X_train), but the error persists. Probably I'm doing something wrong.

[alexjc]

What about y_train? Callstack seems to indicate one is memmap type though.

[iarroyof]

I used the same method:

>>> y_train = loadtxt(filename)
>>> type(y_train)
<type 'numpy.ndarray'>

It is a bit rare because some days ago it worked fine. Although the dataset I used before was much smaller, so I suppose the library considered not necessary using memmap. It is probable a bug is present when the input file is sufficiently large because in mlp.py the data type validation does not consider array has not todense() method.

262             if type(array) != numpy.ndarray:
263                 array = array.todense()

which at the moment only takes place when using memmap is needed. Any idea to overcome it?

[alexjc]

If line 263 is executed, then the type is not ndarray as defined by 262. Can you check this?
type(y_train) == numpy.ndarray

It's possible we need to test if the class is derived from ndarray, not just plain equality. I think this can be done using isinstance(array, numpy.ndarray).

You could also try removing that check to see if it works regardless, and all other required functions (e.g. indexing) are implemented correctly.

[iarroyof]

I'd like to know what do you think @alexjc. I uniquely edited the mlp.py line 263 from array = array.todense() to array = numpy.array(array), it is probably the easiest thing to do. Nevertheless, I am really not sure what I did.

Separately I made an additional test for your suggestion:

In [1]: from tempfile import mkdtemp
In [2]: import os.path as path
In [3]: filename = path.join(mkdtemp(), 'newfile.dat')
In [5]: import numpy as np
In [6]: fp = np.memmap(filename, dtype='float32', mode='w+', shape=(3,4))
In [7]: type(fp)
Out[7]: numpy.core.memmap.memmap
In [8]: isinstance(fp, np.ndarray)
Out[8]: True

so, what do you think is better to do?

[alexjc]

I need to look at this in the context of Pandas DataFrames too... But I would make the following change for now:

262             if not isinstance(array, numpy.ndarray):
263+                assert hasattr(array, 'todense'), "Unknown data type; please convert to ndarray first."

Does it work?

[iarroyof]

The array = numpy.array(array) modification works. I will add what you say also.

[alexjc]

If you make a PR, I'd rather keep .todense() in there explicitly (for now) since it passes all tests with the scipy sparse datastructures. I'll have to look at Pandas separately.

[iarroyof]

Ok, thank you very much.

[alexjc]

Does memmap work with the rest of the code even without casting to numpy.array()?

[iarroyof]

You mean by modifying line 262 as if not isinstance(array, numpy.ndarray): and keeping array.todense()?

[alexjc]

Yeah, does that work?

[iarroyof]

Not indeed. It gives an error:

_64.egg/sklearn/grid_search.py", line 996, in fit
    return self._fit(X, y, sampled_params)
  File "/usr/local/lib/python2.7/dist-packages/scikit_learn-0.17-py2.7-linux-x86_64.egg/sklearn/grid_search.py", line 553, in _fit
    for parameters in parameter_iterable
  File "/usr/local/lib/python2.7/dist-packages/scikit_learn-0.17-py2.7-linux-x86_64.egg/sklearn/externals/joblib/parallel.py", line 812, in __call__
    self.retrieve()
  File "/usr/local/lib/python2.7/dist-packages/scikit_learn-0.17-py2.7-linux-x86_64.egg/sklearn/externals/joblib/parallel.py", line 762, in retrieve
    raise exception
sklearn.externals.joblib.my_exceptions.JoblibAssertionError: JoblibAssertionError
___________________________________________________________________________
Multiprocessing exception:
...........................................................................
/home/iarroyof/nlp-pipeline/nn.py in <module>()
     58 
     59 for t in xrange(N):
     60     crash = True
     61     while(crash):
     62         try:
---> 63             rs.fit(X_train, y_train)
     64             crash = False
     65         except RuntimeError:
     66             sys.stderr.write("--------------------- [Crashed by RunTimeERROR] --------------------- \n")
     67             crash = True

...........................................................................
/usr/local/lib/python2.7/dist-packages/scikit_learn-0.17-py2.7-linux-x86_64.egg/sklearn/grid_search.py in fit(self=RandomizedSearchCV(cv=None, error_score='raise',...e, refit=True,
          scoring=None, verbose=0), X=array([[ -7.19239362e+05,  -2.57225021e+05,  -5....758269e+00,   6.37787139e-02,  -6.47838211e-02]]), y=array([ 2.6 ,  2.6 ,  3.8 ,  4.2 ,  0.8 ,  1.4 ,... 4.2 ,  0.  ,  2.25,
        2.6 ,  2.  ,  3.  ]))
    991 
    992         """
    993         sampled_params = ParameterSampler(self.param_distributions,
    994                                           self.n_iter,
    995                                           random_state=self.random_state)
--> 996         return self._fit(X, y, sampled_params)
        self._fit = <bound method RandomizedSearchCV._fit of Randomi..., refit=True,
          scoring=None, verbose=0)>
        X = array([[ -7.19239362e+05,  -2.57225021e+05,  -5....758269e+00,   6.37787139e-02,  -6.47838211e-02]])
        y = array([ 2.6 ,  2.6 ,  3.8 ,  4.2 ,  0.8 ,  1.4 ,... 4.2 ,  0.  ,  2.25,
        2.6 ,  2.  ,  3.  ])
        sampled_params = <sklearn.grid_search.ParameterSampler object>
    997 
    998 
    999 
   1000 

...........................................................................
/usr/local/lib/python2.7/dist-packages/scikit_learn-0.17-py2.7-linux-x86_64.egg/sklearn/grid_search.py in _fit(self=RandomizedSearchCV(cv=None, error_score='raise',...e, refit=True,
          scoring=None, verbose=0), X=array([[ -7.19239362e+05,  -2.57225021e+05,  -5....758269e+00,   6.37787139e-02,  -6.47838211e-02]]), y=array([ 2.6 ,  2.6 ,  3.8 ,  4.2 ,  0.8 ,  1.4 ,... 4.2 ,  0.  ,  2.25,
        2.6 ,  2.  ,  3.  ]), parameter_iterable=<sklearn.grid_search.ParameterSampler object>)
    548         )(
    549             delayed(_fit_and_score)(clone(base_estimator), X, y, self.scorer_,
    550                                     train, test, self.verbose, parameters,
    551                                     self.fit_params, return_parameters=True,
    552                                     error_score=self.error_score)
--> 553                 for parameters in parameter_iterable
        parameters = undefined
        parameter_iterable = <sklearn.grid_search.ParameterSampler object>
    554                 for train, test in cv)
    555 
    556         # Out is a list of triplet: score, estimator, n_test_samples
    557         n_fits = len(out)

...........................................................................
/usr/local/lib/python2.7/dist-packages/scikit_learn-0.17-py2.7-linux-x86_64.egg/sklearn/externals/joblib/parallel.py in __call__(self=Parallel(n_jobs=16), iterable=<generator object <genexpr>>)
    807             if pre_dispatch == "all" or n_jobs == 1:
    808                 # The iterable was consumed all at once by the above for loop.
    809                 # No need to wait for async callbacks to trigger to
    810                 # consumption.
    811                 self._iterating = False
--> 812             self.retrieve()
        self.retrieve = <bound method Parallel.retrieve of Parallel(n_jobs=16)>
    813             # Make sure that we get a last message telling us we are done
    814             elapsed_time = time.time() - self._start_time
    815             self._print('Done %3i out of %3i | elapsed: %s finished',
    816                         (len(self._output), len(self._output),

---------------------------------------------------------------------------
Sub-process traceback:
---------------------------------------------------------------------------
AssertionError                                     Mon Feb 22 21:11:40 2016
PID: 18416                                    Python 2.7.6: /usr/bin/python
...........................................................................
/usr/local/lib/python2.7/dist-packages/scikit_learn-0.17-py2.7-linux-x86_64.egg/sklearn/externals/joblib/parallel.pyc in __call__(self=<sklearn.externals.joblib.parallel.BatchedCalls object>)
     67     def __init__(self, iterator_slice):
     68         self.items = list(iterator_slice)
     69         self._size = len(self.items)
     70 
     71     def __call__(self):
---> 72         return [func(*args, **kwargs) for func, args, kwargs in self.items]
     73 
     74     def __len__(self):
     75         return self._size
     76 

...........................................................................
/usr/local/lib/python2.7/dist-packages/scikit_learn-0.17-py2.7-linux-x86_64.egg/sklearn/cross_validation.pyc in _fit_and_score(estimator=Regressor(batch_size=1, callback=None, debug=Fal...e, warning=None, weight_decay=None, weights=None), X=memmap([[ -7.19239362e+05,  -2.57225021e+05,  -5...758269e+00,   6.37787139e-02,  -6.47838211e-02]]), y=array([ 2.6 ,  2.6 ,  3.8 ,  4.2 ,  0.8 ,  1.4 ,... 4.2 ,  0.  ,  2.25,
        2.6 ,  2.  ,  3.  ]), scorer=<function _passthrough_scorer>, train=array([175, 176, 177, 178, 179, 180, 181, 182, 1...15, 516, 517, 518, 519, 520, 521, 522, 523, 524]), test=array([  0,   1,   2,   3,   4,   5,   6,   7,  ..., 167, 168,
       169, 170, 171, 172, 173, 174]), verbose=0, parameters={'hidden0__type': 'Rectifier', 'hidden0__units': 16, 'hidden1__type': 'Sigmoid', 'hidden1__units': 39, 'learning_momentum': 0.3153464892915808, 'learning_rate': 0.023125104835940262, 'learning_rule': 'adagrad', 'output__type': 'Softmax', 'regularize': None}, fit_params={}, return_train_score=False, return_parameters=True, error_score='raise')
   1526 
   1527     try:
   1528         if y_train is None:
   1529             estimator.fit(X_train, **fit_params)
   1530         else:
-> 1531             estimator.fit(X_train, y_train, **fit_params)
   1532 
   1533     except Exception as e:
   1534         if error_score == 'raise':
   1535             raise

...........................................................................
/home/iarroyof/scikit-neuralnetwork/sknn/mlp.pyc in fit(self=Regressor(batch_size=1, callback=None, debug=Fal...e, warning=None, weight_decay=None, weights=None), X=memmap([[ -1.71586427e+07,   2.45456060e+06,   6...758269e+00,   6.37787139e-02,  -6.47838211e-02]]), y=array([ 4.4 ,  2.2 ,  3.4 ,  4.  ,  2.6 ,  4.2 ,... 4.8 ,  4.2 ,  0.  ,  2.25,  2.6 ,  2.  ,  3.  ]), w=None)
    288         """
    289 
    290         if self.valid_set is not None:
    291             self.valid_set = self._reshape(*self.valid_set)
    292 
--> 293         return super(Regressor, self)._fit(X, y, w)
    294 
    295     def predict(self, X):
    296         """Calculate predictions for specified inputs.
    297 

...........................................................................
/home/iarroyof/scikit-neuralnetwork/sknn/mlp.pyc in _fit(self=Regressor(batch_size=1, callback=None, debug=Fal...e, warning=None, weight_decay=None, weights=None), X=memmap([[ -1.71586427e+07,   2.45456060e+06,   6...758269e+00,   6.37787139e-02,  -6.47838211e-02]]), y=array([[ 4.4 ],
       [ 2.2 ],
       [ 3.4 ],
...
       [ 2.6 ],
       [ 2.  ],
       [ 3.  ]]), w=None)
    206             "Expecting same number of input and output samples."
    207         data_shape, data_size = X.shape, X.size+y.size
    208         X, y = self._reshape(X, y)
    209 
    210         if not self.is_initialized:
--> 211             X, y = self._initialize(X, y, w)
    212 
    213         log.info("Training on dataset of {:,} samples with {:,} total size.".format(data_shape[0], data_size))
    214         if data_shape[1:] != X.shape[1:]:
    215             log.warning("  - Reshaping input array from {} to {}.".format(data_shape, X.shape))

...........................................................................
/home/iarroyof/scikit-neuralnetwork/sknn/mlp.pyc in _initialize(self=Regressor(batch_size=1, callback=None, debug=Fal...e, warning=None, weight_decay=None, weights=None), X=memmap([[ -1.71586427e+07,   2.45456060e+06,   6...758269e+00,   6.37787139e-02,  -6.47838211e-02]]), y=array([[ 4.4 ],
       [ 2.2 ],
       [ 3.4 ],
...
       [ 2.6 ],
       [ 2.  ],
       [ 3.  ]]), w=None)
     37             "This neural network has already been initialized."
     38         self._create_specs(X, y)
     39 
     40         backend.setup()
     41         self._backend = backend.MultiLayerPerceptronBackend(self)
---> 42         return self._backend._initialize_impl(X, y, w)
     43 
     44     def _check_layer(self, layer, required, optional=[]):
     45         required.extend(['name', 'type'])
     46         for r in required:

...........................................................................
/home/iarroyof/scikit-neuralnetwork/sknn/backend/lasagne/mlp.py in _initialize_impl(self=<sknn.backend.lasagne.mlp.MultiLayerPerceptronBackend object>, X=memmap([[ -1.71586427e+07,   2.45456060e+06,   6...758269e+00,   6.37787139e-02,  -6.47838211e-02]]), y=array([[ 4.4 ],
       [ 2.2 ],
       [ 3.4 ],
...
       [ 2.6 ],
       [ 2.  ],
       [ 3.  ]]), w=None)
    232         params = []
    233         for spec, mlp_layer in zip(self.layers, self.mlp):
    234             if spec.frozen: continue
    235             params.extend(mlp_layer.get_params())
    236 
--> 237         self.trainer, self.validator = self._create_mlp_trainer(params)
        self.trainer = None
        self.validator = None
        self._create_mlp_trainer = <bound method MultiLayerPerceptronBackend._creat....lasagne.mlp.MultiLayerPerceptronBackend object>>
        params = [W, b, W, b, W, b]
    238         return X, y
    239 
    240     def _predict_impl(self, X):
    241         if self.is_convolution():

...........................................................................
/home/iarroyof/scikit-neuralnetwork/sknn/backend/lasagne/mlp.py in _create_mlp_trainer(self=<sknn.backend.lasagne.mlp.MultiLayerPerceptronBackend object>, params=[W, b, W, b, W, b])
     71         cost_symbol = self.cost_function(self.trainer_output, self.data_output)
     72         cost_symbol = lasagne.objectives.aggregate(cost_symbol.T, self.data_mask, mode='mean')
     73 
     74         if self.regularizer is not None:
     75             cost_symbol = cost_symbol + self.regularizer
---> 76         return self._create_trainer_function(params, cost_symbol)
        self._create_trainer_function = <bound method MultiLayerPerceptronBackend._creat....lasagne.mlp.MultiLayerPerceptronBackend object>>
        params = [W, b, W, b, W, b]
        cost_symbol = Elemwise{true_div,no_inplace}.0
     77 
     78     def _create_trainer_function(self, params, cost):
     79         if self.learning_rule in ('sgd', 'adagrad', 'adadelta', 'rmsprop', 'adam'):
     80             lr = getattr(lasagne.updates, self.learning_rule)

...........................................................................
/home/iarroyof/scikit-neuralnetwork/sknn/backend/lasagne/mlp.py in _create_trainer_function(self=<sknn.backend.lasagne.mlp.MultiLayerPerceptronBackend object>, params=[W, b, W, b, W, b], cost=Elemwise{true_div,no_inplace}.0)
     88                 "Learning rule type `%s` is not supported." % self.learning_rule)
     89 
     90         trainer = theano.function([self.data_input, self.data_output, self.data_mask], cost,
     91                                    updates=self._learning_rule,
     92                                    on_unused_input='ignore',
---> 93                                    allow_input_downcast=True)
     94 
     95         compare = self.cost_function(self.network_output, self.data_correct).mean()
     96         validator = theano.function([self.data_input, self.data_correct], compare,
     97                                     allow_input_downcast=True)

...........................................................................
/usr/local/lib/python2.7/dist-packages/theano/compile/function.pyc in function(inputs=[X, y, m], outputs=Elemwise{true_div,no_inplace}.0, mode=None, updates=OrderedDict([(<TensorType(float64, matrix)>, Ele...no_inplace}.0), (b, Elemwise{sub,no_inplace}.0)]), givens=[], no_default_updates=False, accept_inplace=False, name='/home/iarroyof/scikit-neuralnetwork/sknn/backend/lasagne/mlp.py:93', rebuild_strict=True, allow_input_downcast=True, profile=None, on_unused_input=u'ignore')
    261                 no_default_updates=no_default_updates,
    262                 accept_inplace=accept_inplace, name=name,
    263                 rebuild_strict=rebuild_strict,
    264                 allow_input_downcast=allow_input_downcast,
    265                 on_unused_input=on_unused_input,
--> 266                 profile=profile)
    267     # We need to add the flag check_aliased inputs if we have any mutable or
    268     # borrowed used defined inputs
    269     fn._check_for_aliased_inputs = check_for_aliased_inputs
    270     return fn

...........................................................................
/usr/local/lib/python2.7/dist-packages/theano/compile/pfunc.pyc in pfunc(params=[X, y, m], outputs=Elemwise{true_div,no_inplace}.0, mode=None, updates=OrderedDict([(<TensorType(float64, matrix)>, Ele...no_inplace}.0), (b, Elemwise{sub,no_inplace}.0)]), givens=[], no_default_updates=False, accept_inplace=False, name='/home/iarroyof/scikit-neuralnetwork/sknn/backend/lasagne/mlp.py:93', rebuild_strict=True, allow_input_downcast=True, profile=False, on_unused_input=u'ignore')
    506                     mutable=False, borrow=True, shared=True)
    507         inputs.append(si)
    508 
    509     return orig_function(inputs, cloned_outputs, mode,
    510             accept_inplace=accept_inplace, name=name, profile=profile,
--> 511             on_unused_input=on_unused_input)
    512 
    513 
    514 def _pfunc_param_to_in(param, strict=False, allow_downcast=None):
    515     if isinstance(param, Constant):

...........................................................................
/usr/local/lib/python2.7/dist-packages/theano/compile/function_module.pyc in orig_function(inputs=[In(X), In(y), In(m), In(W -> Elemwise{sub,no_inplace}.0), In(b -> Elemwise{sub,no_inplace}.0), In(W -> Elemwise{sub,no_inplace}.0), In(b -> Elemwise{sub,no_inplace}.0), In(W -> Elemwise{sub,no_inplace}.0), In(b -> Elemwise{sub,no_inplace}.0), In(<TensorType(float64, matrix)> -> Elemwise{add,no_inplace}.0), In(<TensorType(float64, vector)> -> Elemwise{add,no_inplace}.0), In(<TensorType(float64, matrix)> -> Elemwise{add,no_inplace}.0), In(<TensorType(float64, vector)> -> Elemwise{add,no_inplace}.0), In(<TensorType(float64, matrix)> -> Elemwise{add,no_inplace}.0), In(<TensorType(float64, vector)> -> Elemwise{add,no_inplace}.0)], outputs=Out(Elemwise{true_div,no_inplace}.0,False), mode=<theano.compile.mode.Mode object>, accept_inplace=False, name='/home/iarroyof/scikit-neuralnetwork/sknn/backend/lasagne/mlp.py:93', profile=False, on_unused_input=u'ignore')
   1460         fn = Maker(inputs,
   1461                    outputs,
   1462                    mode,
   1463                    accept_inplace=accept_inplace,
   1464                    profile=profile,
-> 1465                    on_unused_input=on_unused_input).create(
   1466                        defaults)
   1467 
   1468     t2 = time.time()
   1469     if profile:

...........................................................................
/usr/local/lib/python2.7/dist-packages/theano/compile/function_module.pyc in __init__(self=<theano.compile.function_module.FunctionMaker object>, inputs=[In(X), In(y), In(m), In(W -> Elemwise{sub,no_inplace}.0), In(b -> Elemwise{sub,no_inplace}.0), In(W -> Elemwise{sub,no_inplace}.0), In(b -> Elemwise{sub,no_inplace}.0), In(W -> Elemwise{sub,no_inplace}.0), In(b -> Elemwise{sub,no_inplace}.0), In(<TensorType(float64, matrix)> -> Elemwise{add,no_inplace}.0), In(<TensorType(float64, vector)> -> Elemwise{add,no_inplace}.0), In(<TensorType(float64, matrix)> -> Elemwise{add,no_inplace}.0), In(<TensorType(float64, vector)> -> Elemwise{add,no_inplace}.0), In(<TensorType(float64, matrix)> -> Elemwise{add,no_inplace}.0), In(<TensorType(float64, vector)> -> Elemwise{add,no_inplace}.0)], outputs=[Out(Elemwise{true_div,no_inplace}.0,False)], mode=<theano.compile.mode.Mode object>, accept_inplace=False, function_builder=<class 'theano.compile.function_module.Function'>, profile=None, on_unused_input=u'ignore', fgraph=[*3 -> Elemwise{true_div,no_inplace}(Elemwise{tr...,no_inplace}(<TensorType(float64, vector)>, *16)])
   1155                 # now optimize the graph
   1156                 if theano.config.cache_optimizations:
   1157                     optimizer_profile = self.optimize_graph_with_cache(
   1158                         optimizer, inputs, outputs)
   1159                 else:    
-> 1160                     optimizer_profile = optimizer(fgraph)
   1161                     
   1162                 end_optimizer = time.time()
   1163                 opt_time = end_optimizer - start_optimizer
   1164                 if profile:

...........................................................................
/usr/local/lib/python2.7/dist-packages/theano/gof/opt.pyc in __call__(self=[<theano.tensor.opt.ShapeOptimizer object at 0x7...gof.opt.MergeOptimizer object at 0x7f39d2791d10>], fgraph=[*3 -> Elemwise{true_div,no_inplace}(Elemwise{tr...,no_inplace}(<TensorType(float64, vector)>, *16)])
     85 
     86     def __call__(self, fgraph):
     87         """WRITEME
     88         Same as self.optimize(fgraph)
     89         """
---> 90         return self.optimize(fgraph)
     91 
     92     def add_requirements(self, fgraph):
     93         """WRITEME
     94         Add features to the fgraph that are required to apply the optimization.

...........................................................................
/usr/local/lib/python2.7/dist-packages/theano/gof/opt.pyc in optimize(self=[<theano.tensor.opt.ShapeOptimizer object at 0x7...gof.opt.MergeOptimizer object at 0x7f39d2791d10>], fgraph=[*3 -> Elemwise{true_div,no_inplace}(Elemwise{tr...,no_inplace}(<TensorType(float64, vector)>, *16)], *args=(), **kwargs={})
     76         """
     77         self.add_requirements(fgraph)
     78         try:
     79             orig = theano.tensor.basic.constant.enable
     80             theano.tensor.basic.constant.enable = False
---> 81             ret = self.apply(fgraph, *args, **kwargs)
     82         finally:
     83             theano.tensor.basic.constant.enable = orig
     84         return ret
     85 

...........................................................................
/usr/local/lib/python2.7/dist-packages/theano/gof/opt.pyc in apply(self=[<theano.tensor.opt.ShapeOptimizer object at 0x7...gof.opt.MergeOptimizer object at 0x7f39d2791d10>], fgraph=[*3 -> Elemwise{true_div,no_inplace}(Elemwise{tr...,no_inplace}(<TensorType(float64, vector)>, *16)])
    190         nb_node_before = len(fgraph.apply_nodes)
    191         sub_profs = []
    192         for optimizer in self:
    193             try:
    194                 t0 = time.time()
--> 195                 sub_prof = optimizer.optimize(fgraph)
    196                 l.append(float(time.time() - t0))
    197                 sub_profs.append(sub_prof)
    198                 if fgraph.profile:
    199                     sub_validate_time.append(fgraph.profile.validate_time)

...........................................................................
/usr/local/lib/python2.7/dist-packages/theano/gof/opt.pyc in optimize(self=<theano.gof.opt.EquilibriumOptimizer object>, fgraph=[*3 -> Elemwise{true_div,no_inplace}(Elemwise{tr...,no_inplace}(<TensorType(float64, vector)>, *16)], *args=(), **kwargs={})
     76         """
     77         self.add_requirements(fgraph)
     78         try:
     79             orig = theano.tensor.basic.constant.enable
     80             theano.tensor.basic.constant.enable = False
---> 81             ret = self.apply(fgraph, *args, **kwargs)
     82         finally:
     83             theano.tensor.basic.constant.enable = orig
     84         return ret
     85 

...........................................................................
/usr/local/lib/python2.7/dist-packages/theano/gof/opt.pyc in apply(self=<theano.gof.opt.EquilibriumOptimizer object>, fgraph=[*3 -> Elemwise{true_div,no_inplace}(Elemwise{tr...,no_inplace}(<TensorType(float64, vector)>, *16)], start_from=[Elemwise{true_div,no_inplace}.0, Elemwise{sub,no_inplace}.0, Elemwise{sub,no_inplace}.0, Elemwise{sub,no_inplace}.0, Elemwise{sub,no_inplace}.0, Elemwise{sub,no_inplace}.0, Elemwise{sub,no_inplace}.0, Elemwise{add,no_inplace}.0, Elemwise{add,no_inplace}.0, Elemwise{add,no_inplace}.0, Elemwise{add,no_inplace}.0, Elemwise{add,no_inplace}.0, Elemwise{add,no_inplace}.0])
   1832                     for lopt in (self.local_optimizers_all +
   1833                                  self.local_optimizers_map.get(type(node.op), []) +
   1834                                  self.local_optimizers_map.get(node.op, [])):
   1835                         nb = change_tracker.nb_imported
   1836                         t_opt = time.time()
-> 1837                         lopt_change = self.process_node(fgraph, node, lopt)
   1838                         time_opts[lopt] += time.time() - t_opt
   1839                         if lopt_change:
   1840                             process_count.setdefault(lopt, 0)
   1841                             process_count[lopt] += 1

...........................................................................
/usr/local/lib/python2.7/dist-packages/theano/gof/opt.pyc in process_node(self=<theano.gof.opt.EquilibriumOptimizer object>, fgraph=[*3 -> Elemwise{true_div,no_inplace}(Elemwise{tr...,no_inplace}(<TensorType(float64, vector)>, *16)], node=DimShuffle{x}(TensorConstant{2.0}), lopt=<theano.gof.opt.FromFunctionLocalOptimizer object>)
   1493             replacements = lopt.transform(node)
   1494         except Exception, e:
   1495             if self.failure_callback is not None:
   1496                 self.failure_callback(e, self,
   1497                                       [(x, None) for x in node.outputs],
-> 1498                                       lopt)
   1499                 return False
   1500             else:
   1501                 raise
   1502         if replacements is False or replacements is None:

...........................................................................
/usr/local/lib/python2.7/dist-packages/theano/gof/opt.pyc in warn_inplace(exc=AssertionError(), nav=<theano.gof.opt.EquilibriumOptimizer object>, repl_pairs=[(DimShuffle{x}.0, None)], local_opt=<theano.gof.opt.FromFunctionLocalOptimizer object>)
   1391 
   1392         ignore InconsistencyErrors, print traceback
   1393         """
   1394         if isinstance(exc, InconsistencyError):
   1395             return
-> 1396         return NavigatorOptimizer.warn(exc, nav, repl_pairs, local_opt)
   1397 
   1398     @staticmethod
   1399     def warn_ignore(exc, nav, repl_pairs, local_opt):
   1400         """failure_callback for NavigatorOptimizer: ignore all errors

...........................................................................
/usr/local/lib/python2.7/dist-packages/theano/gof/opt.pyc in warn(exc=AssertionError(), nav=<theano.gof.opt.EquilibriumOptimizer object>, repl_pairs=[(DimShuffle{x}.0, None)], local_opt=<theano.gof.opt.FromFunctionLocalOptimizer object>)
   1381         if config.on_opt_error == 'pdb':
   1382             pdb.post_mortem(sys.exc_info()[2])
   1383         elif isinstance(exc, AssertionError) or config.on_opt_error == 'raise':
   1384             # We always crash on AssertionError because something may be
   1385             # seriously wrong if such an exception is raised.
-> 1386             raise exc
   1387 
   1388     @staticmethod
   1389     def warn_inplace(exc, nav, repl_pairs, local_opt):
   1390         """failure_callback for NavigatorOptimizer

AssertionError: 

I suppose it is absolutely necessary carting any resulting data type to numpy array.