Nifti1Image design prevents efficient caching

arthurmensch commented 7 years ago

This is a long-standing issue, that we have already tried to adress without success.


from nilearn.datasets import fetch_adhd
from sklearn.externals.joblib import Memory
from nilearn.input_data import MultiNiftiMasker

mem = Memory(cachedir='joblib')
mem.clear()

dataset = fetch_adhd().func[:2]

masker = MultiNiftiMasker(n_jobs=2, memory=mem).fit(dataset)
masker.transform(dataset)
masker = MultiNiftiMasker(n_jobs=1, memory=mem).fit(dataset)
masker.transform(dataset)

yields

/opt/miniconda3/bin/python /home/arthur/work/repos/nilearn/examples/test.py
/opt/miniconda3/lib/python3.5/site-packages/sklearn/externals/joblib/logger.py:77: DeprecationWarning: The 'warn' function is deprecated, use 'warning' instead
  logging.warn("[%s]: %s" % (self, msg))
WARNING:root:[Memory(cachedir='joblib/joblib')]: Flushing completely the cache
/home/arthur/work/repos/nilearn/nilearn/_utils/cache_mixin.py:272: UserWarning: memory_level is currently set to 0 but a Memory object has been provided. Setting memory_level to 1.
  warnings.warn("memory_level is currently set to 0 but "
________________________________________________________________________________
[Memory] Calling nilearn.masking.compute_multi_background_mask...
compute_multi_background_mask([ '/home/arthur/data/nilearn_data/adhd/data/0010042/0010042_rest_tshift_RPI_voreg_mni.nii.gz',
  '/home/arthur/data/nilearn_data/adhd/data/0010064/0010064_rest_tshift_RPI_voreg_mni.nii.gz'], n_jobs=2, memory=Memory(cachedir='joblib/joblib'), target_shape=None, verbose=0, target_affine=None)
________________________________________________________________________________
[Memory] Calling nilearn.image.image._compute_mean...
_compute_mean(<nibabel.nifti1.Nifti1Image object at 0x7f22ac2f9f60>, target_shape=None, smooth=False, target_affine=None)
________________________________________________________________________________
[Memory] Calling nilearn.image.image._compute_mean...
_compute_mean(<nibabel.nifti1.Nifti1Image object at 0x7f22ac2f9d30>, target_shape=None, smooth=False, target_affine=None)
_____________________________________________________compute_mean - 0.4s, 0.0min
_____________________________________________________compute_mean - 0.4s, 0.0min
____________________________________compute_multi_background_mask - 1.6s, 0.0min
________________________________________________________________________________
[Memory] Calling nilearn.image.resampling.resample_img...
resample_img(<nibabel.nifti1.Nifti1Image object at 0x7f22ad366908>, target_shape=None, interpolation='nearest', copy=False, target_affine=None)
_____________________________________________________resample_img - 0.0s, 0.0min
________________________________________________________________________________
[Memory] Calling nilearn.input_data.nifti_masker.filter_and_mask...
filter_and_mask(<nibabel.nifti1.Nifti1Image object at 0x7f22ac307f98>, <nibabel.nifti1.Nifti1Image object at 0x7f22ac307160>, { 'detrend': False,
  'high_pass': None,
  'low_pass': None,
  'smoothing_fwhm': None,
  'standardize': False,
  't_r': None,
  'target_affine': None,
  'target_shape': None}, confounds=None, copy=True, memory=Memory(cachedir='joblib/joblib'), verbose=0, memory_level=1)
________________________________________________________________________________
[Memory] Calling nilearn.input_data.nifti_masker.filter_and_mask...
filter_and_mask(<nibabel.nifti1.Nifti1Image object at 0x7f22ac307eb8>, <nibabel.nifti1.Nifti1Image object at 0x7f22ac307780>, { 'detrend': False,
  'high_pass': None,
  'low_pass': None,
  'smoothing_fwhm': None,
  'standardize': False,
  't_r': None,
  'target_affine': None,
  'target_shape': None}, confounds=None, copy=True, memory=Memory(cachedir='joblib/joblib'), verbose=0, memory_level=1)
__________________________________________________filter_and_mask - 1.6s, 0.0min
__________________________________________________filter_and_mask - 1.6s, 0.0min
________________________________________________________________________________
[Memory] Calling nilearn.input_data.nifti_masker.filter_and_mask...
filter_and_mask(<nibabel.nifti1.Nifti1Image object at 0x7f22ac3072b0>, <nibabel.nifti1.Nifti1Image object at 0x7f22ac2f9a90>, { 'detrend': False,
  'high_pass': None,
  'low_pass': None,
  'smoothing_fwhm': None,
  'standardize': False,
  't_r': None,
  'target_affine': None,
  'target_shape': None}, confounds=None, memory=Memory(cachedir='joblib/joblib'), memory_level=1, verbose=0, copy=True)
__________________________________________________filter_and_mask - 1.0s, 0.0min
________________________________________________________________________________
[Memory] Calling nilearn.input_data.nifti_masker.filter_and_mask...
filter_and_mask(<nibabel.nifti1.Nifti1Image object at 0x7f22ac344c18>, <nibabel.nifti1.Nifti1Image object at 0x7f22ac2f9a90>, { 'detrend': False,
  'high_pass': None,
  'low_pass': None,
  'smoothing_fwhm': None,
  'standardize': False,
  't_r': None,
  'target_affine': None,
  'target_shape': None}, confounds=None, memory=Memory(cachedir='joblib/joblib'), memory_level=1, verbose=0, copy=True)
__________________________________________________filter_and_mask - 1.0s, 0.0min

Process finished with exit code 0

This indicates that the cached mask function is not properly recalled the second time.

With the following monkey-patch, it works properly:

import nibabel
from nibabel import Nifti1Image as NibabelNifti1Image
from nibabel import load as nibabel_load
import numpy as np

def load(filename, **kwargs):
    img = nibabel_load(filename, **kwargs)
    img.__class__ = Nifti1Image
    return img

class Nifti1Image(NibabelNifti1Image):
    def __getstate__(self):
        state = {'dataobj': np.asarray(self._dataobj),
                 'header': self.header,
                 'affine': self.affine,
                 'extra': self.extra}
        return state

    def __setstate__(self, state):
        new_self = Nifti1Image(dataobj=state['dataobj'],
                               affine=state['affine'],
                               header=state['header'],
                               extra=state['extra'])
        self.__dict__ = new_self.__dict__

def monkey_patch_nifti_image():
    nibabel.load = load

from nilearn.datasets import fetch_adhd
from sklearn.externals.joblib import Memory
from nilearn.input_data import MultiNiftiMasker

monkey_patch_nifti_image()

mem = Memory(cachedir='joblib')
mem.clear()

dataset = fetch_adhd().func[:2]

masker = MultiNiftiMasker(n_jobs=2, memory=mem).fit(dataset)
masker.transform(dataset)
masker = MultiNiftiMasker(n_jobs=1, memory=mem).fit(dataset)
masker.transform(dataset)

/opt/miniconda3/bin/python /home/arthur/work/repos/nilearn/examples/test.py
/opt/miniconda3/lib/python3.5/site-packages/sklearn/externals/joblib/logger.py:77: DeprecationWarning: The 'warn' function is deprecated, use 'warning' instead
  logging.warn("[%s]: %s" % (self, msg))
WARNING:root:[Memory(cachedir='joblib/joblib')]: Flushing completely the cache
/home/arthur/work/repos/nilearn/nilearn/_utils/cache_mixin.py:272: UserWarning: memory_level is currently set to 0 but a Memory object has been provided. Setting memory_level to 1.
  warnings.warn("memory_level is currently set to 0 but "
________________________________________________________________________________
[Memory] Calling nilearn.masking.compute_multi_background_mask...
compute_multi_background_mask([ '/home/arthur/data/nilearn_data/adhd/data/0010042/0010042_rest_tshift_RPI_voreg_mni.nii.gz',
  '/home/arthur/data/nilearn_data/adhd/data/0010064/0010064_rest_tshift_RPI_voreg_mni.nii.gz'], n_jobs=2, verbose=0, target_shape=None, target_affine=None, memory=Memory(cachedir='joblib/joblib'))
________________________________________________________________________________
[Memory] Calling nilearn.image.image._compute_mean...
_compute_mean(<__main__.Nifti1Image object at 0x7f3ea9140cf8>, target_affine=None, target_shape=None, smooth=False)
________________________________________________________________________________
[Memory] Calling nilearn.image.image._compute_mean...
_compute_mean(<__main__.Nifti1Image object at 0x7f3ea9140f98>, target_affine=None, target_shape=None, smooth=False)
_____________________________________________________compute_mean - 1.2s, 0.0min
_____________________________________________________compute_mean - 1.2s, 0.0min
____________________________________compute_multi_background_mask - 3.3s, 0.1min
________________________________________________________________________________
[Memory] Calling nilearn.image.resampling.resample_img...
resample_img(<__main__.Nifti1Image object at 0x7f3eaa1aa8d0>, copy=False, target_affine=None, target_shape=None, interpolation='nearest')
_____________________________________________________resample_img - 0.0s, 0.0min
________________________________________________________________________________
[Memory] Calling nilearn.input_data.nifti_masker.filter_and_mask...
filter_and_mask(<__main__.Nifti1Image object at 0x7f3ea9140fd0>, <__main__.Nifti1Image object at 0x7f3ea914b7f0>, { 'detrend': False,
  'high_pass': None,
  'low_pass': None,
  'smoothing_fwhm': None,
  'standardize': False,
  't_r': None,
  'target_affine': None,
  'target_shape': None}, copy=True, verbose=0, memory=Memory(cachedir='joblib/joblib'), confounds=None, memory_level=1)
________________________________________________________________________________
[Memory] Calling nilearn.input_data.nifti_masker.filter_and_mask...
filter_and_mask(<__main__.Nifti1Image object at 0x7f3ea914b828>, <__main__.Nifti1Image object at 0x7f3ea914b048>, { 'detrend': False,
  'high_pass': None,
  'low_pass': None,
  'smoothing_fwhm': None,
  'standardize': False,
  't_r': None,
  'target_affine': None,
  'target_shape': None}, copy=True, verbose=0, memory=Memory(cachedir='joblib/joblib'), confounds=None, memory_level=1)
__________________________________________________filter_and_mask - 1.8s, 0.0min
__________________________________________________filter_and_mask - 1.6s, 0.0min

Process finished with exit code 0

Integrating the patch in nilearn does not break any test. Should I try to merge this ? I guess we should do a non regression test based on the code snippet above.

The issue is pretty preocuppying as all users are likely to have double-size masking cache on their disk

arthurmensch commented 7 years ago

Otherwise we could use (ping@ogrisel @GaelVaroquaux). It would be safe and efficient.

class UnsafeNifti1Image(Nifti1Image):
    def __getstate__(self):
        if self.get_filename() is not None:
            filename = self.get_filename()
            stat = os.stat(filename)
            last_modified = stat.st_mtime
            state = {'filename': filename,
                     'last_modified': last_modified}
        else:
            state = {'dataobj': np.asarray(self._dataobj),
                     'header': self.header,
                     'affine': self.affine,
                     'extra': self.extra}
        return state

    def __setstate__(self, state):
        if 'filename' in state:
            print('unpickling')
            new_self = Nifti1Image.from_filename(state['filename'])
        else:
            new_self = Nifti1Image(dataobj=state['dataobj'],
                                   affine=state['affine'],
                                   header=state['header'],
                                   extra=state['extra'])
        self.__dict__ = new_self.__dict__

arthurmensch commented 7 years ago

Here is the full gist that:

Pickles and unpickles Nifti1Image using only their relevant state
Hash them using either these states or, if they exist, the filenames and the modification time of the associated .nii(.gz) files.

https://gist.github.com/arthurmensch/3331024db4ce08510b3bd51a0469c322

The real proper way to integrate it would be to

be able to provide a custom hash function for each of the arguments of a cache function in joblib (a feature that has already been requested according to @ogrisel)
Propose the modification to the nibabel team - they should take it as it is the proper way of pickling Nifti1Image...
Use monkey-patches in nilearn for previous versions of nibabel and joblib

ping @GaelVaroquaux what's your opinion on this ?

Remi-Gau commented 9 months ago

Maybe worth rechecking if this issue is still valid

nilearn / nilearn

Nifti1Image design prevents efficient caching #1362