scikit-learn joblib bug: multiprocessing pool self.value out of range for 'i' format code, only with large numpy arrays

Question

My code runs fine with smaller test samples, like 10000 rows of data in X_train, y_train. When I call it for millions of rows, I get the resulting error. Is the bug in a package, or can I do something differently? I am using Python 2.7.7 from Anaconda 2.0.1, and I put the pool.py from Anaconda's multiprocessing package and parallel.py from scikit-learn's external package on my Dropbox for you.

The test script is:

import numpy as np
import sklearn
from sklearn.linear_model import SGDClassifier
from sklearn import grid_search
import multiprocessing as mp


def main():
    print("Started.")

    print("numpy:", np.__version__)
    print("sklearn:", sklearn.__version__)

    n_samples = 1000000
    n_features = 1000

    X_train = np.random.randn(n_samples, n_features)
    y_train = np.random.randint(0, 2, size=n_samples)

    print("input data size: %.3fMB" % (X_train.nbytes / 1e6))

    model = SGDClassifier(penalty='elasticnet', n_iter=10, shuffle=True)
    param_grid = [{
        'alpha' : 10.0 ** -np.arange(1,7),
        'l1_ratio': [.05, .15, .5, .7, .9, .95, .99, 1],
    }]
    gs = grid_search.GridSearchCV(model, param_grid, n_jobs=8, verbose=100)
    gs.fit(X_train, y_train)
    print(gs.grid_scores_)

if __name__=='__main__':
    mp.freeze_support()
    main()

This results in the output:

Vendor:  Continuum Analytics, Inc.
Package: mkl
Message: trial mode expires in 28 days
Started.
('numpy:', '1.8.1')
('sklearn:', '0.15.0b1')
input data size: 8000.000MB
Fitting 3 folds for each of 48 candidates, totalling 144 fits
Memmaping (shape=(1000000L, 1000L), dtype=float64) to new file c:\users\laszlos\appdata\local\temp\4\joblib_memmaping_pool_6172_78765976\6172-284752304-75223296-0.pkl
Failed to save <type 'numpy.ndarray'> to .npy file:
Traceback (most recent call last):
  File "C:\Anaconda\lib\site-packages\sklearn\externals\joblib\numpy_pickle.py", line 240, in save
    obj, filename = self._write_array(obj, filename)
  File "C:\Anaconda\lib\site-packages\sklearn\externals\joblib\numpy_pickle.py", line 203, in _write_array
    self.np.save(filename, array)
  File "C:\Anaconda\lib\site-packages\numpy\lib\npyio.py", line 453, in save
    format.write_array(fid, arr)
  File "C:\Anaconda\lib\site-packages\numpy\lib\format.py", line 406, in write_array
    array.tofile(fp)
ValueError: 1000000000 requested and 268435456 written

Memmaping (shape=(1000000L, 1000L), dtype=float64) to old file c:\users\laszlos\appdata\local\temp\4\joblib_memmaping_pool_6172_78765976\6172-284752304-75223296-0.pkl
Vendor:  Continuum Analytics, Inc.
Package: mkl
Message: trial mode expires in 28 days
Vendor:  Continuum Analytics, Inc.
Package: mkl
Message: trial mode expires in 28 days
Vendor:  Continuum Analytics, Inc.
Package: mkl
Message: trial mode expires in 28 days
Vendor:  Continuum Analytics, Inc.
Package: mkl
Message: trial mode expires in 28 days
Vendor:  Continuum Analytics, Inc.
Package: mkl
Message: trial mode expires in 28 days
Vendor:  Continuum Analytics, Inc.
Package: mkl
Message: trial mode expires in 28 days
Vendor:  Continuum Analytics, Inc.
Package: mkl
Message: trial mode expires in 28 days
Vendor:  Continuum Analytics, Inc.
Package: mkl
Message: trial mode expires in 28 days
Traceback (most recent call last):
  File "S:\laszlo\gridsearch_largearray.py", line 33, in <module>
    main()
  File "S:\laszlo\gridsearch_largearray.py", line 28, in main
    gs.fit(X_train, y_train)
  File "C:\Anaconda\lib\site-packages\sklearn\grid_search.py", line 597, in fit
    return self._fit(X, y, ParameterGrid(self.param_grid))
  File "C:\Anaconda\lib\site-packages\sklearn\grid_search.py", line 379, in _fit
    for parameters in parameter_iterable
  File "C:\Anaconda\lib\site-packages\sklearn\externals\joblib\parallel.py", line 651, in __call__
    self.retrieve()
  File "C:\Anaconda\lib\site-packages\sklearn\externals\joblib\parallel.py", line 503, in retrieve
    self._output.append(job.get())
  File "C:\Anaconda\lib\multiprocessing\pool.py", line 558, in get
    raise self._value
struct.error: integer out of range for 'i' format code

EDIT: ogrisel's answer does work with manual memory mapping with scikit-learn-0.15.0b1. Don't forget to run only one script at once, otherwise you can still run out of memory and have too many threads. (My run take ~60 GB on data of size ~12.5 GB in CSV, with 8 threads.)

Answer 1

As a workaround you can try to memory map your data explicitly & manually as explained in the joblib documentation.

Edit #1: Here is the important part:

from sklearn.externals import joblib

joblib.dump(X_train, some_filename)
X_train = joblib.load(some_filename, mmap_mode='r+')

Then pass this memmap'ed data to GridSearchCV under scikit-learn 0.15+.

Edit #2: Furthermore: if you use the 32bit version of Anaconda, you will be limited to 2GB per python process which can also limit the memory.

I just found a bug for numpy.save under Python 3.4 but even when fixed the subsequent call to mmap will fail with:

OSError: [WinError 8] Not enough storage is available to process this command

So please use a 64 bit version of Python (with Anaconda as AFAIK there is currently no other 64bit packages for numpy / scipy / scikit-learn==0.15.0b1 at this time).

Edit #3: I found another issue that might be causing excessive memory usage under windows: currently joblib.Parallel memory maps input data with mmap_mode='c' by default: this copy-on-write setting seems to cause windows to exhaust the paging file and sometimes triggers "[error 1455] the paging file is too small for this operation to complete" errors. Setting mmap_mode='r' or mmap_mode='r+' does not trigger that problem. I will run tests to see if I can change the default mode in the next version of joblib.