Pickling large NumPy array

Question

I have a large 3d numpy array that I'd like to preserve. My first approach is simply to use pickle, but this seems to lead to a poorly explained error.

test_rand = np.random.random((100000,200,50))
with open('models/test.pkl', 'wb') as save_file:
    pickle.dump(test_rand, save_file, -1)

---------------------------------------------------------------------------
error                                     Traceback (most recent call last)
<ipython-input-18-511e30b08440> in <module>()
      1 with open('models/test.pkl', 'wb') as save_file:
----> 2         pickle.dump(test_rand, save_file, -1)
      3 

C:\Users\g1dak02\AppData\Local\Continuum\Anaconda\lib\pickle.pyc in dump(obj, file, protocol)
   1368 
   1369 def dump(obj, file, protocol=None):
-> 1370     Pickler(file, protocol).dump(obj)
   1371 
   1372 def dumps(obj, protocol=None):

C:\Users\g1dak02\AppData\Local\Continuum\Anaconda\lib\pickle.pyc in dump(self, obj)
    222         if self.proto >= 2:
    223             self.write(PROTO + chr(self.proto))
--> 224         self.save(obj)
    225         self.write(STOP)
    226 

C:\Users\g1dak02\AppData\Local\Continuum\Anaconda\lib\pickle.pyc in save(self, obj)
    329 
    330         # Save the reduce() output and finally memoize the object
--> 331         self.save_reduce(obj=obj, *rv)
    332 
    333     def persistent_id(self, obj):

C:\Users\g1dak02\AppData\Local\Continuum\Anaconda\lib\pickle.pyc in save_reduce(self, func, args, state, listitems, dictitems, obj)
    417 
    418         if state is not None:
--> 419             save(state)
    420             write(BUILD)
    421 

C:\Users\g1dak02\AppData\Local\Continuum\Anaconda\lib\pickle.pyc in save(self, obj)
    284         f = self.dispatch.get(t)
    285         if f:
--> 286             f(self, obj) # Call unbound method with explicit self
    287             return
    288 

C:\Users\g1dak02\AppData\Local\Continuum\Anaconda\lib\pickle.pyc in save_tuple(self, obj)
    560         write(MARK)
    561         for element in obj:
--> 562             save(element)
    563 
    564         if id(obj) in memo:

C:\Users\g1dak02\AppData\Local\Continuum\Anaconda\lib\pickle.pyc in save(self, obj)
    284         f = self.dispatch.get(t)
    285         if f:
--> 286             f(self, obj) # Call unbound method with explicit self
    287             return
    288 

C:\Users\g1dak02\AppData\Local\Continuum\Anaconda\lib\pickle.pyc in save_string(self, obj, pack)
    484                 self.write(SHORT_BINSTRING + chr(n) + obj)
    485             else:
--> 486                 self.write(BINSTRING + pack("<i", n) + obj)
    487         else:
    488             self.write(STRING + repr(obj) + '\n')

error: integer out of range for 'i' format code

So the two questions I have are as follows:

• What is actually going on in this error?
• How should I go about saving the array to disk?

I am using Python 2.7.8 and NumPy 1.9.0.

Answer 1

With regard to #1, it's a bug… and an old one at that. There's an enlightening, albeit surprisingly old, discussion about this here: http://python.6.x6.nabble.com/test-gzip-test-tarfile-failure-om-AMD64-td1830323.html

The reasons for the error are here: http://www.littleredbat.net/mk/files/grimoire.html#contents_item_2.1

The simplest and most basic type are integers, which are represented as a C long. Their size is therefore dependent on the platform you're using; on a 32-bit machine, they can range from -2147483647 to 2147483647. Python programs can determine the highest possible value for an integer by looking at sys.maxint; the lowest possible value will usually be -sys.maxint - 1.

This error is not a common one, as most people when faced with a very large numpy array, will use np.save or np.savez to take advantage of the reduced pickle format for numpy arrays (see the __reduce__ method for a numpy array, which is what np.save calls under the covers).

To show that it's just about the array being too large for pickle…

>>> import numpy as np
>>> import pickle
>>> test_rand = np.random.random((100000,200,50))
>>> x = pickle.dumps(test_rand[:20000], -1)
>>> x = pickle.dumps(test_rand[:30000], -1)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/Users/mmckerns/lib/python2.7/site-packages/dill-0.2.3.dev0-py2.7.egg/dill/dill.py", line 194, in dumps
    dump(obj, file, protocol, byref, fmode)#, strictio)
  File "/Users/mmckerns/lib/python2.7/site-packages/dill-0.2.3.dev0-py2.7.egg/dill/dill.py", line 184, in dump
    pik.dump(obj)
  File "/opt/local/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/pickle.py", line 224, in dump
    self.save(obj)
  File "/opt/local/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/pickle.py", line 286, in save
    f(self, obj) # Call unbound method with explicit self
  File "/Users/mmckerns/lib/python2.7/site-packages/dill-0.2.3.dev0-py2.7.egg/dill/dill.py", line 181, in save_numpy_array
    pik.save_reduce(_create_array, (f, args, state, npdict), obj=obj)
  File "/opt/local/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/pickle.py", line 401, in save_reduce
    save(args)
  File "/opt/local/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/pickle.py", line 286, in save
    f(self, obj) # Call unbound method with explicit self
  File "/opt/local/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/pickle.py", line 562, in save_tuple
    save(element)
  File "/opt/local/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/pickle.py", line 286, in save
    f(self, obj) # Call unbound method with explicit self
  File "/opt/local/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/pickle.py", line 562, in save_tuple
    save(element)
  File "/opt/local/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/pickle.py", line 286, in save
    f(self, obj) # Call unbound method with explicit self
  File "/opt/local/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/pickle.py", line 486, in save_string
    self.write(BINSTRING + pack("<i", n) + obj)
struct.error: 'i' format requires -2147483648 <= number <= 2147483647
>>> 

however, this works for the full array...

>>> x = test_rand.__reduce__()
>>> type(x)
<type 'tuple'>
>>> x[0]     
<built-in function _reconstruct>
>>> x[1]
(<type 'numpy.ndarray'>, (0,), 'b')
>>> x[2][0:3]
(1, (100000, 200, 50), dtype('float64'))
>>> len(x[2][4])
8000000000
>>> x[2][4][:100]
'Y\xa4}\xdf\x84\xdf\xe1?\xfe\x1fd\xe3\xf2\xab\xe2?\x80\xe4\xfe\x17\xfb\xd6\xc2?\xd73\x92\xc9N]\xe8?\x90\xbc\xe3@\xdcO\xc9?\x18\x9dX\x12MG\xc4?(\x0f\x8f\xf9}\xf6\xb1?\xd0\x90O\xe2\x9b\xf1\xed?_\x99\x06\xacY\x9e\xe2?\xe7\xf8\x15\xa8\x13\x91\xe2?\x96}\xffH\xda\xc3\xd4?@\t\xae_"\xe0\xda?y<%\x8a'

And if you'd like to burn out your fan, print x.

What you'll also notice is the function in x[0] gets saved along with the data. It's a self-contained function that can produce a numpy array from the pickled data.

Answer 2

As an alternative to pickle, especially for very large datasets, you may wish to consider a Python interface to a binary data format such as HDF5 (e.g., h5py). For a discussion of its pros and cons, see this question and the first answer.