What to do if ʻObject arrays cannot be loaded when allow_pickle = False` occurs in numpy.load ()

--numpy 1.16.3 or later

phenomenon

Python code example

np.load('/path/to/file.npy')

Examples of errors that occur

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-37-1db66562b57b> in <module>
----> 1 np.load('tmp.npy')

~/venv/aep/lib/python3.7/site-packages/numpy/lib/npyio.py in load(file, mmap_mode, allow_pickle, fix_imports, encoding)
    451             else:
    452                 return format.read_array(fid, allow_pickle=allow_pickle,
--> 453                                          pickle_kwargs=pickle_kwargs)
    454         else:
    455             # Try a pickle

~/venv/aep/lib/python3.7/site-packages/numpy/lib/format.py in read_array(fp, allow_pickle, pickle_kwargs)
    720         # The array contained Python objects. We need to unpickle the data.
    721         if not allow_pickle:
--> 722             raise ValueError("Object arrays cannot be loaded when "
    723                              "allow_pickle=False")
    724         if pickle_kwargs is None:

ValueError: Object arrays cannot be loaded when allow_pickle=False

Cause

Since numpy v1.16.3, the behavior of thenumpy.load ()function has changed.

• np.load - NumPy v1.17 Manual

Solution

After confirming that there are no ** security concerns ** described later, specify the ʻallow_pickle` option as shown below.

np.load('/path/to/file.npy', allow_pickle=True)

Commentary

numpy matrix and dtype

The numpy matrix (np.ndarray) can store strings and Python objects as well as numbers. The type of stored value is reflected in the attribute dtype.

• Data type objects (dtype)

Vulnerability in numpy v1.16.0

A vulnerability has been reported that could allow malicious code to be executed when serializing a numpy matrix (a file that serializes) containing Python objects with np.load (). (However, there is a counterargument regarding this vulnerability)

• CVE-2019-6446

Therefore, from v1.16.3, the default behavior ofnp.load ()is changed as described above, and if dtype is a Python object, if ʻallow_pickle = False, ValueError` is thrown. It was way. It can be said that it is a specification change to push it to the safer side.

Security concerns

As a matter of course, don't np.load (allow_pickle = True) for ** untrusted files **. As mentioned in the previous section, it is possible to execute arbitrary code.

There is usually no problem with ad hoc code such as data formatting by Jupyter and machine learning [^ 1]. Note that application developers use Python.