Pickle - Cpickle.Unpicklingerror: Invalid Load Key, '?'

cPickle.UnpicklingError: invalid load key, ' '.?

The argument you've passed to cPickle.load() has to be a .pkl file.
mnist.pkl is provided inside of mnist.pkl.gz

So, you have to open that .gz first. Try this:

import gzip
f = gzip.open('mnist.pkl.gz', 'rb')
train_set, valid_set, test_set = cPickle.load(f)

cPickle: UnpicklingError: invalid load key, 'A'

When storing multiple objects (by repeated dump, not from containers) Pickle will store objects sequentially in pickle files, so if an object is broken it can be removed without corrupting the others.

In principle, the pickle format is pseudo-documented in pickle.py. For most cases, the opcodes at the beginning of the module are sufficient to piece together what is happening. Basically, pickle files are an instruction on how to build objects.

How readable a pickle file is depends on its pickle format - 0 is doable, everything above is difficult. Whether you can fix or must delete depends entirely on this. What's consistent is that each individual pickle ends with a dot (.). For example, b'Va\np0\n.' and b'\x80\x04\x95\x05\x00\x00\x00\x00\x00\x00\x00\x8c\x01a\x94.' both are the character '"a"', but in protocol 0 and 4.

The simplest form of recovery is to count the number of objects you can load:

with open('/my/pickle.pkl', 'rb') as pkl_source:
    idx = 1
    while True:
        pickle.load(pkl_source)
        print(idx)
        idx += 1

Then open the pickle file, skip as many objects and remove everything up to the next ..

What causes the error _pickle.UnpicklingError: invalid load key, ' '. ?

pickling is recursive, not sequential. Thus, to pickle a list, pickle will start to pickle the containing list, then pickle the first element… diving into the first element and pickling dependencies and sub-elements until the first element is serialized. Then moves on to the next element of the list, and so on, until it finally finishes the list and finishes serializing the enclosing list. In short, it's hard to treat a recursive pickle as sequential, except for some special cases. It's better to use a smarter pattern on your dump, if you want to load in a special way.

The most common pickle, it to pickle everything with a single dump to a file -- but then you have to load everything at once with a single load. However, if you open a file handle and do multiple dump calls (e.g. one for each element of the list, or a tuple of selected elements), then your load will mirror that… you open the file handle and do multiple load calls until you have all the list elements and can reconstruct the list. It's still not easy to selectively load only certain list elements, however. To do that, you'd probably have to store your list elements as a dict (with the index of the element or chunk as the key) using a package like klepto, which can break up a pickled dict into several files transparently, and enables easy loading of specific elements.

Saving and loading multiple objects in pickle file?

---

Related Topics