I want to be able to save my array subclass to a npy file, and recover the result later.

Something like:

>>> class MyArray(np.ndarray): pass
>>> data = MyArray(np.arange(10))
>>> np.save('fname', data)
>>> data2 = np.load('fname')
>>> assert isinstance(data2, MyArray)  # raises AssertionError

the docs says (emphasis mine):

The format explicitly does not need to:

• [...]
• Fully handle arbitrary subclasses of numpy.ndarray. Subclasses will beaccepted for writing, but only the array data will be written out. Aregular numpy.ndarray object will be created upon reading the file.The API can be used to build a format for a particular subclass, butthat is out of scope for the general NPY format.

So is it possible to make the above code not raise an AssertionError?

I don't see evidence that np.save handles array subclasses.

I tried to save a np.matrix with it, and got back a ndarray.

I tried to save a np.ma array, and got an error

NotImplementedError: MaskedArray.tofile() not implemented yet.

Saving is done by np.lib.npyio.format.write_array, which does

_write_array_header()   # save dtype, shape etc

if dtype is object it uses pickle.dump(array, fp ...)

otherwise it does array.tofile(fp). tofile handles writing the data buffer.

I think pickle.dump of an array ends up using np.save, but I don't recall how that's triggered.

I can for example pickle an array, and load it:

In [657]: f=open('test','wb')
In [658]: pickle.Pickler(f).dump(x)
In [659]: f.close()
In [660]: np.load('test')
In [664]: f=open('test','rb')
In [665]: pickle.load(f)

This pickle dump/load sequence works for test np.ma, np.matrix and sparse.coo_matrix cases. So that's probably the direction to explore for your own subclass.

Searching on numpy and pickle I found Preserve custom attributes when pickling subclass of numpy array. The answer involves a custom .__reduce__ and .__setstate__.