Hi, I was trying to read directly from a compressed file:

data = np.array([9, 8, 7, 6, 5, 4, 3, 2, 1])
dt = data.type
with gzip.open("datafile.gz", "wb") as outfile:
    outfile.write(data.tobytes())

with gzip.open("datafile.gz", "rb") as infile:
     data = np.fromfile(infile, dtype=dt)
     print(data)

Unfortunately, this returns garbage data:
[6542475788951259935 7594864974085029634 1008947487324530028 3113290099057413416 691954333390038676 6865743131250406218]

The workaround is to load the data into a buffer first, and then let np.frombuffer read it:

with gzip.open("datafile.gz", "rb") as infile:
     data = np.frombuffer(infile.read(), dtype=dt)
     print(data)

Returns: [9 8 7 6 5 4 3 2 1] as expected.

I think silently returning corrupt data is pretty much the worst possible behaviour in this case. Either of these options would be an improvement:

• Teach np.fromfile() to read from gzip (and possible other compression formats) file objects
• Raise an error explaining that NumPy can't deal with this kind of file object
• At least clearly document what kind of "file object" np.fromfile() expects.

Right now, the file parameter of np.fromfile is documented as:

file : file or str
Open file object or filename.

And the file object returned from gzip.open is a file object, but apparently not the right kind. That's pretty confusing and should at least be changed in the documentation.

Operating System: Fedora 27, 64-Bit
Python version: 3.6.4
NumPy version: 1.13.3