sparse matrix add() fail

[benureau]

add() on a sparse matrix fails.

$ python test_voc.py
.E
======================================================================
ERROR: test_add_coherent (__main__.TestAdd)
Test that all implementations are coherent
----------------------------------------------------------------------
Traceback (most recent call last):
  File "test_voc.py", line 59, in test_add_coherent
    v.add(3, 4, 1)
  File "/Users/fabien/research/enc/coda/naminggamesal/lib/ngvoc/__init__.py", line 32, in mod_fun
    ans = tempfun(obj_self, *args)
  File "/Users/fabien/research/enc/coda/naminggamesal/lib/ngvoc/matrix.py", line 77, in add
    self._content[m,w] = val
  File "/Users/fabien/.pyenv/versions/2.7.6/lib/python2.7/site-packages/scipy/sparse/lil.py", line 286, in __setitem__
    i, j, x, self.dtype)
  File "_csparsetools.pyx", line 102, in scipy.sparse._csparsetools.lil_insert (scipy/sparse/_csparsetools.c:3424)
KeyError: dtype('float16')

----------------------------------------------------------------------

It seems due to this line:

naminggamesal/lib/ngvoc/matrix.py

Line 24 in c46cab9

self._content=np.matrix(np.zeros((self._M,self._W),dtype=np.float16))

Is there a reason to use 'float16' explicitly ?

[wschuell]

I put float16 data type so that it would take less memory, as required level of precision is really low (but shouldn't be just integer or binary, for future use of scores per associations). The reason is that databases would take several times less memory.
I think this problem has been solved in latest numpy versions, as i don't get this error.

[benureau]

Fails for me on both 1.9.0 et 1.9.2. Python version 2.7.6 on OS.X:

Python 2.7.6 (default, Mar 10 2014, 20:11:44)
[GCC 4.2.1 Compatible Apple LLVM 5.0 (clang-500.2.79)] on darwin

[benureau]

Was wrong, this line is the culprit:

naminggamesal/lib/ngvoc/matrix.py

Line 243 in c46cab9

self._content=sparse.lil_matrix((self._M,self._W),dtype=np.float16)

Changing it to:

        self._content=sparse.lil_matrix((self._M,self._W))

solves the problem.