Use pseudo-inverse computation

[chuckwondo]

Use numpy.linalg.pinv instead of numpy.linalg.inv to avoid "singular matrix" errors.

Upgrade numpy in Dockerfile to avoid "hanging" during computation of pseudo-inverse.

Fixes #7

[chuckwondo]

@arthurduf, please take a look at this. I suggest pulling the branch and testing.

[arthurduf]

All good for me

[nemo794]

Just a heads up that numpy.linalg.pinv takes a significantly longer to run than numpy.linalg.inv. (That said, I confirmed via htop that pinv does use all available cores, so we're good there.)

Is this what we want?

>>> import timeit
>>> inv = '''\
... import numpy as np
... pos = np.random.randint(0, 10, size=(5000,5000))
... np.linalg.inv(pos)
... '''
>>> pinv = '''\
... import numpy as np
... pos = np.random.randint(0, 10, size=(5000,5000))
... np.linalg.pinv(pos)
... '''
>>> timeit.timeit(inv, number=1)
4.161739458999364
>>> timeit.timeit(pinv, number=1)
43.71366904099705

[chuckwondo]

Just a heads up that numpy.linalg.pinv takes a significantly longer to run than numpy.linalg.inv. (That said, I confirmed via htop that pinv does use all available cores, so we're good there.)

Is this what we want?

>>> import timeit
>>> inv = '''\
... import numpy as np
... pos = np.random.randint(0, 10, size=(5000,5000))
... np.linalg.inv(pos)
... '''
>>> pinv = '''\
... import numpy as np
... pos = np.random.randint(0, 10, size=(5000,5000))
... np.linalg.pinv(pos)
... '''
>>> timeit.timeit(inv, number=1)
4.161739458999364
>>> timeit.timeit(pinv, number=1)
43.71366904099705

Generally speaking, as long as this has the intended effect of consuming resources, I'd say this is fine. The actual computation is not what we care about, is it? Unless this particular calculation is going to cause the larger bboxes to take inordinately long (hours?) to complete, this should be fine, no?