Make a new release?

[cvjjm]

It seems that the 1.3 release of autograd that is available via pypi has a few bugs that are fixed in the current master.

As an example consider this code:

from autograd import numpy as np
from autograd import jacobian
def g(m):
    return np.linalg.eigh(m)[1]
m = np.array([[1, 2],[2, 3]])
dg_dm = jacobian(g)(m)
print(dg_dm)

delta = 1e-6
dg_dm_fd = np.zeros(list(m.shape)*2)
for x in range(len(m)):
    for y in range(len(m)):
        if y > x:
            continue
        displacement = np.zeros_like(m, dtype=float)
        displacement[x, y] = delta
        if x != y:
            displacement[y, x] = delta
        #print(delta, displacement, (g(m+displacement) - g(m-displacement))/(2*delta))
        dg_dm_fd[x, y] += (g(m+displacement) - g(m-displacement))/(2*delta)

dg_dm_fd = dg_dm_fd.transpose((2, 3, 0, 1))
print(dg_dm_fd)
assert np.allclose(dg_dm, dg_dm_fd), str(dg_dm - dg_dm_fd)

which works correctly with master but fails with pip install autograd.

Would be awesome if a 1.4 release could be made so that code that depends on this working can declare a dependency on autograd 1.4. Thanks very much!

[j-towns]

Done, hope that helps!

[cvjjm]

Awesome!