all output are nan?!

[Hualin]

hey, thanks for you code. but I run the sammontest.py, both raw and distance output are nan, because E-E_new array is all nan.

at first glance i am not sure where to fix it, since there Dinv but you catch the inf error and replace it with 0.

[tompollard]

Hi Hualin, thanks for your message. I haven't looked at the code in a while, but I'll check this out when I get a chance. The code could do definitely do with some cleaning up.

The functionality was based on a Matlab function, which I remember would return NaNs in the following circumstances:

• when there are two identical data points in the input array.
• where there was a NaN in the input array

Do either of these issues apply here? i.e. please could you try removing (1) all NaNs from your input array and (2) any duplicate data points from your input array and let me know if this helps?

[Hualin]

Cool, thanks for your fast reply. I guess sum of E_new or E is NaN because there was a value NaN in the array of E. Perhaps, delta = D-d.

I believe I should do some cleanup routine on distance matrix for D and d, it should fix something. I will be back here.

[Hualin]

Hi there. I believe I got it fixed. Though I drop something:

1. dropinputdist = "distance", always give vector space. (it can be hacked if one need to give distance matrix directly)
2. drop euclid function, use scipy.spetial.distance.cdist instead.
3. change some naming conversions: x => X, y=>Y, D=>xD, d=>yD, Dinv=>xDinv, dinv=>yDinv (the code below use these names)
4. make some change about xDInv build:
   yD = distance.cdist(Y,Y)

np.fill_diagonal(xD, 1)
np.fill_diagonal(yD, 1)

xDinv = 1.0/xD
yDinv = 1.0/yD
np.fill_diagonal(xDinv, 0)
np.fill_diagonal(yDinv, 0)
xDinv[np.isnan(xDinv)] = 0
yDinv[np.isnan(yDinv)] = 0
xDinv[np.isinf(xDinv)] = 0
yDinv[np.isinf(yDinv)] = 0

and somehow. those changes make error gone! I didn't touch the optimisation process.

cheers,

[tompollard]

Thanks Hualin - please could you put the suggested changes into a pull request?

[tompollard]

This should have been addressed in the merged pull request.