CODEWITHSUNDEEP
pythonpython-3.xnumpynumba
pythonweb
pythonweb

Reputation: 1073

numba function will not compile

so I have a function in numba that is not compiling for some reason (calculates area under ROC curve). I am also not sure how I would be able to debug the function because my debugger is not working when I set a breakpoint in the numba decorated function.

Here is the function:

@nb.njit()
def auc_numba(fcst, obs):
    L = obs.size

    i_ord = fcst.argsort()

    sumV = 0.
    sumV2 = 0.
    sumW = 0.
    sumW2 = 0.
    n = 0
    m = 0
    i = 0

    while True:
        nn = mm = 0
        while True:
            j = i_ord[i]
            if obs[j]:
                mm += 1
            else:
                nn += 1
            if i == L - 1:
                break
            jp1 = i_ord[i + 1]
            if fcst[j] != fcst[jp1]:
                break
            i += 1
        sumW += nn * (m + mm / 2.0)
        sumW2 += nn * (m + mm / 2.0) * (m + mm / 2.0)
        sumV += mm * (n + nn / 2.0)
        sumV2 += mm * (n + nn / 2.0) * (n + nn / 2.0)
        n += nn
        m += mm
        i += 1
        if i >= L:
            break

    theta = sumV / (m * n)
    v = sumV2 / ((m - 1) * n * n) - sumV * sumV / (m * (m - 1) * n * n)
    w = sumW2 / ((n - 1) * m * m) - sumW * sumW / (n * (n - 1) * m * m)

    sd_auc = np.sqrt(v / m + w / n)

    return np.array([theta, sd_auc])

What I am thinking is that there is something wrong with the while loops that I have implemented. It's possible that the types are wrong, hence the break is not being activated and the function is running forever.

Here is some sample data to test:

obs = np.array([1, 0, 1, 1, 0, 1, 0, 1, 1, 0, 0, 1, 0, 0, 1, 1, 1, 0, 1, 1, 1, 0, 1, 1, 0, 0, 1])
fcst = np.array([0.7083333, 0.5416667, 0.875, 0.5833333, 0.2083333, 0.8333333, 0.1666667, 0.9583333, 0.625, 0.1666667, 0.5, 1.0, 0.6666667, 0.2083333, 0.875, 0.75, 0.625, 0.3333333, 0.8333333, 0.2083333, 0.125, 0.0, 0.875, 0.8333333, 0.125, 0.5416667, 0.75])

When I run this without the decorator I get [0.89488636 0.06561209] which are the correct values.

So I guess if I could just get some help understanding why it is not compiling and maybe some tips on how to debug in numba?

Upvotes: 2

Views: 226

Answers (1)

JoshAdel
JoshAdel

Reputation: 68732

There is something strange going on with the double while True loop. For whatever reason (and I don't understand it), if you create two variables x and y at the top and then:

x = 1
y = 0

while True:
    nn = mm = 0
    while x > y:

and keep everything else the same, the code works. I'm going to submit an issue to the Numba tracker since this seems like a bug to me.

Update: The numba issue can be found here

Upvotes: 3

Related Questions