dvolkow / mle Goto Github PK

Right, you've said the method works not very well with poly degree > 1, and it would be understandable to see some function like vr(params, degree) that works with any degree, but degree=1 is hardcoded for now; but I dont see even a support for higher degrees. That's bad, please, generalize this.

In mle_io.py:

f = open(DATASET_FILE, 'r')
reader = csv.reader(f)
#reading

In modern python you'd rather use context manager semantics:

with open(DATASET_FILE, 'r') as f:
    reader = csv.reader(f)
   #reading

This is better due to automatic cleanup in any case after leaving the with scope.

I will mark non-critical issues as [codestyle], as there are no issue labels defined in the repo (or I can't assign one).

Generally, numpy is much faster then python native arrays (altho in this case I didn't achieve significant speed-up by merely using numpy everywhere).

Numpy lets you to carry operations on whole arrays (much like in R):

vr = vr_calc_all(R_0, A) + v_sun_all(u, v, w) # this is array already

sigma = stdev(vr)

L     = - len(data.l) * math.log(math.sqrt(2 * math.pi) * sigma)
vr_diff = vr - data.vr
vr_diff_sqr = np.power(vr_diff, 2)
L -= np.sum(vr_diff_sqr) / (2 * sigma ** 2)

I also changed lists to numpy arrays everywhere to achieve array-wise operations, like:

def r_all(self, r0):
    if r0 != self.cached_value[0]:
        r0s = np.repeat(r0, self.distcos.shape[0])
        self.cached_value = (r0, np.sqrt(np.power(r0s, 2) + self.distcos - self.dblcos * r0))
    return self.cached_value[1]

'minimize' function you use supports the 'callback' argument it calls after each iteration, you can insert debug output there.
The difference with what you did is that it is called only after each full iteration; the likelyhood_log function, on the other hand, is called many times per iteration (every time minimizer needs any value for calculations, say, it's called multiple times only to numerically calculate a singe derivative value).

• setup.py
• setup.cfg

In python, you can assign values from tuple-style object like this:

R_0, A, u, v, w = parameters

This is more pythonic then assigning them line-by-line.

For compare and control results.

Now model provide only linear form estimation polinom.

Need for some optimizers into scipy methods, as jac parameter.

Please, provide full reference results you achieved via LSE earlier.

Now only point-type result.