pure python valid region does not behave as shown in docs

[butlerpd]

If there are regions of parameter space which are not valid, a model function can exclude them from fitting according to the docs. However the C models do it differently and require a valid = (Boolean expression) while the python model is told to return np.nan.

Recent attempts at using this second version shows it does not actually work for fitting. Upon investigation @pkienzle believes that the following code placed at line 225 of kernelpy.py] might fix the problem

total = form()
if np.isnan(total).any():
    total = np.zeros(nq, 'd')

That said, a preferred solution would be to make the two interfaces the same. That however would require parsing a C style Boolean expression.

[pkienzle]

Should avoid computing the form volume and form radius since they are also likely to be invalid. The code becomes:

if call_details.num_active == 0:
    weight_norm = 1.0
    total = form()
    if np.isnan(total).any():
        total = np.zeros(nq, 'd')
        weighted_shell = weighted_form = weighted_radius = 0
    else:
        weighted_shell, weighted_form = form_volume()
        weighted_radius = form_radius()

[pkienzle]

Current code looks like returning NaN should work:

sasmodels/sasmodels/kernelpy.py

Lines 268 to 278 in 00fd024

	Iq = np.asarray(form(), 'd')
	if np.isnan(Iq).any():
	continue
	# Update value and norm.
	total += weight * Iq
	weight_norm += weight
	unweighted_shell, unweighted_form = form_volume()
	weighted_shell += weight * unweighted_shell
	weighted_form += weight * unweighted_form
	weighted_radius += weight * form_radius()

For valid = ... we could do a simple string substitution with "!"=>" not ", "&&" => " and ", and "||" => " or ", then use eval() on the expression. If the result is true then proceed with the loop otherwise continue. Various checkers will raise flags that eval() is unsafe; we ought to ignore them since the rest of the model is running arbitrary python code supplied by the user, but they can be persistent.

Note that this is the same place we would need to insert a call to transform the function arguments if we want reparameterize() to support python models.