1 | """ |
---|
2 | Python driver for python kernels |
---|
3 | |
---|
4 | Calls the kernel with a vector of $q$ values for a single parameter set. |
---|
5 | Polydispersity is supported by looping over different parameter sets and |
---|
6 | summing the results. The interface to :class:`PyModel` matches those for |
---|
7 | :class:`kernelcl.GpuModel` and :class:`kerneldll.DllModel`. |
---|
8 | """ |
---|
9 | from __future__ import division, print_function |
---|
10 | |
---|
11 | import numpy as np # type: ignore |
---|
12 | from numpy import pi, cos #type: ignore |
---|
13 | |
---|
14 | from . import details |
---|
15 | from .generate import F64 |
---|
16 | from .kernel import KernelModel, Kernel |
---|
17 | |
---|
18 | try: |
---|
19 | from typing import Union, Callable |
---|
20 | except: |
---|
21 | pass |
---|
22 | else: |
---|
23 | DType = Union[None, str, np.dtype] |
---|
24 | |
---|
25 | class PyModel(KernelModel): |
---|
26 | """ |
---|
27 | Wrapper for pure python models. |
---|
28 | """ |
---|
29 | def __init__(self, model_info): |
---|
30 | # Make sure Iq and Iqxy are available and vectorized |
---|
31 | _create_default_functions(model_info) |
---|
32 | self.info = model_info |
---|
33 | |
---|
34 | def make_kernel(self, q_vectors): |
---|
35 | q_input = PyInput(q_vectors, dtype=F64) |
---|
36 | kernel = self.info.Iqxy if q_input.is_2d else self.info.Iq |
---|
37 | return PyKernel(kernel, self.info, q_input) |
---|
38 | |
---|
39 | def release(self): |
---|
40 | """ |
---|
41 | Free resources associated with the model. |
---|
42 | """ |
---|
43 | pass |
---|
44 | |
---|
45 | class PyInput(object): |
---|
46 | """ |
---|
47 | Make q data available to the gpu. |
---|
48 | |
---|
49 | *q_vectors* is a list of q vectors, which will be *[q]* for 1-D data, |
---|
50 | and *[qx, qy]* for 2-D data. Internally, the vectors will be reallocated |
---|
51 | to get the best performance on OpenCL, which may involve shifting and |
---|
52 | stretching the array to better match the memory architecture. Additional |
---|
53 | points will be evaluated with *q=1e-3*. |
---|
54 | |
---|
55 | *dtype* is the data type for the q vectors. The data type should be |
---|
56 | set to match that of the kernel, which is an attribute of |
---|
57 | :class:`GpuProgram`. Note that not all kernels support double |
---|
58 | precision, so even if the program was created for double precision, |
---|
59 | the *GpuProgram.dtype* may be single precision. |
---|
60 | |
---|
61 | Call :meth:`release` when complete. Even if not called directly, the |
---|
62 | buffer will be released when the data object is freed. |
---|
63 | """ |
---|
64 | def __init__(self, q_vectors, dtype): |
---|
65 | self.nq = q_vectors[0].size |
---|
66 | self.dtype = dtype |
---|
67 | self.is_2d = (len(q_vectors) == 2) |
---|
68 | if self.is_2d: |
---|
69 | self.q = np.empty((self.nq, 2), dtype=dtype) |
---|
70 | self.q[:, 0] = q_vectors[0] |
---|
71 | self.q[:, 1] = q_vectors[1] |
---|
72 | else: |
---|
73 | self.q = np.empty(self.nq, dtype=dtype) |
---|
74 | self.q[:self.nq] = q_vectors[0] |
---|
75 | |
---|
76 | def release(self): |
---|
77 | """ |
---|
78 | Free resources associated with the model inputs. |
---|
79 | """ |
---|
80 | self.q = None |
---|
81 | |
---|
82 | class PyKernel(Kernel): |
---|
83 | """ |
---|
84 | Callable SAS kernel. |
---|
85 | |
---|
86 | *kernel* is the DllKernel object to call. |
---|
87 | |
---|
88 | *model_info* is the module information |
---|
89 | |
---|
90 | *q_input* is the DllInput q vectors at which the kernel should be |
---|
91 | evaluated. |
---|
92 | |
---|
93 | The resulting call method takes the *pars*, a list of values for |
---|
94 | the fixed parameters to the kernel, and *pd_pars*, a list of (value,weight) |
---|
95 | vectors for the polydisperse parameters. *cutoff* determines the |
---|
96 | integration limits: any points with combined weight less than *cutoff* |
---|
97 | will not be calculated. |
---|
98 | |
---|
99 | Call :meth:`release` when done with the kernel instance. |
---|
100 | """ |
---|
101 | def __init__(self, kernel, model_info, q_input): |
---|
102 | self.dtype = np.dtype('d') |
---|
103 | self.info = model_info |
---|
104 | self.q_input = q_input |
---|
105 | self.res = np.empty(q_input.nq, q_input.dtype) |
---|
106 | self.kernel = kernel |
---|
107 | self.dim = '2d' if q_input.is_2d else '1d' |
---|
108 | |
---|
109 | partable = model_info.parameters |
---|
110 | kernel_parameters = (partable.iqxy_parameters if q_input.is_2d |
---|
111 | else partable.iq_parameters) |
---|
112 | volume_parameters = partable.form_volume_parameters |
---|
113 | |
---|
114 | # Create an array to hold the parameter values. There will be a |
---|
115 | # single array whose values are updated as the calculator goes |
---|
116 | # through the loop. Arguments to the kernel and volume functions |
---|
117 | # will use views into this vector, relying on the fact that a |
---|
118 | # an array of no dimensions acts like a scalar. |
---|
119 | parameter_vector = np.empty(len(partable.call_parameters)-2, 'd') |
---|
120 | |
---|
121 | # Create views into the array to hold the arguments |
---|
122 | offset = 0 |
---|
123 | kernel_args, volume_args = [], [] |
---|
124 | for p in partable.kernel_parameters: |
---|
125 | if p.length == 1: |
---|
126 | # Scalar values are length 1 vectors with no dimensions. |
---|
127 | v = parameter_vector[offset:offset+1].reshape(()) |
---|
128 | else: |
---|
129 | # Vector values are simple views. |
---|
130 | v = parameter_vector[offset:offset+p.length] |
---|
131 | offset += p.length |
---|
132 | if p in kernel_parameters: |
---|
133 | kernel_args.append(v) |
---|
134 | if p in volume_parameters: |
---|
135 | volume_args.append(v) |
---|
136 | |
---|
137 | # Hold on to the parameter vector so we can use it to call kernel later. |
---|
138 | # This may also be required to preserve the views into the vector. |
---|
139 | self._parameter_vector = parameter_vector |
---|
140 | |
---|
141 | # Generate a closure which calls the kernel with the views into the |
---|
142 | # parameter array. |
---|
143 | if q_input.is_2d: |
---|
144 | form = model_info.Iqxy |
---|
145 | qx, qy = q_input.q[:,0], q_input.q[:,1] |
---|
146 | self._form = lambda: form(qx, qy, *kernel_args) |
---|
147 | else: |
---|
148 | form = model_info.Iq |
---|
149 | q = q_input.q |
---|
150 | self._form = lambda: form(q, *kernel_args) |
---|
151 | |
---|
152 | # Generate a closure which calls the form_volume if it exists. |
---|
153 | form_volume = model_info.form_volume |
---|
154 | self._volume = ((lambda: form_volume(*volume_args)) if form_volume |
---|
155 | else (lambda: 1.0)) |
---|
156 | |
---|
157 | def __call__(self, call_details, values, cutoff, magnetic): |
---|
158 | assert isinstance(call_details, details.CallDetails) |
---|
159 | res = _loops(self._parameter_vector, self._form, self._volume, |
---|
160 | self.q_input.nq, call_details, values, cutoff) |
---|
161 | return res |
---|
162 | |
---|
163 | def release(self): |
---|
164 | """ |
---|
165 | Free resources associated with the kernel. |
---|
166 | """ |
---|
167 | self.q_input.release() |
---|
168 | self.q_input = None |
---|
169 | |
---|
170 | def _loops(parameters, form, form_volume, nq, details, |
---|
171 | values, cutoff): |
---|
172 | # type: (np.ndarray, Callable[[], np.ndarray], Callable[[], float], int, details.CallDetails, np.ndarray, np.ndarray, float) -> None |
---|
173 | ################################################################ |
---|
174 | # # |
---|
175 | # !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! # |
---|
176 | # !! !! # |
---|
177 | # !! KEEP THIS CODE CONSISTENT WITH KERNEL_TEMPLATE.C !! # |
---|
178 | # !! !! # |
---|
179 | # !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!! # |
---|
180 | # # |
---|
181 | ################################################################ |
---|
182 | NPARS = len(parameters) |
---|
183 | parameters[:] = values[2:NPARS+2] |
---|
184 | scale, background = values[0], values[1] |
---|
185 | if details.num_active == 0: |
---|
186 | norm = float(form_volume()) |
---|
187 | if norm > 0.0: |
---|
188 | return (scale/norm)*form() + background |
---|
189 | else: |
---|
190 | return np.ones(nq, 'd')*background |
---|
191 | |
---|
192 | pd_value = values[2+NPARS:2+NPARS+details.pd_sum] |
---|
193 | pd_weight = values[2+NPARS+details.pd_sum:] |
---|
194 | |
---|
195 | pd_norm = 0.0 |
---|
196 | spherical_correction = 1.0 |
---|
197 | partial_weight = np.NaN |
---|
198 | weight =np.NaN |
---|
199 | |
---|
200 | p0_par = details.pd_par[0] |
---|
201 | p0_is_theta = (p0_par == details.theta_par) |
---|
202 | p0_length = details.pd_length[0] |
---|
203 | p0_index = p0_length |
---|
204 | p0_offset = details.pd_offset[0] |
---|
205 | |
---|
206 | pd_par = details.pd_par[:details.num_active] |
---|
207 | pd_offset = details.pd_offset[:details.num_active] |
---|
208 | pd_stride = details.pd_stride[:details.num_active] |
---|
209 | pd_length = details.pd_length[:details.num_active] |
---|
210 | |
---|
211 | total = np.zeros(nq, 'd') |
---|
212 | for loop_index in range(details.pd_prod): |
---|
213 | # update polydispersity parameter values |
---|
214 | if p0_index == p0_length: |
---|
215 | pd_index = (loop_index//pd_stride)%pd_length |
---|
216 | parameters[pd_par] = pd_value[pd_offset+pd_index] |
---|
217 | partial_weight = np.prod(pd_weight[pd_offset+pd_index][1:]) |
---|
218 | if details.theta_par >= 0: |
---|
219 | spherical_correction = max(abs(cos(pi/180 * parameters[details.theta_par])), 1e-6) |
---|
220 | p0_index = loop_index%p0_length |
---|
221 | |
---|
222 | weight = partial_weight * pd_weight[p0_offset + p0_index] |
---|
223 | parameters[p0_par] = pd_value[p0_offset + p0_index] |
---|
224 | if p0_is_theta: |
---|
225 | spherical_correction = max(abs(cos(pi/180 * parameters[p0_par])), 1e-6) |
---|
226 | p0_index += 1 |
---|
227 | if weight > cutoff: |
---|
228 | # Call the scattering function |
---|
229 | # Assume that NaNs are only generated if the parameters are bad; |
---|
230 | # exclude all q for that NaN. Even better would be to have an |
---|
231 | # INVALID expression like the C models, but that is too expensive. |
---|
232 | I = form() |
---|
233 | if np.isnan(I).any(): continue |
---|
234 | |
---|
235 | # update value and norm |
---|
236 | weight *= spherical_correction |
---|
237 | total += weight * I |
---|
238 | pd_norm += weight * form_volume() |
---|
239 | |
---|
240 | if pd_norm > 0.0: |
---|
241 | return (scale/pd_norm)*total + background |
---|
242 | else: |
---|
243 | return np.ones(nq, 'd')*background |
---|
244 | |
---|
245 | |
---|
246 | def _create_default_functions(model_info): |
---|
247 | """ |
---|
248 | Autogenerate missing functions, such as Iqxy from Iq. |
---|
249 | |
---|
250 | This only works for Iqxy when Iq is written in python. :func:`make_source` |
---|
251 | performs a similar role for Iq written in C. This also vectorizes |
---|
252 | any functions that are not already marked as vectorized. |
---|
253 | """ |
---|
254 | _create_vector_Iq(model_info) |
---|
255 | _create_vector_Iqxy(model_info) # call create_vector_Iq() first |
---|
256 | |
---|
257 | |
---|
258 | def _create_vector_Iq(model_info): |
---|
259 | """ |
---|
260 | Define Iq as a vector function if it exists. |
---|
261 | """ |
---|
262 | Iq = model_info.Iq |
---|
263 | if callable(Iq) and not getattr(Iq, 'vectorized', False): |
---|
264 | #print("vectorizing Iq") |
---|
265 | def vector_Iq(q, *args): |
---|
266 | """ |
---|
267 | Vectorized 1D kernel. |
---|
268 | """ |
---|
269 | return np.array([Iq(qi, *args) for qi in q]) |
---|
270 | vector_Iq.vectorized = True |
---|
271 | model_info.Iq = vector_Iq |
---|
272 | |
---|
273 | def _create_vector_Iqxy(model_info): |
---|
274 | """ |
---|
275 | Define Iqxy as a vector function if it exists, or default it from Iq(). |
---|
276 | """ |
---|
277 | Iq, Iqxy = model_info.Iq, model_info.Iqxy |
---|
278 | if callable(Iqxy): |
---|
279 | if not getattr(Iqxy, 'vectorized', False): |
---|
280 | #print("vectorizing Iqxy") |
---|
281 | def vector_Iqxy(qx, qy, *args): |
---|
282 | """ |
---|
283 | Vectorized 2D kernel. |
---|
284 | """ |
---|
285 | return np.array([Iqxy(qxi, qyi, *args) for qxi, qyi in zip(qx, qy)]) |
---|
286 | vector_Iqxy.vectorized = True |
---|
287 | model_info.Iqxy = vector_Iqxy |
---|
288 | elif callable(Iq): |
---|
289 | #print("defaulting Iqxy") |
---|
290 | # Iq is vectorized because create_vector_Iq was already called. |
---|
291 | def default_Iqxy(qx, qy, *args): |
---|
292 | """ |
---|
293 | Default 2D kernel. |
---|
294 | """ |
---|
295 | return Iq(np.sqrt(qx**2 + qy**2), *args) |
---|
296 | default_Iqxy.vectorized = True |
---|
297 | model_info.Iqxy = default_Iqxy |
---|
298 | |
---|