1 | """ |
---|
2 | Core model handling routines. |
---|
3 | """ |
---|
4 | |
---|
5 | from os.path import basename, dirname, join as joinpath, splitext |
---|
6 | from glob import glob |
---|
7 | import imp |
---|
8 | |
---|
9 | import numpy as np |
---|
10 | |
---|
11 | from . import models |
---|
12 | from . import weights |
---|
13 | from . import generate |
---|
14 | # TODO: remove circular references between product and core |
---|
15 | # product uses call_ER/call_VR, core uses make_product_info/ProductModel |
---|
16 | #from . import product |
---|
17 | from . import mixture |
---|
18 | from . import kernelpy |
---|
19 | from . import kerneldll |
---|
20 | try: |
---|
21 | from . import kernelcl |
---|
22 | HAVE_OPENCL = True |
---|
23 | except: |
---|
24 | HAVE_OPENCL = False |
---|
25 | |
---|
26 | __all__ = [ |
---|
27 | "list_models", "load_model_info", "precompile_dll", |
---|
28 | "build_model", "make_kernel", "call_kernel", "call_ER_VR", |
---|
29 | ] |
---|
30 | |
---|
31 | def list_models(): |
---|
32 | """ |
---|
33 | Return the list of available models on the model path. |
---|
34 | """ |
---|
35 | root = dirname(__file__) |
---|
36 | files = sorted(glob(joinpath(root, 'models', "[a-zA-Z]*.py"))) |
---|
37 | available_models = [basename(f)[:-3] for f in files] |
---|
38 | return available_models |
---|
39 | |
---|
40 | def isstr(s): |
---|
41 | """ |
---|
42 | Return True if *s* is a string-like object. |
---|
43 | """ |
---|
44 | try: s + '' |
---|
45 | except: return False |
---|
46 | return True |
---|
47 | |
---|
48 | def load_model(model_name, **kw): |
---|
49 | """ |
---|
50 | Load model info and build model. |
---|
51 | """ |
---|
52 | return build_model(load_model_info(model_name), **kw) |
---|
53 | |
---|
54 | def load_model_info_from_path(path): |
---|
55 | # Pull off the last .ext if it exists; there may be others |
---|
56 | name = basename(splitext(path)[0]) |
---|
57 | |
---|
58 | # Not cleaning name since don't need to be able to reload this |
---|
59 | # model later |
---|
60 | # Should probably turf the model from sys.modules after we are done... |
---|
61 | |
---|
62 | # Placing the model in the 'sasmodels.custom' name space, even |
---|
63 | # though it doesn't actually exist. imp.load_source doesn't seem |
---|
64 | # to care. |
---|
65 | kernel_module = imp.load_source('sasmodels.custom.'+name, path) |
---|
66 | |
---|
67 | # Now that we have the module, we can load it as usual |
---|
68 | model_info = generate.make_model_info(kernel_module) |
---|
69 | return model_info |
---|
70 | |
---|
71 | def load_model_info(model_name): |
---|
72 | """ |
---|
73 | Load a model definition given the model name. |
---|
74 | |
---|
75 | This returns a handle to the module defining the model. This can be |
---|
76 | used with functions in generate to build the docs or extract model info. |
---|
77 | """ |
---|
78 | parts = model_name.split('+') |
---|
79 | if len(parts) > 1: |
---|
80 | model_info_list = [load_model_info(p) for p in parts] |
---|
81 | return mixture.make_mixture_info(model_info_list) |
---|
82 | |
---|
83 | parts = model_name.split('*') |
---|
84 | if len(parts) > 1: |
---|
85 | from . import product |
---|
86 | # Note: currently have circular reference |
---|
87 | if len(parts) > 2: |
---|
88 | raise ValueError("use P*S to apply structure factor S to model P") |
---|
89 | P_info, Q_info = [load_model_info(p) for p in parts] |
---|
90 | return product.make_product_info(P_info, Q_info) |
---|
91 | |
---|
92 | #import sys; print "\n".join(sys.path) |
---|
93 | __import__('sasmodels.models.'+model_name) |
---|
94 | kernel_module = getattr(models, model_name, None) |
---|
95 | return generate.make_model_info(kernel_module) |
---|
96 | |
---|
97 | |
---|
98 | def build_model(model_info, dtype=None, platform="ocl"): |
---|
99 | """ |
---|
100 | Prepare the model for the default execution platform. |
---|
101 | |
---|
102 | This will return an OpenCL model, a DLL model or a python model depending |
---|
103 | on the model and the computing platform. |
---|
104 | |
---|
105 | *model_info* is the model definition structure returned from |
---|
106 | :func:`load_model_info`. |
---|
107 | |
---|
108 | *dtype* indicates whether the model should use single or double precision |
---|
109 | for the calculation. Any valid numpy single or double precision identifier |
---|
110 | is valid, such as 'single', 'f', 'f32', or np.float32 for single, or |
---|
111 | 'double', 'd', 'f64' and np.float64 for double. If *None*, then use |
---|
112 | 'single' unless the model defines single=False. |
---|
113 | |
---|
114 | *platform* should be "dll" to force the dll to be used for C models, |
---|
115 | otherwise it uses the default "ocl". |
---|
116 | """ |
---|
117 | composition = model_info.get('composition', None) |
---|
118 | if composition is not None: |
---|
119 | composition_type, parts = composition |
---|
120 | models = [build_model(p, dtype=dtype, platform=platform) for p in parts] |
---|
121 | if composition_type == 'mixture': |
---|
122 | return mixture.MixtureModel(model_info, models) |
---|
123 | elif composition_type == 'product': |
---|
124 | from . import product |
---|
125 | P, S = models |
---|
126 | return product.ProductModel(model_info, P, S) |
---|
127 | else: |
---|
128 | raise ValueError('unknown mixture type %s'%composition_type) |
---|
129 | |
---|
130 | ## for debugging: |
---|
131 | ## 1. uncomment open().write so that the source will be saved next time |
---|
132 | ## 2. run "python -m sasmodels.direct_model $MODELNAME" to save the source |
---|
133 | ## 3. recomment the open.write() and uncomment open().read() |
---|
134 | ## 4. rerun "python -m sasmodels.direct_model $MODELNAME" |
---|
135 | ## 5. uncomment open().read() so that source will be regenerated from model |
---|
136 | # open(model_info['name']+'.c','w').write(source) |
---|
137 | # source = open(model_info['name']+'.cl','r').read() |
---|
138 | source = generate.make_source(model_info) |
---|
139 | if dtype is None: |
---|
140 | dtype = 'single' if model_info['single'] else 'double' |
---|
141 | if callable(model_info.get('Iq', None)): |
---|
142 | return kernelpy.PyModel(model_info) |
---|
143 | if (platform == "dll" |
---|
144 | or not HAVE_OPENCL |
---|
145 | or not kernelcl.environment().has_type(dtype)): |
---|
146 | return kerneldll.load_dll(source, model_info, dtype) |
---|
147 | else: |
---|
148 | return kernelcl.GpuModel(source, model_info, dtype) |
---|
149 | |
---|
150 | def precompile_dll(model_name, dtype="double"): |
---|
151 | """ |
---|
152 | Precompile the dll for a model. |
---|
153 | |
---|
154 | Returns the path to the compiled model, or None if the model is a pure |
---|
155 | python model. |
---|
156 | |
---|
157 | This can be used when build the windows distribution of sasmodels |
---|
158 | (which may be missing the OpenCL driver and the dll compiler), or |
---|
159 | otherwise sharing models with windows users who do not have a compiler. |
---|
160 | |
---|
161 | See :func:`sasmodels.kerneldll.make_dll` for details on controlling the |
---|
162 | dll path and the allowed floating point precision. |
---|
163 | """ |
---|
164 | model_info = load_model_info(model_name) |
---|
165 | source = generate.make_source(model_info) |
---|
166 | return kerneldll.make_dll(source, model_info, dtype=dtype) if source else None |
---|
167 | |
---|
168 | |
---|
169 | def make_kernel(model, q_vectors): |
---|
170 | """ |
---|
171 | Return a computation kernel from the model definition and the q input. |
---|
172 | """ |
---|
173 | return model(q_vectors) |
---|
174 | |
---|
175 | def get_weights(parameter, values): |
---|
176 | """ |
---|
177 | Generate the distribution for parameter *name* given the parameter values |
---|
178 | in *pars*. |
---|
179 | |
---|
180 | Uses "name", "name_pd", "name_pd_type", "name_pd_n", "name_pd_sigma" |
---|
181 | from the *pars* dictionary for parameter value and parameter dispersion. |
---|
182 | """ |
---|
183 | value = values.get(parameter.name, parameter.default) |
---|
184 | if parameter.type not in ('volume', 'orientation'): |
---|
185 | return np.array([value]), np.array([1.0]) |
---|
186 | relative = parameter.type == 'volume' |
---|
187 | limits = parameter.limits |
---|
188 | disperser = values.get(parameter.name+'_pd_type', 'gaussian') |
---|
189 | npts = values.get(parameter.name+'_pd_n', 0) |
---|
190 | width = values.get(parameter.name+'_pd', 0.0) |
---|
191 | nsigma = values.get(parameter.name+'_pd_nsigma', 3.0) |
---|
192 | value, weight = weights.get_weights( |
---|
193 | disperser, npts, width, nsigma, value, limits, relative) |
---|
194 | return value, weight / np.sum(weight) |
---|
195 | |
---|
196 | def dispersion_mesh(pars): |
---|
197 | """ |
---|
198 | Create a mesh grid of dispersion parameters and weights. |
---|
199 | |
---|
200 | Returns [p1,p2,...],w where pj is a vector of values for parameter j |
---|
201 | and w is a vector containing the products for weights for each |
---|
202 | parameter set in the vector. |
---|
203 | """ |
---|
204 | value, weight = zip(*pars) |
---|
205 | value = [v.flatten() for v in np.meshgrid(*value)] |
---|
206 | weight = np.vstack([v.flatten() for v in np.meshgrid(*weight)]) |
---|
207 | weight = np.prod(weight, axis=0) |
---|
208 | return value, weight |
---|
209 | |
---|
210 | def call_kernel(kernel, pars, cutoff=0, mono=False): |
---|
211 | """ |
---|
212 | Call *kernel* returned from :func:`make_kernel` with parameters *pars*. |
---|
213 | |
---|
214 | *cutoff* is the limiting value for the product of dispersion weights used |
---|
215 | to perform the multidimensional dispersion calculation more quickly at a |
---|
216 | slight cost to accuracy. The default value of *cutoff=0* integrates over |
---|
217 | the entire dispersion cube. Using *cutoff=1e-5* can be 50% faster, but |
---|
218 | with an error of about 1%, which is usually less than the measurement |
---|
219 | uncertainty. |
---|
220 | |
---|
221 | *mono* is True if polydispersity should be set to none on all parameters. |
---|
222 | """ |
---|
223 | if mono: |
---|
224 | values = [pars.get(p.name, p.default) for p in kernel.info['parameters']] |
---|
225 | weights = [1.0]*len(values) |
---|
226 | else: |
---|
227 | vw_pairs = [get_weights(p, pars) for p in kernel.info['parameters']] |
---|
228 | values, weights = zip(*vw_pairs) |
---|
229 | |
---|
230 | #TODO: This is what we thought to do if max([len(w) for w in weights]) > 1: |
---|
231 | print("from") |
---|
232 | import pprint; pprint.pprint(weights) |
---|
233 | print("to") |
---|
234 | if max([len(w) for w in weights]) > 1: |
---|
235 | details = generate.poly_details(kernel.info, weights) |
---|
236 | else: |
---|
237 | details = kernel.info['mono_details'] |
---|
238 | |
---|
239 | weights, values = [np.hstack(v) for v in (weights, values)] |
---|
240 | |
---|
241 | weights = weights.astype(dtype=kernel.dtype) |
---|
242 | values = values.astype(dtype=kernel.dtype) |
---|
243 | |
---|
244 | return kernel(details, weights, values, cutoff) |
---|
245 | |
---|
246 | def call_ER_VR(model_info, vol_pars): |
---|
247 | """ |
---|
248 | Return effect radius and volume ratio for the model. |
---|
249 | |
---|
250 | *info* is either *kernel.info* for *kernel=make_kernel(model,q)* |
---|
251 | or *model.info*. |
---|
252 | |
---|
253 | *pars* are the parameters as expected by :func:`call_kernel`. |
---|
254 | """ |
---|
255 | ER = model_info.get('ER', None) |
---|
256 | VR = model_info.get('VR', None) |
---|
257 | value, weight = dispersion_mesh(vol_pars) |
---|
258 | |
---|
259 | individual_radii = ER(*value) if ER else 1.0 |
---|
260 | whole, part = VR(*value) if VR else (1.0, 1.0) |
---|
261 | |
---|
262 | effect_radius = np.sum(weight*individual_radii) / np.sum(weight) |
---|
263 | volume_ratio = np.sum(weight*part)/np.sum(weight*whole) |
---|
264 | return effect_radius, volume_ratio |
---|
265 | |
---|
266 | |
---|
267 | def call_ER(model_info, values): |
---|
268 | """ |
---|
269 | Call the model ER function using *values*. *model_info* is either |
---|
270 | *model.info* if you have a loaded model, or *kernel.info* if you |
---|
271 | have a model kernel prepared for evaluation. |
---|
272 | """ |
---|
273 | ER = model_info.get('ER', None) |
---|
274 | if ER is None: |
---|
275 | return 1.0 |
---|
276 | else: |
---|
277 | vol_pars = [get_weights(parameter, values) |
---|
278 | for parameter in model_info['parameters'] |
---|
279 | if parameter.type == 'volume'] |
---|
280 | value, weight = dispersion_mesh(vol_pars) |
---|
281 | individual_radii = ER(*value) |
---|
282 | #print(values[0].shape, weights.shape, fv.shape) |
---|
283 | return np.sum(weight*individual_radii) / np.sum(weight) |
---|
284 | |
---|
285 | def call_VR(model_info, values): |
---|
286 | """ |
---|
287 | Call the model VR function using *pars*. |
---|
288 | *info* is either *model.info* if you have a loaded model, or *kernel.info* |
---|
289 | if you have a model kernel prepared for evaluation. |
---|
290 | """ |
---|
291 | VR = model_info.get('VR', None) |
---|
292 | if VR is None: |
---|
293 | return 1.0 |
---|
294 | else: |
---|
295 | vol_pars = [get_weights(parameter, values) |
---|
296 | for parameter in model_info['parameters'] |
---|
297 | if parameter.type == 'volume'] |
---|
298 | value, weight = dispersion_mesh(vol_pars) |
---|
299 | whole, part = VR(*value) |
---|
300 | return np.sum(weight*part)/np.sum(weight*whole) |
---|
301 | |
---|
302 | # TODO: remove call_ER, call_VR |
---|
303 | |
---|