1 | """ |
---|
2 | Core model handling routines. |
---|
3 | """ |
---|
4 | __all__ = ["list_models", "load_model_definition", "precompile_dll", |
---|
5 | "load_model", "make_kernel", "call_kernel", "call_ER", "call_VR" ] |
---|
6 | |
---|
7 | from os.path import basename, dirname, join as joinpath |
---|
8 | from glob import glob |
---|
9 | |
---|
10 | import numpy as np |
---|
11 | |
---|
12 | from . import models |
---|
13 | from . import weights |
---|
14 | from . import generate |
---|
15 | |
---|
16 | from . import kernelpy |
---|
17 | from . import kerneldll |
---|
18 | try: |
---|
19 | from . import kernelcl |
---|
20 | HAVE_OPENCL = True |
---|
21 | except: |
---|
22 | HAVE_OPENCL = False |
---|
23 | |
---|
24 | |
---|
25 | def list_models(): |
---|
26 | """ |
---|
27 | Return the list of available models on the model path. |
---|
28 | """ |
---|
29 | root = dirname(__file__) |
---|
30 | files = sorted(glob(joinpath(root, 'models', "[a-zA-Z]*.py"))) |
---|
31 | available_models = [basename(f)[:-3] for f in files] |
---|
32 | return available_models |
---|
33 | |
---|
34 | |
---|
35 | def load_model_definition(model_name): |
---|
36 | """ |
---|
37 | Load a model definition given the model name. |
---|
38 | """ |
---|
39 | __import__('sasmodels.models.'+model_name) |
---|
40 | model_definition = getattr(models, model_name, None) |
---|
41 | return model_definition |
---|
42 | |
---|
43 | |
---|
44 | def precompile_dll(model_name, dtype="double"): |
---|
45 | """ |
---|
46 | Precompile the dll for a model. |
---|
47 | |
---|
48 | Returns the path to the compiled model. |
---|
49 | |
---|
50 | This can be used when build the windows distribution of sasmodels |
---|
51 | (which may be missing the OpenCL driver and the dll compiler), or |
---|
52 | otherwise sharing models with windows users who do not have a compiler. |
---|
53 | |
---|
54 | See :func:`sasmodels.kerneldll.make_dll` for details on controlling the |
---|
55 | dll path and the allowed floating point precision. |
---|
56 | """ |
---|
57 | model_definition = load_model_definition(model_name) |
---|
58 | source, info = generate.make(model_definition) |
---|
59 | return kerneldll.make_dll(source, info, dtype=dtype) |
---|
60 | |
---|
61 | |
---|
62 | def isstr(s): |
---|
63 | try: s + '' |
---|
64 | except: return False |
---|
65 | return True |
---|
66 | |
---|
67 | def load_model(model_definition, dtype="single", platform="ocl"): |
---|
68 | """ |
---|
69 | Prepare the model for the default execution platform. |
---|
70 | |
---|
71 | This will return an OpenCL model, a DLL model or a python model depending |
---|
72 | on the model and the computing platform. |
---|
73 | |
---|
74 | *model_definition* is the python module which defines the model. If the |
---|
75 | model name is given instead, then :func:`load_model_definition` will be |
---|
76 | called with the model name. |
---|
77 | |
---|
78 | *dtype* indicates whether the model should use single or double precision |
---|
79 | for the calculation. Any valid numpy single or double precision identifier |
---|
80 | is valid, such as 'single', 'f', 'f32', or np.float32 for single, or |
---|
81 | 'double', 'd', 'f64' and np.float64 for double. |
---|
82 | |
---|
83 | *platform* should be "dll" to force the dll to be used for C models, |
---|
84 | otherwise it uses the default "ocl". |
---|
85 | """ |
---|
86 | if isstr(model_definition): |
---|
87 | model_definition = load_model_definition(model_definition) |
---|
88 | source, info = generate.make(model_definition) |
---|
89 | if callable(info.get('Iq', None)): |
---|
90 | return kernelpy.PyModel(info) |
---|
91 | |
---|
92 | ## for debugging: |
---|
93 | ## 1. uncomment open().write so that the source will be saved next time |
---|
94 | ## 2. run "python -m sasmodels.direct_model $MODELNAME" to save the source |
---|
95 | ## 3. recomment the open.write() and uncomment open().read() |
---|
96 | ## 4. rerun "python -m sasmodels.direct_model $MODELNAME" |
---|
97 | ## 5. uncomment open().read() so that source will be regenerated from model |
---|
98 | # open(info['name']+'.c','w').write(source) |
---|
99 | # source = open(info['name']+'.cl','r').read() |
---|
100 | |
---|
101 | if (platform=="dll" |
---|
102 | or not HAVE_OPENCL |
---|
103 | or not kernelcl.environment().has_type(dtype)): |
---|
104 | return kerneldll.load_dll(source, info, dtype) |
---|
105 | else: |
---|
106 | return kernelcl.GpuModel(source, info, dtype) |
---|
107 | |
---|
108 | def make_kernel(model, q_vectors): |
---|
109 | """ |
---|
110 | Return a computation kernel from the model definition and the q input. |
---|
111 | """ |
---|
112 | model_input = model.make_input(q_vectors) |
---|
113 | return model(model_input) |
---|
114 | |
---|
115 | def get_weights(info, pars, name): |
---|
116 | """ |
---|
117 | Generate the distribution for parameter *name* given the parameter values |
---|
118 | in *pars*. |
---|
119 | |
---|
120 | Uses "name", "name_pd", "name_pd_type", "name_pd_n", "name_pd_sigma" |
---|
121 | from the *pars* dictionary for parameter value and parameter dispersion. |
---|
122 | """ |
---|
123 | relative = name in info['partype']['pd-rel'] |
---|
124 | limits = info['limits'][name] |
---|
125 | disperser = pars.get(name+'_pd_type', 'gaussian') |
---|
126 | value = pars.get(name, info['defaults'][name]) |
---|
127 | npts = pars.get(name+'_pd_n', 0) |
---|
128 | width = pars.get(name+'_pd', 0.0) |
---|
129 | nsigma = pars.get(name+'_pd_nsigma', 3.0) |
---|
130 | value,weight = weights.get_weights( |
---|
131 | disperser, npts, width, nsigma, value, limits, relative) |
---|
132 | return value, weight / np.sum(weight) |
---|
133 | |
---|
134 | def dispersion_mesh(pars): |
---|
135 | """ |
---|
136 | Create a mesh grid of dispersion parameters and weights. |
---|
137 | |
---|
138 | Returns [p1,p2,...],w where pj is a vector of values for parameter j |
---|
139 | and w is a vector containing the products for weights for each |
---|
140 | parameter set in the vector. |
---|
141 | """ |
---|
142 | value, weight = zip(*pars) |
---|
143 | if len(value) > 1: |
---|
144 | value = [v.flatten() for v in np.meshgrid(*value)] |
---|
145 | weight = np.vstack([v.flatten() for v in np.meshgrid(*weight)]) |
---|
146 | weight = np.prod(weight, axis=0) |
---|
147 | return value, weight |
---|
148 | |
---|
149 | def call_kernel(kernel, pars, cutoff=0): |
---|
150 | """ |
---|
151 | Call *kernel* returned from :func:`make_kernel` with parameters *pars*. |
---|
152 | |
---|
153 | *cutoff* is the limiting value for the product of dispersion weights used |
---|
154 | to perform the multidimensional dispersion calculation more quickly at a |
---|
155 | slight cost to accuracy. The default value of *cutoff=0* integrates over |
---|
156 | the entire dispersion cube. Using *cutoff=1e-5* can be 50% faster, but |
---|
157 | with an error of about 1%, which is usually less than the measurement |
---|
158 | uncertainty. |
---|
159 | """ |
---|
160 | fixed_pars = [pars.get(name, kernel.info['defaults'][name]) |
---|
161 | for name in kernel.fixed_pars] |
---|
162 | pd_pars = [get_weights(kernel.info, pars, name) for name in kernel.pd_pars] |
---|
163 | return kernel(fixed_pars, pd_pars, cutoff=cutoff) |
---|
164 | |
---|
165 | def call_ER(info, pars): |
---|
166 | """ |
---|
167 | Call the model ER function using *pars*. |
---|
168 | |
---|
169 | *info* is either *model.info* if you have a loaded model, or *kernel.info* |
---|
170 | if you have a model kernel prepared for evaluation. |
---|
171 | """ |
---|
172 | ER = info.get('ER', None) |
---|
173 | if ER is None: |
---|
174 | return 1.0 |
---|
175 | else: |
---|
176 | vol_pars = [get_weights(info, pars, name) |
---|
177 | for name in info['partype']['volume']] |
---|
178 | value, weight = dispersion_mesh(vol_pars) |
---|
179 | individual_radii = ER(*value) |
---|
180 | #print(values[0].shape, weights.shape, fv.shape) |
---|
181 | return np.sum(weight*individual_radii) / np.sum(weight) |
---|
182 | |
---|
183 | def call_VR(info, pars): |
---|
184 | """ |
---|
185 | Call the model VR function using *pars*. |
---|
186 | |
---|
187 | *info* is either *model.info* if you have a loaded model, or *kernel.info* |
---|
188 | if you have a model kernel prepared for evaluation. |
---|
189 | """ |
---|
190 | VR = info.get('VR', None) |
---|
191 | if VR is None: |
---|
192 | return 1.0 |
---|
193 | else: |
---|
194 | vol_pars = [get_weights(info, pars, name) |
---|
195 | for name in info['partype']['volume']] |
---|
196 | value, weight = dispersion_mesh(vol_pars) |
---|
197 | whole,part = VR(*value) |
---|
198 | return np.sum(weight*part)/np.sum(weight*whole) |
---|
199 | |
---|