| 1 | #!/usr/bin/env python |
|---|
| 2 | # -*- coding: utf-8 -*- |
|---|
| 3 | |
|---|
| 4 | import sys |
|---|
| 5 | import math |
|---|
| 6 | from os.path import basename, dirname, join as joinpath |
|---|
| 7 | import glob |
|---|
| 8 | |
|---|
| 9 | import numpy as np |
|---|
| 10 | |
|---|
| 11 | from sasmodels.bumps_model import BumpsModel, plot_data, tic |
|---|
| 12 | from sasmodels import core |
|---|
| 13 | from sasmodels import kerneldll |
|---|
| 14 | from sasmodels.convert import revert_model |
|---|
| 15 | kerneldll.ALLOW_SINGLE_PRECISION_DLLS = True |
|---|
| 16 | |
|---|
| 17 | # List of available models |
|---|
| 18 | ROOT = dirname(__file__) |
|---|
| 19 | MODELS = [basename(f)[:-3] |
|---|
| 20 | for f in sorted(glob.glob(joinpath(ROOT,"sasmodels","models","[a-zA-Z]*.py")))] |
|---|
| 21 | |
|---|
| 22 | |
|---|
| 23 | def sasview_model(model_definition, **pars): |
|---|
| 24 | """ |
|---|
| 25 | Load a sasview model given the model name. |
|---|
| 26 | """ |
|---|
| 27 | # convert model parameters from sasmodel form to sasview form |
|---|
| 28 | #print "old",sorted(pars.items()) |
|---|
| 29 | modelname, pars = revert_model(model_definition, pars) |
|---|
| 30 | #print "new",sorted(pars.items()) |
|---|
| 31 | sas = __import__('sas.models.'+modelname) |
|---|
| 32 | ModelClass = getattr(getattr(sas.models,modelname,None),modelname,None) |
|---|
| 33 | if ModelClass is None: |
|---|
| 34 | raise ValueError("could not find model %r in sas.models"%modelname) |
|---|
| 35 | model = ModelClass() |
|---|
| 36 | |
|---|
| 37 | for k,v in pars.items(): |
|---|
| 38 | if k.endswith("_pd"): |
|---|
| 39 | model.dispersion[k[:-3]]['width'] = v |
|---|
| 40 | elif k.endswith("_pd_n"): |
|---|
| 41 | model.dispersion[k[:-5]]['npts'] = v |
|---|
| 42 | elif k.endswith("_pd_nsigma"): |
|---|
| 43 | model.dispersion[k[:-10]]['nsigmas'] = v |
|---|
| 44 | elif k.endswith("_pd_type"): |
|---|
| 45 | model.dispersion[k[:-8]]['type'] = v |
|---|
| 46 | else: |
|---|
| 47 | model.setParam(k, v) |
|---|
| 48 | return model |
|---|
| 49 | |
|---|
| 50 | def randomize(p, v): |
|---|
| 51 | """ |
|---|
| 52 | Randomizing parameter. |
|---|
| 53 | |
|---|
| 54 | Guess the parameter type from name. |
|---|
| 55 | """ |
|---|
| 56 | if any(p.endswith(s) for s in ('_pd_n','_pd_nsigma','_pd_type')): |
|---|
| 57 | return v |
|---|
| 58 | elif any(s in p for s in ('theta','phi','psi')): |
|---|
| 59 | # orientation in [-180,180], orientation pd in [0,45] |
|---|
| 60 | if p.endswith('_pd'): |
|---|
| 61 | return 45*np.random.rand() |
|---|
| 62 | else: |
|---|
| 63 | return 360*np.random.rand() - 180 |
|---|
| 64 | elif 'sld' in p: |
|---|
| 65 | # sld in in [-0.5,10] |
|---|
| 66 | return 10.5*np.random.rand() - 0.5 |
|---|
| 67 | elif p.endswith('_pd'): |
|---|
| 68 | # length pd in [0,1] |
|---|
| 69 | return np.random.rand() |
|---|
| 70 | else: |
|---|
| 71 | # values from 0 to 2*x for all other parameters |
|---|
| 72 | return 2*np.random.rand()*(v if v != 0 else 1) |
|---|
| 73 | |
|---|
| 74 | def randomize_model(name, pars, seed=None): |
|---|
| 75 | if seed is None: |
|---|
| 76 | seed = np.random.randint(1e9) |
|---|
| 77 | np.random.seed(seed) |
|---|
| 78 | # Note: the sort guarantees order of calls to random number generator |
|---|
| 79 | pars = dict((p,randomize(p,v)) for p,v in sorted(pars.items())) |
|---|
| 80 | # The capped cylinder model has a constraint on its parameters |
|---|
| 81 | if name == 'capped_cylinder' and pars['cap_radius'] < pars['radius']: |
|---|
| 82 | pars['radius'],pars['cap_radius'] = pars['cap_radius'],pars['radius'] |
|---|
| 83 | return pars, seed |
|---|
| 84 | |
|---|
| 85 | def parlist(pars): |
|---|
| 86 | return "\n".join("%s: %s"%(p,v) for p,v in sorted(pars.items())) |
|---|
| 87 | |
|---|
| 88 | def suppress_pd(pars): |
|---|
| 89 | """ |
|---|
| 90 | Suppress theta_pd for now until the normalization is resolved. |
|---|
| 91 | |
|---|
| 92 | May also suppress complete polydispersity of the model to test |
|---|
| 93 | models more quickly. |
|---|
| 94 | """ |
|---|
| 95 | for p in pars: |
|---|
| 96 | if p.endswith("_pd"): pars[p] = 0 |
|---|
| 97 | |
|---|
| 98 | def eval_sasview(name, pars, data, Nevals=1): |
|---|
| 99 | model = sasview_model(name, **pars) |
|---|
| 100 | toc = tic() |
|---|
| 101 | for _ in range(Nevals): |
|---|
| 102 | if hasattr(data, 'qx_data'): |
|---|
| 103 | value = model.evalDistribution([data.qx_data, data.qy_data]) |
|---|
| 104 | else: |
|---|
| 105 | value = model.evalDistribution(data.x) |
|---|
| 106 | average_time = toc()*1000./Nevals |
|---|
| 107 | return value, average_time |
|---|
| 108 | |
|---|
| 109 | def eval_opencl(model_definition, pars, data, dtype='single', Nevals=1, cutoff=0): |
|---|
| 110 | try: |
|---|
| 111 | model = core.load_model(model_definition, dtype=dtype, platform="ocl") |
|---|
| 112 | except Exception,exc: |
|---|
| 113 | print exc |
|---|
| 114 | print "... trying again with single precision" |
|---|
| 115 | model = core.load_model(model_definition, dtype='single', platform="ocl") |
|---|
| 116 | problem = BumpsModel(data, model, cutoff=cutoff, **pars) |
|---|
| 117 | toc = tic() |
|---|
| 118 | for _ in range(Nevals): |
|---|
| 119 | #pars['scale'] = np.random.rand() |
|---|
| 120 | problem.update() |
|---|
| 121 | value = problem.theory() |
|---|
| 122 | average_time = toc()*1000./Nevals |
|---|
| 123 | return value, average_time |
|---|
| 124 | |
|---|
| 125 | def eval_ctypes(model_definition, pars, data, dtype='double', Nevals=1, cutoff=0): |
|---|
| 126 | model = core.load_model(model_definition, dtype=dtype, platform="dll") |
|---|
| 127 | problem = BumpsModel(data, model, cutoff=cutoff, **pars) |
|---|
| 128 | toc = tic() |
|---|
| 129 | for _ in range(Nevals): |
|---|
| 130 | problem.update() |
|---|
| 131 | value = problem.theory() |
|---|
| 132 | average_time = toc()*1000./Nevals |
|---|
| 133 | return value, average_time |
|---|
| 134 | |
|---|
| 135 | def make_data(qmax, is2D, Nq=128, view='log'): |
|---|
| 136 | if is2D: |
|---|
| 137 | from sasmodels.bumps_model import empty_data2D, set_beam_stop |
|---|
| 138 | data = empty_data2D(np.linspace(-qmax, qmax, Nq)) |
|---|
| 139 | set_beam_stop(data, 0.004) |
|---|
| 140 | index = ~data.mask |
|---|
| 141 | else: |
|---|
| 142 | from sasmodels.bumps_model import empty_data1D |
|---|
| 143 | if view == 'log': |
|---|
| 144 | qmax = math.log10(qmax) |
|---|
| 145 | q = np.logspace(qmax-3, qmax, Nq) |
|---|
| 146 | else: |
|---|
| 147 | q = np.linspace(0.001*qmax, qmax, Nq) |
|---|
| 148 | data = empty_data1D(q) |
|---|
| 149 | index = slice(None, None) |
|---|
| 150 | return data, index |
|---|
| 151 | |
|---|
| 152 | def compare(name, pars, Ncpu, Nocl, opts, set_pars): |
|---|
| 153 | view = 'linear' if '-linear' in opts else 'log' if '-log' in opts else 'q4' if '-q4' in opts else 'log' |
|---|
| 154 | |
|---|
| 155 | opt_values = dict(split |
|---|
| 156 | for s in opts for split in ((s.split('='),)) |
|---|
| 157 | if len(split) == 2) |
|---|
| 158 | # Sort out data |
|---|
| 159 | qmax = 10.0 if '-exq' in opts else 1.0 if '-highq' in opts else 0.2 if '-midq' in opts else 0.05 |
|---|
| 160 | Nq = int(opt_values.get('-Nq', '128')) |
|---|
| 161 | is2D = not "-1d" in opts |
|---|
| 162 | data, index = make_data(qmax, is2D, Nq, view=view) |
|---|
| 163 | |
|---|
| 164 | |
|---|
| 165 | # modelling accuracy is determined by dtype and cutoff |
|---|
| 166 | dtype = 'double' if '-double' in opts else 'single' |
|---|
| 167 | cutoff = float(opt_values.get('-cutoff','1e-5')) |
|---|
| 168 | |
|---|
| 169 | # randomize parameters |
|---|
| 170 | pars.update(set_pars) |
|---|
| 171 | if '-random' in opts or '-random' in opt_values: |
|---|
| 172 | seed = int(opt_values['-random']) if '-random' in opt_values else None |
|---|
| 173 | pars, seed = randomize_model(name, pars, seed=seed) |
|---|
| 174 | print "Randomize using -random=%i"%seed |
|---|
| 175 | |
|---|
| 176 | # parameter selection |
|---|
| 177 | if '-mono' in opts: |
|---|
| 178 | suppress_pd(pars) |
|---|
| 179 | if '-pars' in opts: |
|---|
| 180 | print "pars",parlist(pars) |
|---|
| 181 | |
|---|
| 182 | model_definition = core.load_model_definition(name) |
|---|
| 183 | # OpenCl calculation |
|---|
| 184 | if Nocl > 0: |
|---|
| 185 | ocl, ocl_time = eval_opencl(model_definition, pars, data, |
|---|
| 186 | dtype=dtype, cutoff=cutoff, Nevals=Nocl) |
|---|
| 187 | print "opencl t=%.1f ms, intensity=%.0f"%(ocl_time, sum(ocl[index])) |
|---|
| 188 | #print max(ocl), min(ocl) |
|---|
| 189 | |
|---|
| 190 | # ctypes/sasview calculation |
|---|
| 191 | if Ncpu > 0 and "-ctypes" in opts: |
|---|
| 192 | cpu, cpu_time = eval_ctypes(model_definition, pars, data, |
|---|
| 193 | dtype=dtype, cutoff=cutoff, Nevals=Ncpu) |
|---|
| 194 | comp = "ctypes" |
|---|
| 195 | print "ctypes t=%.1f ms, intensity=%.0f"%(cpu_time, sum(cpu[index])) |
|---|
| 196 | elif Ncpu > 0: |
|---|
| 197 | cpu, cpu_time = eval_sasview(model_definition, pars, data, Ncpu) |
|---|
| 198 | comp = "sasview" |
|---|
| 199 | print "sasview t=%.1f ms, intensity=%.0f"%(cpu_time, sum(cpu[index])) |
|---|
| 200 | |
|---|
| 201 | # Compare, but only if computing both forms |
|---|
| 202 | if Nocl > 0 and Ncpu > 0: |
|---|
| 203 | #print "speedup %.2g"%(cpu_time/ocl_time) |
|---|
| 204 | #print "max |ocl/cpu|", max(abs(ocl/cpu)), "%.15g"%max(abs(ocl)), "%.15g"%max(abs(cpu)) |
|---|
| 205 | #cpu *= max(ocl/cpu) |
|---|
| 206 | resid, relerr = np.zeros_like(ocl), np.zeros_like(ocl) |
|---|
| 207 | resid[index] = (ocl - cpu)[index] |
|---|
| 208 | relerr[index] = resid[index]/cpu[index] |
|---|
| 209 | #bad = (relerr>1e-4) |
|---|
| 210 | #print relerr[bad],cpu[bad],ocl[bad],data.qx_data[bad],data.qy_data[bad] |
|---|
| 211 | def stats(label,err): |
|---|
| 212 | sorted_err = np.sort(abs(err)) |
|---|
| 213 | p50 = int((len(err)-1)*0.50) |
|---|
| 214 | p98 = int((len(err)-1)*0.98) |
|---|
| 215 | data = [ |
|---|
| 216 | "max:%.3e"%sorted_err[-1], |
|---|
| 217 | "median:%.3e"%sorted_err[p50], |
|---|
| 218 | "98%%:%.3e"%sorted_err[p98], |
|---|
| 219 | "rms:%.3e"%np.sqrt(np.mean(err**2)), |
|---|
| 220 | "zero-offset:%+.3e"%np.mean(err), |
|---|
| 221 | ] |
|---|
| 222 | print label," ".join(data) |
|---|
| 223 | stats("|ocl-%s|"%comp+(" "*(3+len(comp))), resid[index]) |
|---|
| 224 | stats("|(ocl-%s)/%s|"%(comp,comp), relerr[index]) |
|---|
| 225 | |
|---|
| 226 | # Plot if requested |
|---|
| 227 | if '-noplot' in opts: return |
|---|
| 228 | import matplotlib.pyplot as plt |
|---|
| 229 | if Ncpu > 0: |
|---|
| 230 | if Nocl > 0: plt.subplot(131) |
|---|
| 231 | plot_data(data, cpu, view=view) |
|---|
| 232 | plt.title("%s t=%.1f ms"%(comp,cpu_time)) |
|---|
| 233 | cbar_title = "log I" |
|---|
| 234 | if Nocl > 0: |
|---|
| 235 | if Ncpu > 0: plt.subplot(132) |
|---|
| 236 | plot_data(data, ocl, view=view) |
|---|
| 237 | plt.title("opencl t=%.1f ms"%ocl_time) |
|---|
| 238 | cbar_title = "log I" |
|---|
| 239 | if Ncpu > 0 and Nocl > 0: |
|---|
| 240 | plt.subplot(133) |
|---|
| 241 | if '-abs' in opts: |
|---|
| 242 | err,errstr,errview = resid, "abs err", "linear" |
|---|
| 243 | else: |
|---|
| 244 | err,errstr,errview = abs(relerr), "rel err", "log" |
|---|
| 245 | #err,errstr = ocl/cpu,"ratio" |
|---|
| 246 | plot_data(data, err, view=errview) |
|---|
| 247 | plt.title("max %s = %.3g"%(errstr, max(abs(err[index])))) |
|---|
| 248 | cbar_title = errstr if errview=="linear" else "log "+errstr |
|---|
| 249 | if is2D: |
|---|
| 250 | h = plt.colorbar() |
|---|
| 251 | h.ax.set_title(cbar_title) |
|---|
| 252 | |
|---|
| 253 | if Ncpu > 0 and Nocl > 0 and '-hist' in opts: |
|---|
| 254 | plt.figure() |
|---|
| 255 | v = relerr[index] |
|---|
| 256 | v[v==0] = 0.5*np.min(np.abs(v[v!=0])) |
|---|
| 257 | plt.hist(np.log10(np.abs(v)), normed=1, bins=50); |
|---|
| 258 | plt.xlabel('log10(err), err = | F(q) single - F(q) double| / | F(q) double |'); |
|---|
| 259 | plt.ylabel('P(err)') |
|---|
| 260 | plt.title('Comparison of single and double precision models for %s'%name) |
|---|
| 261 | |
|---|
| 262 | plt.show() |
|---|
| 263 | |
|---|
| 264 | # =========================================================================== |
|---|
| 265 | # |
|---|
| 266 | USAGE=""" |
|---|
| 267 | usage: compare.py model [Nopencl] [Nsasview] [options...] [key=val] |
|---|
| 268 | |
|---|
| 269 | Compare the speed and value for a model between the SasView original and the |
|---|
| 270 | OpenCL rewrite. |
|---|
| 271 | |
|---|
| 272 | model is the name of the model to compare (see below). |
|---|
| 273 | Nopencl is the number of times to run the OpenCL model (default=5) |
|---|
| 274 | Nsasview is the number of times to run the Sasview model (default=1) |
|---|
| 275 | |
|---|
| 276 | Options (* for default): |
|---|
| 277 | |
|---|
| 278 | -plot*/-noplot plots or suppress the plot of the model |
|---|
| 279 | -single*/-double uses double precision for comparison |
|---|
| 280 | -lowq*/-midq/-highq/-exq use q values up to 0.05, 0.2, 1.0, 10.0 |
|---|
| 281 | -Nq=128 sets the number of Q points in the data set |
|---|
| 282 | -1d/-2d* computes 1d or 2d data |
|---|
| 283 | -preset*/-random[=seed] preset or random parameters |
|---|
| 284 | -mono/-poly* force monodisperse/polydisperse |
|---|
| 285 | -ctypes/-sasview* whether cpu is tested using sasview or ctypes |
|---|
| 286 | -cutoff=1e-5*/value cutoff for including a point in polydispersity |
|---|
| 287 | -pars/-nopars* prints the parameter set or not |
|---|
| 288 | -abs/-rel* plot relative or absolute error |
|---|
| 289 | -linear/-log/-q4 intensity scaling |
|---|
| 290 | -hist/-nohist* plot histogram of relative error |
|---|
| 291 | |
|---|
| 292 | Key=value pairs allow you to set specific values to any of the model |
|---|
| 293 | parameters. |
|---|
| 294 | |
|---|
| 295 | Available models: |
|---|
| 296 | |
|---|
| 297 | %s |
|---|
| 298 | """ |
|---|
| 299 | |
|---|
| 300 | NAME_OPTIONS = set([ |
|---|
| 301 | 'plot','noplot', |
|---|
| 302 | 'single','double', |
|---|
| 303 | 'lowq','midq','highq','exq', |
|---|
| 304 | '2d','1d', |
|---|
| 305 | 'preset','random', |
|---|
| 306 | 'poly','mono', |
|---|
| 307 | 'sasview','ctypes', |
|---|
| 308 | 'nopars','pars', |
|---|
| 309 | 'rel','abs', |
|---|
| 310 | 'linear', 'log', 'q4', |
|---|
| 311 | 'hist','nohist', |
|---|
| 312 | ]) |
|---|
| 313 | VALUE_OPTIONS = [ |
|---|
| 314 | # Note: random is both a name option and a value option |
|---|
| 315 | 'cutoff', 'random', 'Nq', |
|---|
| 316 | ] |
|---|
| 317 | |
|---|
| 318 | def get_demo_pars(name): |
|---|
| 319 | import sasmodels.models |
|---|
| 320 | __import__('sasmodels.models.'+name) |
|---|
| 321 | model = getattr(sasmodels.models, name) |
|---|
| 322 | pars = getattr(model, 'demo', None) |
|---|
| 323 | if pars is None: pars = dict((p[0],p[2]) for p in model.parameters) |
|---|
| 324 | return pars |
|---|
| 325 | |
|---|
| 326 | def main(): |
|---|
| 327 | opts = [arg for arg in sys.argv[1:] if arg.startswith('-')] |
|---|
| 328 | args = [arg for arg in sys.argv[1:] if not arg.startswith('-')] |
|---|
| 329 | models = "\n ".join("%-15s"%v for v in MODELS) |
|---|
| 330 | if len(args) == 0: |
|---|
| 331 | print(USAGE%models) |
|---|
| 332 | sys.exit(1) |
|---|
| 333 | if args[0] not in MODELS: |
|---|
| 334 | print "Model %r not available. Use one of:\n %s"%(args[0],models) |
|---|
| 335 | sys.exit(1) |
|---|
| 336 | |
|---|
| 337 | invalid = [o[1:] for o in opts |
|---|
| 338 | if o[1:] not in NAME_OPTIONS |
|---|
| 339 | and not any(o.startswith('-%s='%t) for t in VALUE_OPTIONS)] |
|---|
| 340 | if invalid: |
|---|
| 341 | print "Invalid options: %s"%(", ".join(invalid)) |
|---|
| 342 | sys.exit(1) |
|---|
| 343 | |
|---|
| 344 | # Get demo parameters from model definition, or use default parameters |
|---|
| 345 | # if model does not define demo parameters |
|---|
| 346 | name = args[0] |
|---|
| 347 | pars = get_demo_pars(name) |
|---|
| 348 | |
|---|
| 349 | Nopencl = int(args[1]) if len(args) > 1 else 5 |
|---|
| 350 | Nsasview = int(args[2]) if len(args) > 2 else 1 |
|---|
| 351 | |
|---|
| 352 | # Fill in default polydispersity parameters |
|---|
| 353 | pds = set(p.split('_pd')[0] for p in pars if p.endswith('_pd')) |
|---|
| 354 | for p in pds: |
|---|
| 355 | if p+"_pd_nsigma" not in pars: pars[p+"_pd_nsigma"] = 3 |
|---|
| 356 | if p+"_pd_type" not in pars: pars[p+"_pd_type"] = "gaussian" |
|---|
| 357 | |
|---|
| 358 | # Fill in parameters given on the command line |
|---|
| 359 | set_pars = {} |
|---|
| 360 | for arg in args[3:]: |
|---|
| 361 | k,v = arg.split('=') |
|---|
| 362 | if k not in pars: |
|---|
| 363 | # extract base name without distribution |
|---|
| 364 | s = set(p.split('_pd')[0] for p in pars) |
|---|
| 365 | print "%r invalid; parameters are: %s"%(k,", ".join(sorted(s))) |
|---|
| 366 | sys.exit(1) |
|---|
| 367 | set_pars[k] = float(v) if not v.endswith('type') else v |
|---|
| 368 | |
|---|
| 369 | compare(name, pars, Nsasview, Nopencl, opts, set_pars) |
|---|
| 370 | |
|---|
| 371 | if __name__ == "__main__": |
|---|
| 372 | main() |
|---|
