| 1 |
""" |
| 2 |
Probability density functions with support for shared and computed parameters. |
| 3 |
|
| 4 |
This module extends the functionality of L{csb.statistics.pdf} with a specialized |
| 5 |
and more sophisticated L{AbstractDensity} -- the L{ParameterizedDensity}, which |
| 6 |
works with L{AbstractParameter} objects rather than simple floats. |
| 7 |
|
| 8 |
Each L{AbstractParameter} holds two properties - L{AbstractParameter.name} and |
| 9 |
L{AbstractParameter.value}: |
| 10 |
|
| 11 |
>>> class Sigma(AbstractParameter): |
| 12 |
>>> def _validate(self, value): |
| 13 |
>>> return float(value) |
| 14 |
>>> def _compute(self, base_value): |
| 15 |
>>> return 1.0 / base_value ** 0.5 |
| 16 |
>>> |
| 17 |
>>> sigma = Sigma(3) |
| 18 |
>>> sigma.name, sigma.value |
| 19 |
sigma, 3 |
| 20 |
|
| 21 |
L{AbstractParameter}s holding a single float value are indistinguishable from |
| 22 |
the simple float parameters used in L{csb.statistics.pdf.BaseDensity}. |
| 23 |
However, each L{AbstractParameter} supports the concept of binding: |
| 24 |
|
| 25 |
>>> sigma.is_virtual |
| 26 |
False |
| 27 |
>>> precision = Precision(1) |
| 28 |
>>> sigma.bind_to(precision) |
| 29 |
>>> sigma.is_virtual |
| 30 |
True |
| 31 |
>>> precision.set(2) # triggers an implicit, lazy update in sigma |
| 32 |
>>> sigma.set(1) |
| 33 |
ParameterizationError: Virtual parameters can't be updated explicitly |
| 34 |
|
| 35 |
The instance of Sigma is now a virtual parameter which receives automatic updates |
| 36 |
from another base parameter using a predefined rule (L{AbstractParameter._compute}). |
| 37 |
This is a lazy, on demand process. As soon as Sigma's computed value is |
| 38 |
requested (via C{s.value}), Sigma will query the parameter it depends on |
| 39 |
(Precision), which in turn will get recomputed first based on its own base, etc. |
| 40 |
Thus, the L{AbstractParameter} model supports a parameter dependency chain with |
| 41 |
linear or tree-like topologies:: |
| 42 |
|
| 43 |
sigma -- y |
| 44 |
/ |
| 45 |
precision -- sigma2 -- x |
| 46 |
|
| 47 |
In this graph precision is a base (non-virtual) parameter and sigma, sigma2, x, and y |
| 48 |
are all virtual (computed). Binding precision to another parameter will immediately |
| 49 |
turn it into a virtual one. However, cycles are not allowed (e.g. precision can't |
| 50 |
be bound to sigma2 or x) and each virtual parameter must have exactly one base. |
| 51 |
|
| 52 |
Computed parameters can then be used to implement custom PDFs with dependent |
| 53 |
parameters within one PDF or spanning multiple PDFs. |
| 54 |
""" |
| 55 |
|
| 56 |
import csb.statistics.pdf as pdf |
| 57 |
|
| 58 |
from abc import abstractmethod |
| 59 |
|
| 60 |
|
| 61 |
class ParameterizationError(ValueError): |
| 62 |
pass |
| 63 |
|
| 64 |
class ParameterValueError(pdf.ParameterValueError): |
| 65 |
pass |
| 66 |
|
| 67 |
|
| 68 |
class ParameterizedDensity(pdf.AbstractDensity): |
| 69 |
""" |
| 70 |
Base abstract class for all PDFs, which operate on simple or computed |
| 71 |
(chained) parameters. Parameters of type different from L{AbstractParameter} |
| 72 |
will trigger TypeError-s. |
| 73 |
""" |
| 74 |
|
| 75 |
def _set(self, param, value): |
| 76 |
|
| 77 |
if not isinstance(value, AbstractParameter): |
| 78 |
raise TypeError(value) |
| 79 |
|
| 80 |
super(ParameterizedDensity, self)._set(param, value) |
| 81 |
|
| 82 |
|
| 83 |
class AbstractParameter(object): |
| 84 |
""" |
| 85 |
Abstract parameterization, which can exist independently or be coupled |
| 86 |
to other parameters upon request. Virtual/coupled/derived parameters cannot |
| 87 |
be overwritten explicitly, but their values will get recomputed once their |
| 88 |
corresponding base parameters get updated. This is a lazy process - parameter |
| 89 |
recalculation happens only when an out of date parameter is requested. This |
| 90 |
triggers a real-time cascaded update which affects all parameters from the |
| 91 |
nearest consistent base down to the current inconsistent node. |
| 92 |
|
| 93 |
Implementing subclasses must override L{AbstractParameter._validate} |
| 94 |
and virtual parameters should additionally override L{AbstractParameter._compute}. |
| 95 |
|
| 96 |
@param value: initial value (defaults to None / AbstractParameter.NULL) |
| 97 |
@type value: object |
| 98 |
@param name: name of parameter (this is the name of the class by default) |
| 99 |
@type name: str |
| 100 |
@param base: optional base parameter to compute this instance from |
| 101 |
@type base: L{AbstractParameter} |
| 102 |
""" |
| 103 |
|
| 104 |
NULL = None |
| 105 |
|
| 106 |
def __init__(self, value=NULL, name=None, base=None): |
| 107 |
|
| 108 |
self._derivatives = set() |
| 109 |
self._base = None |
| 110 |
self._consistent = True |
| 111 |
|
| 112 |
if name is None: |
| 113 |
name = self.__class__.__name__.lower() |
| 114 |
|
| 115 |
self._name = str(name) |
| 116 |
self._value = AbstractParameter.NULL |
| 117 |
|
| 118 |
self._update(value) |
| 119 |
|
| 120 |
if base is not None: |
| 121 |
self.bind_to(base) |
| 122 |
|
| 123 |
@property |
| 124 |
def name(self): |
| 125 |
""" |
| 126 |
Parameter name |
| 127 |
""" |
| 128 |
return self._name |
| 129 |
|
| 130 |
@property |
| 131 |
def value(self): |
| 132 |
""" |
| 133 |
Parameter value (guaranteed to be up to date) |
| 134 |
""" |
| 135 |
self._ensure_consistency() |
| 136 |
return self._value |
| 137 |
|
| 138 |
@property |
| 139 |
def is_virtual(self): |
| 140 |
""" |
| 141 |
True if this parameter is virtual (computed) |
| 142 |
""" |
| 143 |
return self._base is not None |
| 144 |
|
| 145 |
def set(self, value): |
| 146 |
""" |
| 147 |
Update the value of this parameter. This is not possible for |
| 148 |
virtual parameters. |
| 149 |
|
| 150 |
@param value: new value |
| 151 |
@type value: object |
| 152 |
|
| 153 |
@raise ParameterizationError: if this is a virtual parameter |
| 154 |
@raise ParameterValueError: on invalid value |
| 155 |
""" |
| 156 |
if self.is_virtual: |
| 157 |
raise ParameterizationError( |
| 158 |
"Virtual parameters can't be updated explicitly") |
| 159 |
|
| 160 |
self._update(value) |
| 161 |
|
| 162 |
self._invalidate() |
| 163 |
self._consistent = True |
| 164 |
|
| 165 |
def bind_to(self, parameter): |
| 166 |
""" |
| 167 |
Bind the current parameter to a base parameter. This converts |
| 168 |
the current parameter to a virtual one, whose value will get |
| 169 |
implicitly updated to be consistent with its base. |
| 170 |
|
| 171 |
Note that virtual parameters must have exactly one base; computing a |
| 172 |
parameter from multiple bases is not allowed. Cycles are also not |
| 173 |
allowed; the topology must always stay a tree with a non-virtual |
| 174 |
parameter at the root. |
| 175 |
|
| 176 |
@param parameter: base parameter to compute this instance from |
| 177 |
@param parameter: L{AbstractParameter} |
| 178 |
|
| 179 |
@raise ParameterizationError: if this parameter is already virtual |
| 180 |
@raise ParameterizationError: on attempt to create a circular dependency |
| 181 |
|
| 182 |
""" |
| 183 |
|
| 184 |
if not isinstance(parameter, AbstractParameter): |
| 185 |
raise TypeError(parameter) |
| 186 |
|
| 187 |
if parameter.find_base_parameter() is self: |
| 188 |
raise ParameterizationError("Circular dependency detected") |
| 189 |
|
| 190 |
if self.is_virtual: |
| 191 |
msg = "Parameter {0.name} is already bound to {1.name}" |
| 192 |
raise ParameterizationError(msg.format(self, self._base)) |
| 193 |
|
| 194 |
self._set_base(parameter) |
| 195 |
self._base._add_derived(self) |
| 196 |
|
| 197 |
self._invalidate() |
| 198 |
|
| 199 |
def _set_base(self, parameter): |
| 200 |
self._base = parameter |
| 201 |
|
| 202 |
def _add_derived(self, parameter): |
| 203 |
self._derivatives.add(parameter) |
| 204 |
|
| 205 |
def _invalidate(self): |
| 206 |
""" |
| 207 |
Mark self and its virtual children as inconsistent |
| 208 |
""" |
| 209 |
for p in self._derivatives: |
| 210 |
p._invalidate() |
| 211 |
|
| 212 |
self._consistent = False |
| 213 |
|
| 214 |
def _update(self, value): |
| 215 |
""" |
| 216 |
Overwrite the current value of the parameter. This triggers |
| 217 |
an abstract (custom) validation hook, but has no side effects |
| 218 |
(i.e. it doesn't propagate!) |
| 219 |
""" |
| 220 |
sanitized = self._validate(value) |
| 221 |
self._value = sanitized |
| 222 |
|
| 223 |
@abstractmethod |
| 224 |
def _validate(self, value): |
| 225 |
""" |
| 226 |
Validate and sanitize the specified value before assignment. |
| 227 |
@return: sanitized value |
| 228 |
|
| 229 |
@raise ParameterValueError: on invalid value |
| 230 |
""" |
| 231 |
return value |
| 232 |
|
| 233 |
def _compute(self, base_value): |
| 234 |
""" |
| 235 |
Compute a new value for the current parameter given the value |
| 236 |
of a base parameter (assuming self.is_virtual). By default this returns |
| 237 |
the value of the base parameter (i.e. self just inherits the value |
| 238 |
of its base untouched). |
| 239 |
""" |
| 240 |
return base_value |
| 241 |
|
| 242 |
def _ensure_consistency(self): |
| 243 |
""" |
| 244 |
Make sure that the current value is up to date. If it isn't, |
| 245 |
trigger a real-time cascaded update following the path from the |
| 246 |
nearest consistent base down to self. Also mark all nodes consistent |
| 247 |
in the course of doing this update. |
| 248 |
""" |
| 249 |
if not self._consistent: |
| 250 |
path = self._nearest_consistent_base() |
| 251 |
|
| 252 |
for parameter in reversed(path): |
| 253 |
parameter._recompute(consistent=True) |
| 254 |
|
| 255 |
def _recompute(self, consistent=True): |
| 256 |
""" |
| 257 |
If self is virtual, force the current parameter to recompute itself from |
| 258 |
its immediate base. This operation has no side effects and does not |
| 259 |
propagate. |
| 260 |
""" |
| 261 |
if self.is_virtual: |
| 262 |
value = self._compute(self._base._value) |
| 263 |
self._update(value) |
| 264 |
|
| 265 |
if consistent: |
| 266 |
self._consistent = True |
| 267 |
|
| 268 |
def _recompute_derivatives(self): |
| 269 |
""" |
| 270 |
Recompute all derived parameters starting from self and mark |
| 271 |
them consistent. |
| 272 |
""" |
| 273 |
self._recompute(consistent=True) |
| 274 |
|
| 275 |
for p in self._derivatives: |
| 276 |
p._recompute_derivatives() |
| 277 |
|
| 278 |
def _nearest_consistent_base(self): |
| 279 |
""" |
| 280 |
Compute and return the path from self to the nearest consistent |
| 281 |
base parameter. |
| 282 |
|
| 283 |
@return: path, leaf-to-root |
| 284 |
@rtype: list of L{AbstractParameter} |
| 285 |
""" |
| 286 |
root = self |
| 287 |
path = [self] |
| 288 |
|
| 289 |
while not root._consistent: |
| 290 |
root = root._base |
| 291 |
path.append(root) |
| 292 |
|
| 293 |
return path |
| 294 |
|
| 295 |
def find_base_parameter(self): |
| 296 |
""" |
| 297 |
Find and return the non-virtual base parameter that is the root |
| 298 |
of the current hierarchy. If self is not virtual, return self. |
| 299 |
|
| 300 |
@return: base parameter |
| 301 |
@rtype: L{AbstractParameter} |
| 302 |
""" |
| 303 |
root = self |
| 304 |
|
| 305 |
while root.is_virtual: |
| 306 |
root = root._base |
| 307 |
|
| 308 |
return root |
| 309 |
|
| 310 |
|
| 311 |
class Parameter(AbstractParameter): |
| 312 |
""" |
| 313 |
Default parameter implementation which accepts float values only. |
| 314 |
""" |
| 315 |
|
| 316 |
def __init__(self, value=0.0, name=None, base=None): |
| 317 |
super(Parameter, self).__init__(value, name, base) |
| 318 |
|
| 319 |
def _validate(self, value): |
| 320 |
|
| 321 |
try: |
| 322 |
return float(value) |
| 323 |
except (ValueError, TypeError): |
| 324 |
raise ParameterValueError(self.name, value) |
| 325 |
|
| 326 |
|
| 327 |
class NonVirtualParameter(Parameter): |
| 328 |
""" |
| 329 |
A float L{Parameter} that is explicitly non-computed and cannot be |
| 330 |
bound to another L{Parameter}. |
| 331 |
""" |
| 332 |
|
| 333 |
def bind_to(self, parameter): |
| 334 |
raise ParameterizationError( |
| 335 |
"This parameter is explicitly non-computed") |
| 336 |
|
| 337 |
@property |
| 338 |
def is_virtual(self): |
| 339 |
return False |
| 340 |
|
| 341 |
|
| 342 |
|
| 343 |
|
| 344 |
|
| 345 |
|
| 346 |
|