def refresh(self, ind_val=None):
# set attributes
self.ind_list = self.Data.getList(self.ind_attr, False)
self.dep_list = self.Data.getList(self.dep_attr, True)
self.dep = {}
self.ind = list(set(self.ind_list))
self.listToContours = [self.ind.index(i) for i in self.ind_list]
self.DU = DistanceUncertainty(self)
self.indQuantities = [self.ind_list.count(i) for i in self.ind]
# Variables
self.std = np.std(self.ind_list)
self.weightFactor = 1.0 / (2 * ((float(self.weight_std_ratio) * self.std) ** 2))
# The points ordered by their dependent attribute
orderedPoints = zip(self.dep_list, self.ind_list)
orderedPoints.sort()
self.sorted_dep_list, self.sorted_ind_list = zip(*orderedPoints)
if not ind_val is None:
new_indices = [i for i in range(len(orderedPoints)) if self.sorted_ind_list[i] == ind_val]
return new_indices
class ObservedDistribution:
def __init__(self, parser, ind_attr, contours , dep_attr, weight_std_ratio=None, retrain=False, prefix=None, save=True):
#If a None path is provided, make one based on the ind_attr
if prefix is None:
prefix = "ods/"+sanitise(ind_attr)+"/"
if not os.path.exists(prefix):
os.makedirs(prefix)
# If no weight is provided then expect to find one in a filename
if weight_std_ratio is None:
pattern = re.compile(sanitise(ind_attr)+'_[0-9]_'+sanitise(dep_attr))
for f in os.listdir(prefix):
fn,ext = os.path.splitext(f)
if ext == ".od" and pattern.match(sanitise(fn)) is not None:
weight_std_ratio = float(fn.rsplit("_")[-1])
print "Found",f,"and setting weight ratio to",weight_std_ratio
break
if weight_std_ratio is None:
print "No suitable weight factor found in provided OD files, using 0.15."
weight_std_ratio = 0.15
self.weight_std_ratio = weight_std_ratio
self.contours = contours
self.prefix = prefix
self.filename = getFileName(ind_attr, contours, dep_attr, weight_std_ratio)
self.path = os.path.join(prefix,self.filename)
# if this od has already been computed read in the file it was saved to and copy the attributes from that version
if os.path.isfile(self.path) and not retrain:
od = readObject(self.path)
attributes = inspect.getmembers(od)
for a in attributes:
if not type(a[1]) is types.MethodType:
setattr(self, a[0], a[1])
print "read in",self.path
# otherwise build the od as normal
else:
self.retrain(parser, ind_attr, dep_attr, save=save)
def __repr__(self):
return self.filename
def refresh(self, ind_val=None):
# set attributes
self.ind_list = self.Data.getList(self.ind_attr, False)
self.dep_list = self.Data.getList(self.dep_attr, True)
self.dep = {}
self.ind = list(set(self.ind_list))
self.listToContours = [self.ind.index(i) for i in self.ind_list]
self.DU = DistanceUncertainty(self)
self.indQuantities = [self.ind_list.count(i) for i in self.ind]
# Variables
self.std = np.std(self.ind_list)
self.weightFactor = 1.0 / (2 * ((float(self.weight_std_ratio) * self.std) ** 2))
# The points ordered by their dependent attribute
orderedPoints = zip(self.dep_list, self.ind_list)
orderedPoints.sort()
self.sorted_dep_list, self.sorted_ind_list = zip(*orderedPoints)
if not ind_val is None:
new_indices = [i for i in range(len(orderedPoints)) if self.sorted_ind_list[i] == ind_val]
return new_indices
def retrain(self, parser, ind_attr, dep_attr, save=True):
self.bins = [0.5]
# Expand the bins to the number of contours selected (zero contours = just predict the median)
for i in xrange(self.contours):
self.bins = np.concatenate([[self.bins[0]/2],self.bins,[1-self.bins[0]/2]])
# convert self.bins back into a list
if not type(self.bins) == list:
self.bins = self.bins.tolist()
self.bins.sort()
# set parser
self.Data = parser
# set attributes
self.ind_attr = ind_attr
self.dep_attr = dep_attr
self.refresh()
# Parallel computation:
results, self.weights = zip(*Parallel(n_jobs=-1)(delayed(findBins)(value, self.sorted_ind_list, self.sorted_dep_list, weightFunction, self.weightFactor, self.bins, findResults) for value in self.ind))
self.finishTraining(results)
if save and not os.path.isfile(self.path):
self.saveObject(self.path)
def finishTraining(self, results):
self.weights = list(self.weights)
# Get Training Dots
# Make a list of the input values at each bin boundary
for b in self.bins:
self.dep[b] = []
for r in results:
for pair in r:
self.dep[pair[0]].append(pair[1])
# set self.ind and self.dep (these are what ED will train on)
self.ind = np.array(self.ind).T
for b in self.bins:
self.dep[b] = np.array(self.dep[b])
#This section avoids a crash that occurs when all of a self.ind[b] have the same value.
if np.allclose(self.dep[b],self.dep[b][0]):
self.dep[b][0] *= 1.01
print "Ran into a Y-axis contour (",b,") that has no variance in any of the point weights."
print "This probably means the weighting factors are inappropriately large."
print "Avoiding a crash in the SVR by falsely editing the first datapoint by 1% and then proceeding."
# function to find the weights for each point around a single point on the independent axis
def weightFunction(self, value, weightFactor=None):
wf = self.weightFactor
if weightFactor is not None:
wf = 1.0 / (2 * ((weightFactor * self.std) ** 2))
return np.exp(wf*(-(abs(np.array(self.ind_list)-value) ** 2)))
# getter functions
def distanceUncertainty(self, values):
return self.DU.distanceUncertainty(values)
def indAttrName(self, san=False):
if san:
return sanitise(self.ind_attr)
return self.ind_attr
def indAttr(self):
return self.ind
def indAttrList(self):
return self.ind_list
def depAttrName(self, san=False):
if san:
return sanitise(self.dep_attr)
return self.dep_attr
def observedContours(self):
return self.dep
def scaledDepAttr(self):
return self.Data.getList(self.dep_attr, True)
def unscaledDepAttr(self):
return self.Data.getList(self.dep_attr, False)
def limits(self):
return [[min(self.ind),max(self.ind)],[min(self.unscaledDepAttr()),max(self.unscaledDepAttr())]]
# scale and unscale points trained on the scaled dots
def unscalePoints(self, vals):
return vals*np.array(self.Data.getScale(self.dep_attr))+np.array(self.Data.getTranslate(self.dep_attr))
def scalePoints(self,vals):
return vals/np.array(self.Data.getScale(self.dep_attr))-np.array(self.Data.getTranslate(self.dep_attr))
def plotArtefacts(self,stroke=None,fill='black',plot=None,alpha=1):
if plot is None:
plot = pl.figure().add_subplot(1,1,1)
plot.scatter(self.ind_list, self.unscalePoints(self.dep_list), edgecolor=stroke,facecolor=fill,s=2,lw=0.25,alpha=alpha)
return plot
def plotArtefact(self,x=None,y=None,plot=None,alpha=1,ED=None):
if plot is None:
plot = pl.figure().add_subplot(1,1,1)
if x is None:
minx = min(self.ind_list)
maxx = max(self.ind_list)
x = minx+np.random.random()*np.ptp([minx, maxx])
if y is None:
y_pred = ED.getExpectationsAt(x,False)
error_scale = np.mean(y_pred[self.bins[-1]]-y_pred[self.bins[0]])
y = ED.getExpectationsAt(x,False,medianOnly=True)+((np.random.random()*error_scale)-(0.5*error_scale))
text_pos = pl.ylim()[0]+np.ptp(pl.ylim())*0.025
surprise,raw_surprise = ED.surpriseCalc(x,y,None,False)
plot.axvline(x)
plot.scatter(x,y,s=500,c='r',marker='*')
text = "".join([" Hypothetical phone: S=",str(round(surprise,3)),' (raw: ',str(round(abs(raw_surprise),3)),')'])
plot.annotate(text,[x,text_pos],color='b')
return plot
# plot the results
def plotObservedContours(self, title="", plot=None, alpha=1):
if plot is None:
plot = pl.figure().add_subplot(1,1,1)
# median dot size (for Kaz)
med_S = 2
# regular dot size
reg_S = .5
# data size
data_S = 5
centralIndex = self.bins.index(0.5)
for i,b in enumerate(self.bins):
dist_from_med = float(abs(i-centralIndex))/(len(self.bins) *.5)
color = (1-dist_from_med, dist_from_med, 0)
if b == 0.5:
S = med_S
else:
S = reg_S
zipped = zip(self.ind, self.unscalePoints(self.dep[b]))
zipped.sort()
x, y = zip(*zipped)
list(x)
list(y)
#if b == 0.5:
# print x
# print y
# sys.exit()
plot.plot(x, y, color=color, lw=S, alpha=alpha)
if self.ind_attr is None:
plot.set_xlabel('$Year$')
else:
plot.set_xlabel(self.ind_attr)
plot.set_ylabel(self.dep_attr)
if len(title) > 0:
plot.set_title(title)
return plot
def plotWeights(self, value, plot=None, alpha=1):
if plot is None:
plot = pl.figure().add_subplot(1,1,1)
weights = self.weightFunction(value)
weights *= 10000.0/sum(weights)
weights = np.sqrt(weights)
plot.scatter(self.ind_list, self.unscalePoints(self.dep_list), color='r', s=weights, alpha=alpha)
return plot
#Show the current plot(s).
def show(self):
pl.show()
#Save the current plot to a given filename.
def saveFig(self,filename):
#if os.path.isfile(self.prefix+'/'+filename):
# version = 1
# while os.path.isfile(self.prefix+'/'+str(version)+'_'+filename):
# version += 1
# filename = str(version)+'_'+filename
pl.savefig(filename)
print 'Saved',filename
def saveObject(self, filename):
with open(filename, 'wb') as output:
pickle.dump(self, output, pickle.HIGHEST_PROTOCOL)
print "Saved,",filename
