def fitEvents(events, startEvent, stopEvent, opts):
dt = opts['dt']
origTau = opts['tau']
multiFit = opts['multiFit']
waveform = opts['waveform']
tvals = opts['tvals']
dtype = [(n, events[n].dtype) for n in events.dtype.names]
output = np.empty(len(events), dtype=dtype + [
('fitAmplitude', float),
('fitTime', float),
('fitRiseTau', float),
('fitDecayTau', float),
('fitWidth', float),
('fitError', float),
('fitFractionalError', float)
])
offset = 0 ## not all input events will produce output events; offset keeps track of the difference.
outputState = {
'guesses': [],
'eventData': [],
'indexes': [],
'xVals': [],
'yVals': []
}
#print "=========="
for i in range(startEvent, stopEvent):
start = events[i]['time']
sliceLen = events[i]['len']*dt +100.*dt ## Ca2+ events are much longer than 50ms
if i+1 < len(events):
nextStart = events[i+1]['time']
#nextStart = events[i+1]['index']*dt
#print " picking between:", sliceLen, nextStart, '-', start, '=', nextStart-start
sliceLen = min(sliceLen, nextStart-start)
#print " chose:", sliceLen
guessLen = events[i]['len']*dt
#guessLen = sliceLen
tau = origTau
if tau is not None:
guessLen += tau*2.
#print " picking between:", guessLen*3, sliceLen
#sliceLen = min(guessLen*3., sliceLen)
## Figure out from where to pull waveform data that will be fitted
startIndex = np.argwhere(tvals>=start)[0][0]
stopIndex = startIndex + int(sliceLen/dt)
startIndex -= 10 ## pull baseline data from before the event starts
#print " data to fit: indices:", startIndex, stopIndex, 'dt:', dt, "times:", startIndex*dt, stopIndex*dt
eventData = waveform[startIndex:stopIndex]
times = tvals[startIndex:stopIndex]
if len(times) < 4: ## PSP fit requires at least 4 points; skip this one
offset += 1
continue
## reconvolve this chunk of the signal if it was previously deconvolved
if tau is not None:
eventData = functions.expReconvolve(eventData, tau=tau, dt=dt)
## Make guesses as to the shape of the event
mx = eventData.max()
mn = eventData.min()
guessAmp = (mx-mn)*2 ## fit converges more reliably if we start too large
guessRise = guessLen/4.
guessDecay = guessLen/4.
guessStart = times[10]
guessWidth = guessLen*0.75
guessYOffset = eventData[0]
## fitting to exponential rise * decay
## parameters are [amplitude, x-offset, rise tau, fall tau]
guess = [guessStart, guessYOffset, guessRise, guessDecay, guessAmp, guessWidth]
guessFit = [guessYOffset, guessStart, guessRise, guessDecay, guessAmp, guessWidth]
#guess = [amp, times[0], guessLen/4., guessLen/2.] ## careful!
#bounds = [
#sorted((guessAmp * 0.1, guessAmp)),
#sorted((guessStart-guessRise*2, guessStart+guessRise*2)),
#sorted((dt*0.5, guessDecay)),
#sorted((dt*0.5, guessDecay * 50.))
#]
yVals = eventData.view(np.ndarray)
## Set bounds for parameters -
## exppulse parameter order: yOffset, t0, tau1, tau2, amp, width
#yOffset, t0, tau1, tau2, amp, width
bounds=[(-10, 10), ## no bounds on yOffset
(float(events[i]['time']-10*dt), float(events[i]['time']+5*dt)), ## t0 must be near the startpoint found by eventDetection
(0.010, float(opts['riseTauMax'])), ## riseTau must be greater than 10 ms
(0.010, float(opts['decayTauMax'])), ## ditto for decayTau
(0., float(opts['ampMax'])), ## amp must be greater than 0
(0, float(events[i]['len']*dt*2))] ## width
#print "Bounds", bounds
#print "times", times.min(), times.max()
## Use Paul's fitting algorithm so that we can put bounds/constraints on the fit params
#print "event:", i, 'amp bounds:', bounds[4]
fitter = Fitting()
fitResults = fitter.FitRegion([1], 0, times, yVals, fitPars=guessFit, fitFunc='exppulse', bounds=bounds, method='SLSQP', dataType='xy')
fitParams, xPts, yPts, names = fitResults
#print "fitParams:", fitParams
#print "names", names
#fitResult = functions.fit(functions.expPulse, times, yVals, guess, generateResult=True, resultXVals=times)
#fitParams, val, computed, err = fitResult
#print ' fitParams:', fitParams[0]
yOffset, t0, tau1, tau2, amp, width = fitParams[0]
#print "fitResult", fitResult
#computed = fitResult[-2]
computed = fitter.expPulse(fitParams[0], times)
diff = (yVals - computed)
err = (diff**2).sum()
fracError = diff.std() / computed.std()
output[i-offset] = tuple(events[i]) + (amp, t0, tau1, tau2, width) + (err, fracError)
#print "amp:", amp
#print "output:", output[i-offset]
outputState['guesses'].append(guess)
outputState['eventData'].append(eventData)
outputState['indexes'].append(i)
outputState['xVals'].append(times)
outputState['yVals'].append(computed)
if offset > 0:
output = output[:-offset]
outputState['output'] = output
return outputState
def processData(self, data):
return functions.expReconvolve(data)
def processEventFits(events, startEvent, stopEvent, opts):
## This function does all the processing work for EventFitter.
dt = opts['dt']
origTau = opts['tau']
multiFit = opts['multiFit']
waveform = opts['waveform']
tvals = opts['tvals']
nFields = len(events.dtype.fields)
dtype = [(n, events[n].dtype) for n in events.dtype.names]
output = np.empty(len(events), dtype=dtype + [
('fitAmplitude', float),
('fitTime', float),
('fitRiseTau', float),
('fitDecayTau', float),
('fitTimeToPeak', float),
('fitError', float),
('fitFractionalError', float),
('fitLengthOverDecay', float),
])
offset = 0 ## not all input events will produce output events; offset keeps track of the difference.
outputState = {
'guesses': [],
'eventData': [],
'indexes': [],
'xVals': [],
'yVals': []
}
for i in range(startEvent, stopEvent):
start = events[i]['time']
#sliceLen = 50e-3
sliceLen = dt*300. ## Ca2+ events are much longer than 50ms
if i+1 < len(events):
nextStart = events[i+1]['time']
sliceLen = min(sliceLen, nextStart-start)
guessLen = events[i]['len']*dt
tau = origTau
if tau is not None:
guessLen += tau*2.
#print i, guessLen, tau, events[i]['len']*dt
#sliceLen = 50e-3
sliceLen = guessLen
if i+1 < len(events): ## cut slice back if there is another event coming up
nextStart = events[i+1]['time']
sliceLen = min(sliceLen, nextStart-start)
## Figure out from where to pull waveform data that will be fitted
startIndex = np.argwhere(tvals>=start)[0][0]
stopIndex = startIndex + int(sliceLen/dt)
eventData = waveform[startIndex:stopIndex]
times = tvals[startIndex:stopIndex]
#print i, startIndex, stopIndex, dt
if len(times) < 4: ## PSP fit requires at least 4 points; skip this one
offset += 1
continue
## reconvolve this chunk of the signal if it was previously deconvolved
if tau is not None:
eventData = functions.expReconvolve(eventData, tau=tau, dt=dt)
#print i, len(eventData)
## Make guesses as to the shape of the event
mx = eventData.max()
mn = eventData.min()
if mx > -mn:
peakVal = mx
else:
peakVal = mn
guessAmp = peakVal * 2 ## fit converges more reliably if we start too large
guessRise = guessLen/4.
guessDecay = guessLen/2.
guessStart = times[0]
zc = functions.zeroCrossingEvents(eventData - (peakVal/3.))
## eliminate events going the wrong direction
if len(zc) > 0:
if guessAmp > 0:
zc = zc[zc['peak']>0]
else:
zc = zc[zc['peak']<0]
#print zc
## measure properties for the largest event within 10ms of start
zc = zc[zc['index'] < 10e-3/dt]
if len(zc) > 0:
if guessAmp > 0:
zcInd = np.argmax(zc['sum']) ## the largest event in this clip
else:
zcInd = np.argmin(zc['sum']) ## the largest event in this clip
zcEv = zc[zcInd]
#guessLen = dt*zc[zcInd]['len']
guessRise = .1e-3 #dt*zcEv['len'] * 0.2
guessDecay = dt*zcEv['len'] * 0.8
guessStart = times[0] + dt*zcEv['index'] - guessRise*3.
## cull down the data set if possible
cullLen = zcEv['index'] + zcEv['len']*3
if len(eventData) > cullLen:
eventData = eventData[:cullLen]
times = times[:cullLen]
## fitting to exponential rise * decay
## parameters are [amplitude, x-offset, rise tau, fall tau]
guess = [guessAmp, guessStart, guessRise, guessDecay]
#guess = [amp, times[0], guessLen/4., guessLen/2.] ## careful!
bounds = [
sorted((guessAmp * 0.1, guessAmp)),
sorted((guessStart-min(guessRise, 0.01), guessStart+guessRise*2)),
sorted((dt*0.5, guessDecay)),
sorted((dt*0.5, guessDecay * 50.))
]
yVals = eventData.view(np.ndarray)
fit = functions.fitPsp(times, yVals, guess=guess, bounds=bounds, multiFit=multiFit)
computed = functions.pspFunc(fit, times)
peakTime = functions.pspMaxTime(fit[2], fit[3])
diff = (yVals - computed)
err = (diff**2).sum()
fracError = diff.std() / computed.std()
lengthOverDecay = (times[-1] - fit[1]) / fit[3] # ratio of (length of data that was fit : decay constant)
output[i-offset] = tuple(events[i]) + tuple(fit) + (peakTime, err, fracError, lengthOverDecay)
#output['fitTime'] += output['time']
#print fit
#self.events.append(eventData)
outputState['guesses'].append(guess)
outputState['eventData'].append(eventData)
outputState['indexes'].append(i)
outputState['xVals'].append(times)
outputState['yVals'].append(computed)
if offset > 0:
output = output[:-offset]
outputState['output'] = output
return outputState
def processData(self, data):
return functions.expReconvolve(data)
