def knn_1d_density(x,k=10,eps=0.01):
'''
Uses local K nearest neighbors to estimate a density and center of
mass at each point in a distribution. Returns a local density estimator in units of 1/input_units. For example, if a sequence
of times in seconds is provided, the result is an estimate of
the continuous time intensity function in units of Hz.
Parameters
----------
x : ndarray
List of points to model
k : integer
Number of nearest neighbors to use in local density estimate
Default is 10
eps : number
Small correction factor to avoid division by zero
Returns
-------
centers : ndarray
Point location of density estimates
density :
Density values at locations of centers
'''
x=np.float64(np.sort(x))
# Handle duplicates by dithering
duplicates = get_edges(np.diff(x)==0.)+1
duplicates[duplicates>=len(x)-1]=len(x)-2
duplicates[duplicates<=0]=1
for a,b in zip(*duplicates):
n = b-a+1
q0 = x[a]
q1 = (x[a-1]-q0)
q2 = (x[b+1]-q0)
#print(a,b,q0,q1,q2)
x[a:b+1] += np.linspace(q1,q2,n+2)[1:-1]
intervals = np.diff(x)
centers = (x[1:]+x[:-1])*0.5
kernel = np.hanning(min(x.shape[0]-1,k)+2)[1:-1]
kernel /=sum(kernel)
intervals = np.convolve(intervals,kernel,'same')
density = (eps+1.0)/(eps+intervals)
return centers,density
def get_high_beta_events(session,area,channel,epoch,
MINLEN = 40, # ms
BOXLEN = 50, # ms
THSCALE = 1.5, # sigma (standard deviations)
lowf = 10.0, # Hz
highf = 45.0, # Hz
pad = 200, # ms
clip = True,
audit = False
):
'''
get_high_beta_events(session,area,channel,epoch) will identify periods of
elevated beta-frequency power for the given channel.
Thresholds are selected per-channel based on all available trials.
The entire trial time is used when estimating the average beta power.
To avoid recomputing, we extract beta events for all trials at once.
By default events that extend past the edge of the specified epoch will
be clipped. Passing clip=False will discard these events.
returns the event threshold, and a list of event start and stop
times relative to session time (not per-trial or epoch time)
passing audit=True will enable previewing each trial and the isolated
beta events.
>>> thr,events = get_high_beta_events('SPK120925','PMd',50,(6,-1000,0))
'''
# get LFP data
signal = get_raw_lfp_session(session,area,channel)
# esimate threshold for beta events
beta_trials = [get_filtered_lfp(session,area,channel,t,(6,-1000,0),lowf,highf) for t in get_good_trials(session)]
threshold = np.std(beta_trials)*THSCALE
print 'threshold=',threshold
N = len(signal)
event,start,stop = epoch
all_events = []
all_high_beta_times = []
for trial in get_good_trials(session):
evt = get_trial_event(session,area,trial,event)
trialstart = get_trial_event(session,area,trial,4)
epochstart = evt + start + trialstart
epochstop = evt + stop + trialstart
tstart = max(0,epochstart-pad)
tstop = min(N,epochstop +pad)
filtered = bandfilter(signal[tstart:tstop],lowf,highf)
envelope = abs(hilbert(filtered))
smoothed = convolve(envelope,ones(BOXLEN)/float(BOXLEN),'same')
E = array(get_edges(smoothed>threshold))+tstart
E = E[:,(diff(E,1,0)[0]>=MINLEN)
& (E[0,:]<epochstop )
& (E[1,:]>epochstart)]
if audit: print E
if clip:
E[0,:] = np.maximum(E[0,:],epochstart)
E[1,:] = np.minimum(E[1,:],epochstop )
else:
E = E[:,(E[1,:]<=epochstop)&(E[0,:]>=epochstart)]
if audit:
clf()
axvspan(epochstart,epochstop,color=(0,0,0,0.25))
plot(arange(tstart,tstop),filtered,lw=0.7,color='k')
plot(arange(tstart,tstop),envelope,lw=0.7,color='r')
plot(arange(tstart,tstop),smoothed,lw=0.7,color='b')
ylim(-80,80)
for a,b in E.T:
axvspan(a,b,color=(1,0,0,0.5))
axhline(threshold,color='k',lw=1.5)
xlim(tstart,tstop)
draw()
wait()
all_events.extend(E.T)
assert all(diff(E,0,1)>=MINLEN)
return threshold, all_events
