experiment.py

# ==================================================================== #
# Class for running synchronization experiments.
# Author: Eddie Lee edl56@cornell.edu
# ==================================================================== #

from .utils import *
import time
from datetime import datetime
from .axis_neuron import left_hand_col_indices,right_hand_col_indices
from .port import *
import dill
from subprocess import call


def ilogistic(x):
    """-np.log(1/x-1)"""
    return -np.log(1/x-1)

def logistic(x):
    """Squish from real line to [0,1].
    1/(1+np.exp(-x))
    """
    return 1/(1+np.exp(-x))


class HandSyncExperiment(object):
    def __init__(self,duration,trial_type,
                 parts_ix=None,
                 broadcast_port=5001,
                 anPort=7013,
                 fs=30,
                 rotation_angle=0,
                 check_directory=True,
                 verbose=False):
        """
        Parameters
        ----------
        duration : float
            Seconds into past to analyze for real time coherence measure.
        trial_type : str
            'avatar','avatar0','hand','hand0'
        parts_ix : list,None
            Indices of the columns in an_port file to extract and analyze. If None, this will be
            replaced automatically in start by the relevant hand.
        broadcast_port : int,5001
        fs : int
            Sampling frequency for interpolated velocities.
        rotation_angle : float
            Radians by which subject would have to be rotated about z-axis (pointing up) to face along the x-axis.
        check_directory : bool,True
        """
        self.duration = duration
        self.trialType = trial_type
        self.broadcastPort = broadcast_port
        self.rotAngle = rotation_angle
        self.verbose=verbose

        self.pause = []  # times when game was paused
        self.unpause = []  # times when game was resumed
        self.trialStartTimes = [] # times trials (excluding very first fully visible trial) were started
        self.trialEndTimes = [] # times trials end (including very first fully visible trial) were started

        self.anPort=anPort  # port at which to receive AN calculation broadcast

        # Check that data is being broadcast on anPort.
        self._check_an_port() 

        # Clear current directory.
        if len(os.listdir('./'))>0 and check_directory:
            self._clear_cd()

    def _clear_cd(self):
        from shutil import rmtree
        affirm='x'
        while not affirm in 'yn':
            affirm=input("Directory is not empty. Delete files? y/[n]")
        if affirm=='y':
            for f in os.listdir('./'):
                try:
                    os.remove(f)
                except OSError:
                    rmtree(f)
        else:
            raise Exception("There are files in current directory.")
    
    def _check_an_port(self):
        import socket
        import select
        try:
            listenSock = socket.socket(socket.AF_INET,socket.SOCK_DGRAM)
            listenSock.setblocking(0)
            listenSock.bind(('127.0.0.1',self.anPort))
            ready = select.select([listenSock], [], [], 1)
            if not ready[0]:
                raise Exception("No data is being broadcast on port.")
        finally:
            listenSock.close()

    def _load_avatar(self):
        """
        This loads the correct avatar for comparison of performance. The handedness of the subject is
        read in from left_or_right.txt.

        Parameters
        ----------

        Returns
        -------
        avatar : dict
            Dictionary of avatar interpolation splines.
        """
        from .pipeline import extract_motionbuilder_model2

        handedness = open('%s/%s'%(DATADR,'left_or_right')).readline().rstrip()
        
        # NOTE: This relies on the fact that these experiments took data from these avatar trials,
        # but you should fetch just the avatar data from the original recordings.
        if handedness=='left':
            v = extract_motionbuilder_model2('avatar',0,'Right',return_time=False)
        elif handedness=='right':
            v = extract_motionbuilder_model2('avatar',0,'Left',return_time=False)
        else:
            print(handedness)
            raise Exception

        return v

    def _load_avatar(self,return_subject=False):
        """
        This loads the correct avatar for comparison of performance. The handedness of the subject is read in
        from left_or_right.txt.

        Parameters
        ----------
        return_subject : bool,False

        Returns
        -------
        avatar : dict
            Dictionary of avatar interpolation splines.
        """
        from .data_access import subject_settings_v3 as subject_settings
        from .data_access import VRTrial3_1 as VRTrial
        handedness = open('%s/%s'%(DATADR,'left_or_right')).readline().rstrip()
        
        # NOTE: This relies on the fact that these experiments took data from these avatar trials, but you
        # should fetch just the avatar data from the original recordings.
        if handedness=='left':
            person,modelhandedness,rotation,dr = subject_settings(0)
        elif handedness=='right':
            person,modelhandedness,rotation,dr = subject_settings(2)
        else:
            print(handedness)
            raise Exception

        trial = VRTrial(person,modelhandedness,rotation,dr)
        avatar = trial.templateTrial
        
        if return_subject:
            subject = trial.subjectTrial
            return avatar,subject
        return avatar

    def load_avatar(self,reverse_time=False,return_subject=False):
        """
        This loads the correct avatar for comparison of performance. The handedness of the subject is read in
        from left_or_right.txt.

        Parameters
        ----------
        reverse_time : bool,False
            If True, play avatar motion backwards in time.
        return_subject : bool,False

        Returns
        -------
        avatar : dict
            Dictionary of avatar interpolation splines.
        """
        from .pipeline import extract_motionbuilder_model3_3
        handedness = open('%s/%s'%(DATADR,'left_or_right')).readline().rstrip()
        
        if handedness=='left':
            v,t = extract_motionbuilder_model3_3('Right',reverse_time=reverse_time)
        elif handedness=='right':
            v,t = extract_motionbuilder_model3_3('Left',reverse_time=reverse_time)
        else:
            print(handedness)
            raise Exception

        return v

    def wait_for_start(self,dt=.1):
        """
        Wait til start file is written.
        """
        while not os.path.isfile('%s/%s'%(DATADR,'start')):
            time.sleep(dt)

    def wait_for_start_time(self):
        """
        Get the time at which the trial started.

        Returns
        -------
        t0 : datetime
            The time at which the trial was started.
        """
        while not os.path.isfile('%s/%s'%(DATADR,'start_time')):
            time.sleep(.5)
        # Give some time for the initial time to be written by UE4.
        time.sleep(.5)
        with open('%s/%s'%(DATADR,'start_time'),'r') as f:
            t0 = datetime.strptime(f.readline(),'%Y-%m-%dT%H:%M:%S.%f')
        return t0
   
    def wait_for_start_gpr(self):
        """
        Once start_gpr has been written, erase self.subVBroadcast's memory of the history of
        velocity.
        """
        while not os.path.isfile('%s/%s'%(DATADR,'start_gpr')):
            time.sleep(.5)

    def read_pause(self):
        readFile = False
        while not readFile:
            try:
                with open('%s/%s'%(DATADR,'pause_time')) as f:
                    self.pause.append( datetime.strptime(f.readline(),'%Y-%m-%dT%H:%M:%S.%f') )
                readFile = True
            except IOError:
                time.sleep(.02)

    def read_this_setting(self):
        """
        Read in window settings and trial start and end times from this_setting file. Times are appended to
        self.trialStartTimes and self.trialEndTimes.

        Returns
        -------
        thisDuration : float
        thisFraction : float
        """
        # Try to open and read. Sometimes there is a delay in accessibility because
        # the file is being written.
        success = False 
        while not success:
            try:
                with open('%s/%s'%(DATADR,'this_setting')) as f:
                    L = f.readline().split(',')
                    thisDuration,thisFraction = (float(i) for i in L[:2])
                    # start start times with the first trial after full visibility
                    self.trialStartTimes.append( datetime.strptime(L[2],'%Y-%m-%dT%H:%M:%S.%f') )
                    self.trialEndTimes.append( datetime.strptime(L[3],'%Y-%m-%dT%H:%M:%S.%f') )
                success = True
            except IOError:
                pass
        return thisDuration,thisFraction
    
    def delete_file(self,fname,max_wait_time=1,dt=.02):
        """
        Try to delete file in DATADR. Return False if deletion is unsuccessful in given time frame.

        Parameters
        ----------
        fname : str
            Just the file name assuming that it is in DATADR.
        max_wait_time : float
        dt : float
            Time to wait in each iteration of while loop.
        """
        notDeleted = True
        t0 = datetime.now()
        while notDeleted:
            try:
                os.remove('%s/%s'%(DATADR,fname))
                notDeleted = False
            except OSError:
                if (datetime.now()-t0).total_seconds()>max_wait_time:
                    print("Failed to delete %s."%fname)
                    return False
                time.sleep(dt)
        print("%s deleted."%fname)
        return True
    
    def define_update_broadcaster(self,reader,stopEvent,pauseEvent,
                                  windowsInIndexUnits,realTimePerfEval,broadcast,
                                  rotAngle,avatar,t0):
        """Define function for real time performance assessment.

        Parameters
        ----------
        reader : ANReader
        stopEvent : threading.Event
        pauseEvent : threading.Event
        windowsInIndexUnits : int,
        realTimePerfEval : DTWPerformance
        broadcast : DataBroadcaster
        rotAngle : float
        avatar : Interpolation
        t0 : datetime.datetime

        Returns
        -------
        update_broadcaster : function
            Function for updating the performance broadcast port with latest performance value.
        """
        def update_broadcaster(performance,export=False):
            try:
                while not stopEvent.is_set():
                    pauseEvent.wait()
                    v,t,tAsDate = reader.copy_recent()
                    
                    if len(v)>=(windowsInIndexUnits):
                        # Put into standard coordinate system (as in paper). Account for reflection symmetry.
                        v[:,1:] *= -1
                        v[:,:2] = rotate_xy(v[:,:2],rotAngle)

                        tAsDate,_ = remove_pause_intervals(tAsDate.tolist(),list(zip(self.pause,self.unpause)))
                        avv = fetch_matching_avatar_vel(avatar,np.array(tAsDate),t0)
                        
                        # Calculate performance metric.
                        performance.append( realTimePerfEval.raw(v[:,1:],avv[:,1:],dt=1/30) )

                        # Update performance.
                        broadcast.update_payload('%1.2f'%performance[-1])
                        if self.verbose=='detailed':
                            print("new coherence is %s"%broadcast._payload)

                        if export:
                            if not os.path.isdir('realtime_velocities'):
                                os.mkdir('realtime_velocities')
                            dill.dump({'v':v,'avv':avv},open('realtime_velocities/%s.p'%str(export).zfill(4),
                                                             'wb'),-1)
                            export+=1
                    time.sleep(0.2)
            finally:
                print("updateBroadcastThread stopped")
        return update_broadcaster

    def run_cal(self,verbose=False,min_v=0.3,pause_before_run=0.):
        """
        Run calibration recording. Have subjects stand straight facing direction of motion. Then, have them
        jerk both hands forward in parallel lines. Then the respective angles for rotating the vector to be
        along the x-axis will be calculated for the left and right hands and put into self.rotAngle.
                
        Parameters
        ----------
        verbose : bool,False
        min_v : float,0.3
            Subject must be moving at least this fast (m/s) along the xy-plane for the calibration to
            register.
        pause_before_run: float,0.
            Number of seconds to wait before running. Used for debugging.
        """
        from numpy.linalg import norm
        counter = 0
        calSuccess = False
        fname = 'an_port_cal.txt'

        while not calSuccess:
            while os.path.isfile(fname):
                fname = 'an_port_cal_%s.txt'%(str(counter).zfill(2))
                counter+=1
            
            input("Press Enter to calibrate...")

            # Setup thread for recording port data.
            recordThread = threading.Thread(target=record_AN_port,
                                            args=(fname,self.anPort),
                                            kwargs={'start_file':'start_cal','stop_file':'stop_cal'})
            time.sleep(pause_before_run)

            print("Running calibration.")
            recordThread.start()

            # Run calibration for a few seconds to give people a chance to move their hands.
            with open('start_cal','w') as f:
                f.write('')
            time.sleep(5)
            with open('stop_cal','w') as f:
                f.write('')
            
            time.sleep(.5)

            # Delete signal files.
            print("Done with calibration.")
            self.delete_file('start_cal')
            self.delete_file('stop_cal')
            recordThread.join()
            time.sleep(2)

            # Load the data and find which direction the user is facing. Extract from that, the
            # rotation angle needed about the z-axis (pointing up out of the ground) to make the person face the
            # x-axis.
            df = load_AN_port(fname,time_as_dt=False)

            # Get xy vector.
            vright = df.iloc[:,right_hand_col_indices()].values[:,:2]
            vleft = df.iloc[:,left_hand_col_indices()].values[:,:2]
            vright[:,1] *= -1
            vleft[:,1] *= -1
            sright = np.linalg.norm(vright,axis=1)
            sleft = np.linalg.norm(vleft,axis=1)
            
            try:
                # Extract 80 percentile of speed for analysis (as long as it is at least min_v).
                ix = (sright>=np.percentile(sright,80)) & (sright>min_v)
                angleRight = extract_rot_angle(vright[ix])
                ix = (sleft>=np.percentile(sleft,80)) & (sleft>min_v)
                angleLeft = extract_rot_angle(vleft[ix])

                calSuccess = True
            except AssertionError:
                print("Retry calibration.")

        self.rotAngle = [-angleLeft,-angleRight]
        print("Rotation angle to center left hand about x-axis is %1.1f degrees."%(
                self.rotAngle[0]*180/np.pi))
        print("Rotation angle to center right hand about x-axis is %1.1f degrees."%(
                self.rotAngle[1]*180/np.pi))
    
    @staticmethod
    def read_cal(fname,min_v):
        """
        Read calibration recording from file. 
                
        Parameters
        ----------
        fname : str
            Name of file to save as.
        min_v : float,0.3
            Subject must be moving at least this fast (m/s) along the xy-plane for the calibration to
            register.
        """
        from numpy.linalg import norm

        # Load the data and find which direction the user is facing. Extract from that, the
        # rotation angle needed about the z-axis (pointing up out of the ground) to make the person face the
        # x-axis.
        df = load_AN_port(fname,time_as_dt=False)

        # Get xy vector.
        vright = df.iloc[:,right_hand_col_indices()].values[:,:2]
        vleft = df.iloc[:,left_hand_col_indices()].values[:,:2]
        vright[:,1] *= -1
        vleft[:,1] *= -1
        sright = np.linalg.norm(vright,axis=1)
        sleft = np.linalg.norm(vleft,axis=1)
        
        # Extract 80 percentile of speed for analysis (as long as it is at least min_v).
        ix = (sright>=np.percentile(sright,80)) & (sright>min_v)
        angleRight = extract_rot_angle(vright[ix])
        ix = (sleft>=np.percentile(sleft,80)) & (sleft>min_v)
        angleLeft = extract_rot_angle(vleft[ix])

        return [-angleLeft,-angleRight]

    def run_lf(self,trial_duration):
        """
        Run Leader-Follower experiment.

        Parameters
        ----------
        trial_duration : float
            Duration in seconds for which to record data.
        """
        suffix = 0
        while os.path.isfile('an_port_%s.txt'%str(suffix).zfill(2)):
            suffix += 1

        recordThread = threading.Thread(target=record_AN_port,
                                        args=('an_port_%s.txt'%str(suffix).zfill(2),self.anPort),
                                        kwargs={'start_file':'start_lf','stop_file':'end_lf'})
        with open('start_lf','w') as f:
            f.write('')
        recordThread.start()

        time.sleep(trial_duration)

        with open('end_lf','w') as f:
            f.write('')
        recordThread.join()
        self.delete_file('start_lf')
        self.delete_file('end_lf')

    def run_vr(self,
               update_delay=.3,
               initial_window_duration=1.0,initial_vis_fraction=0.5,
               min_window_duration=.5,max_window_duration=2,
               min_vis_fraction=.1,max_vis_fraction=1.,
               gpr_mean_prior=ilogistic(.42),
               reverse_time=False,
               verbose=False,
               export_realtime_velocities=False):
        """
        Run realtime analysis for experiment.
                
        Parameters
        ----------
        update_delay : float,.3
            Number of seconds to wait between updating arrays when calculating realtime coherence.
        initial_window_duration : float,1.0
        initial_vis_fraction : float,0.5
        min_window_duration : float,.5
        max_window_duration : float,2
        min_vis_fraction : float,.1
        max_vis_fraction : float,.9
        gpr_mean_prior : float,ilogistic(.42)
        reverse_time : bool,False
        verbose : bool,False
        export_realtime_velocities : bool,False
            For debugging. On every iteration of the real time performace evaluation, save the
            subject and avatar velocities in the folder realtime_velocities.
        
        Notes
        -----
        Code waits til when start_time becomes available to read in avatar start time.
        
        The threads that are running:
        0. reader thread to read velocities from Axis Neuron UDP port rebroadcast 7011.
        1. updateBroadcastThread: assess subject's performance relative to the avatar and
            update performance value
        2. broadcastThread: broadcast subject's performance to port 5001
        3. recordThread: record AN output from UDP rebroadcast @ self.anPort
        
        Thread communication happens through members that are updated using thread locks.
        
        In while loop, run GPR prediction step and write the next window duration and visible
        fraction to file. Waiting for run_gpr and writing to next_setting.

        When end is written, experiment ends.
        """
        from .data_access import subject_settings_v3
        from .data_access import VRTrial3_1 as VRTrial
        from .coherence import GPREllipsoid,DTWPerformance
        self.wait_for_start()
        
        # Setup routines for calculating coherence.
        gprmodel = GPREllipsoid(mean_performance=gpr_mean_prior,
                                tmin=min_window_duration,tmax=max_window_duration,
                                fmin=min_vis_fraction,fmax=max_vis_fraction)
        realTimePerfEval = DTWPerformance()
        gprPerfEval = DTWPerformance()

        nextDuration = np.around(initial_window_duration,1)
        nextFraction = np.around(initial_vis_fraction,1)
        assert min_window_duration<=nextDuration<=max_window_duration
        assert min_vis_fraction<=nextFraction<=max_vis_fraction

        with open('%s/next_setting'%DATADR,'w') as f:
            f.write('%1.1f,%1.1f'%(nextDuration,nextFraction))
        with open('%s/this_setting'%DATADR,'w') as f:
            f.write('%1.1f,%1.1f,%s,%s'%(0,0,datetime.now().isoformat(),datetime.now().isoformat()))
        with open('%s/left_or_right'%DATADR,'r') as f:
            handedness = f.read().rstrip()  # of subject
        if handedness=='left':
            self.subPartsIx = left_hand_col_indices(False)
            self.avPartsIx = right_hand_col_indices(False)
            rotAngle = self.rotAngle[0]
        else:
            self.subPartsIx = right_hand_col_indices(False)
            self.avPartsIx = left_hand_col_indices(False)
            rotAngle = self.rotAngle[1]
        avatar = self.load_avatar(reverse_time)  # avatar for comparing velocities
        windowsInIndexUnits = int(30*self.duration)
        performance = []  # history of performance
        
        pauseEvent = threading.Event()
        pauseEvent.set()
        self.endEvent = threading.Event()  # Event to be set when end file is written
        
        # Open port for communication with UE4 engine. This will send the current coherence value to
        # UE4.
        self.broadcast = DataBroadcaster(self.broadcastPort)
        self.broadcast.update_payload('-1.0')
        broadcastThread = threading.Thread(target=self.broadcast.broadcast,
                   kwargs={'pause':.2,'verbose':True if verbose=='detailed' else False})
        broadcastThread.start()

        # Setup thread for recording port data.
        recordThread = threading.Thread(target=record_AN_port,args=('an_port.txt',self.anPort))

        # Set up thread for updating value of streaming broadcast of coherence.
        # This relies on reader to fetch data which is declared later.
        self.updateBroadcastEvent = threading.Event()

        # Define function for GPR updating. One thread updates the settings. The other updates GPR.
        def update_settings(reader,gprmodel):
            """
            1) GPR is updated.
            2) Next trial setting is set.
            3) GPR is optimized. This can take a long time so the safest thing to do is to run it
                during a trial.
            4) gprmodel and other data is saved into gpr.p.
            """
            while not self.endEvent.is_set():
                # This is only triggered during pauses in order to prevent unnecessary CPU overhead, but I
                # don't think is necessary.
                pauseEvent.wait()
                
                # Update next trial settings and refresh reader history.
                if os.path.isfile('%s/%s'%(DATADR,'run_gpr')):
                    print("successfully read run_gpr")
                    # Fetch user movement during trial.
                    v,t,tdateHistory=reader.copy_history()
                    # Put output from Axis Neuron into comparable coordinate system accounting for reflection
                    # symmetry.
                    v[:,1:]*=-1
                    v[:,:2]=rotate_xy(v[:,:2],rotAngle)
                    tdateHistory,_=remove_pause_intervals( tdateHistory.tolist(),
                                                           list(zip(self.pause,self.unpause)) )
                    avv=fetch_matching_avatar_vel(avatar,np.array(tdateHistory),t0)
                    
                    # Update GPR with this trial's data.
                    # Try to open and read. Sometimes there is a delay in accessibility because
                    # the file is being written.
                    thisDuration,thisFraction=self.read_this_setting()
                    # Get subject performance ignoring the first few seconds of performance.
                    perf=gprPerfEval.time_average( avv[:,1:],v[:,1:],dt=1/30,bds=[2,np.inf] )
                    # Update GPR. For initial full visibility trial, update values for all values of fraction.
                    if thisDuration==0:
                        gprmodel.update( ilogistic(perf),0.,1. )
                    else:
                        gprmodel.update( ilogistic(perf),thisDuration,thisFraction )

                    # Get next trial settings and output them to a file that is read by UE4 before the start
                    # of the next trial.
                    # NOTE: There is no guarantee that this file is read before the next trial starts.
                    nextDuration,nextFraction=gprmodel.max_uncertainty()
                    if verbose:
                        #print call("ls --time-style='+%d-%m-%Y %H:%M:%S' -l this_setting",shell=True)
                        print("thisDuration: %1.1f\tthisFraction: %1.1f"%(thisDuration,thisFraction))
                        print("nextDuration: %1.1f\tnextFraction: %1.1f"%(nextDuration,nextFraction))
                    open('%s/next_setting'%DATADR,'w').write('%1.1f,%1.1f'%(nextDuration,nextFraction))
                    
                    # Refresh history.
                    self.broadcast.update_payload('-1.0')
                    if verbose:print("Refreshing reader history (update_settings).")
                    reader.refresh()
                    
                    # Optimize hyperparameters of GPR given the latest trial data.
                    # NOTE: This has to finish running before the trial ends. Right now, there is no guarantee
                    # that it will.
                    if verbose:print("Running GPR on this trial...")
                    gprmodel.optimize_hyperparams(verbose=verbose,n_restarts=1)
                    
                    # Cleanup.
                    self.delete_file('run_gpr')
                    dill.dump({'gprmodel':gprmodel,'performance':performance,
                               'pause':self.pause,'unpause':self.unpause,
                               'trialStartTimes':self.trialStartTimes,
                               'trialEndTimes':self.trialEndTimes},
                              open('%s/%s'%(DATADR,'gpr.p'),'wb'),-1)

                time.sleep(.05)

        # Wait til start_time has been written to start experiment.
        t0 = self.wait_for_start_time()

        if verbose:print("Starting threads.")
        recordThread.start()
        with ANReader(self.duration,self.subPartsIx,
                      port=7011,
                      verbose=True if verbose=='detailed' else False,
                      port_buffer_size=8192,
                      recent_buffer_size=(self.duration+1)*30) as reader:
            
            updateBroadcastThread = threading.Thread(
                    target=self.define_update_broadcaster(reader,self.updateBroadcastEvent,pauseEvent,
                                                          windowsInIndexUnits,realTimePerfEval,self.broadcast,
                                                          rotAngle,avatar,t0),
                    args=(performance,export_realtime_velocities,) )

            while reader.len_history()<windowsInIndexUnits:
                if verbose:print("Waiting to collect more data...(%d)"%reader.len_history())
                self.broadcast.update_payload('-1.0')
                time.sleep(.25)
            updateBroadcastThread.start()
           
            # Start GPR thread.
            settingsThread=threading.Thread(target=update_settings,args=(reader,gprmodel))
            settingsThread.start()

            while not os.path.isfile('%s/%s'%(DATADR,'end')):
                # If UE4 has been paused
                if os.path.isfile('%s/%s'%(DATADR,'pause_time')):
                    pauseEvent.clear()
                    if verbose:print("Paused.")
                    self.read_pause()
                    self.delete_file('pause_time')
                    while not os.path.isfile('%s/%s'%(DATADR,'unpause_time')):
                        time.sleep(.01)
                    
                    # Try to open and read unpause_time. Sometimes there is a delay in accessibility because
                    # the file is being written.
                    success = False 
                    while not success:
                        try:
                            with open('%s/%s'%(DATADR,'unpause_time')) as f:
                                self.unpause.append( datetime.strptime(f.readline(),'%Y-%m-%dT%H:%M:%S.%f') )
                            success = True
                            pauseEvent.set()
                            if verbose:print("Unpaused.")
                        except IOError:
                            pass
                    self.delete_file('unpause_time')
                    
                    if verbose:print("Refreshing reader history (pause).")
                    reader.refresh()
                
                time.sleep(.1)
         
        if verbose:print("Ending threads...")
        self.stop()
        updateBroadcastThread.join()
        broadcastThread.join()
        recordThread.join()
        settingsThread.join()

        with open('%s/%s'%(DATADR,'end_port_read'),'w') as f:
            f.write('')

        # Read in last trial setting.
        self.read_this_setting() 
        
        if verbose:print("Saving GPR.")
        dill.dump({'gprmodel':gprmodel,'performance':performance,
                   'pause':self.pause,'unpause':self.unpause,
                   'trialStartTimes':self.trialStartTimes,
                   'trialEndTimes':self.trialEndTimes,
                   'rotAngle':rotAngle},
                  open('%s/%s'%(DATADR,'gpr.p'),'wb'),-1)

        # Give time for UE4 to finish saving files.
        time.sleep(10)

        # Move all files into the left or right directory given by which hand the subject was using.
        if not os.path.isdir(handedness):
            os.mkdir(handedness)
        for f in os.listdir('./'):
            if os.path.isfile(f) or f=='capture':
                os.rename(f,'%s/%s'%(handedness,f))

    def stop(self):
        """Stop all thread that could be running. This does not wait for threads to stop."""
        self.updateBroadcastEvent.set()
        self.broadcast.stopEvent.set()
        self.endEvent.set()
# end HandSyncExperiment



def fetch_matching_avatar_vel(avatar,t,t0=None,verbose=False):
    """
    Get the stretch of avatar velocities that aligns with the velocity data of the subject. 

    Parameters
    ----------
    avatar : Interpolation
        This would be the templateTrial loaded in VRTrial.
    t : array of floats or datetime objects
        Stretch of time to return data from. If t0 is specified, this needs to be datetime objects.
    t0 : datetime,None
    verbose : bool,False

    Returns
    -------
    v : ndarray
        (n_time,3). Avatar's velocity that matches given time stamps relative to the starting time
        t0.
    """
    if not t0 is None:
        # Transform dt to time in seconds.
        t = np.array([i.total_seconds() for i in t-t0])
        assert (t>=0).all()
    if verbose:
        print("Getting avatar times between %1.1fs and %1.1fs."%(t[0],t[-1]))

    # Return part of avatar's trajectory that agrees with the stipulated time bounds.
    return avatar(t)

def remove_pause_intervals(t,pause_intervals,return_removed_ix=False):
    """
    Given a list of time points where data was taken and a list of tuples where the data take was paused,
    return the times at which the data would've been taken if there had been no pause having removed all data
    points that were taken during the specified pause intervals.

    Parameters
    ----------
    t : list
        datetime.datetime objects of when data was recorded. These should be ordered in time.
    pause_intervals : list of tuples
        Each tuple should be (start,end).
    return_removed_ix : bool,False
        If True, return the indices of all the entries in t that were removed.

    Returns
    -------
    tDate : list of datetime.datetime objects
    t : ndarray
        Time is seconds starting from tDate[0]
    removedIx : list
        Indices of elements removed from t.
    """
    t = t[:]
    pause_intervals = pause_intervals[:]
    removedIx=[]
    rangeT=list(range(len(t)))

    for dtix,(t0,t1) in enumerate(pause_intervals):
        assert t0<t1
        
        dt = t1-t0

        # Count up to the beginning of the pause interval.
        counter = 0
        t_ = t[counter]
        while t_<t0 and counter<len(t):
            t_ = t[counter]
            counter += 1

        # Remove all data points within the pause interval.
        if counter>0:
            while t[counter-1] < t1:
                t.pop(counter-1)
                removedIx.append(rangeT.pop(counter-1))
        
        # Subtract the duration of the removed pause interval from the remaining data.
        if counter<len(t):
            for counter in range(counter-1,len(t)):
                t[counter] -= dt
            for dtix in range(dtix+1,len(pause_intervals)):
                pause_intervals[dtix] = (pause_intervals[dtix][0]-dt,pause_intervals[dtix][1]-dt)
    if return_removed_ix:
        return t,np.concatenate([[0],np.cumsum([i.total_seconds() for i in np.diff(t)])]),removedIx
    return t,np.concatenate([[0],np.cumsum([i.total_seconds() for i in np.diff(t)])])

def extract_rot_angle(v,noise_threshold=.4,min_points=10):
    """
    Take average normalized vector and use that to calculate rotation angle of vector.  There can be a set of
    velocities for forward and then backwards movement that are both used to get a better approximation of the
    direction of motion  Rotation angle is along direction of initial movement.

    This needs to be negated to get the angle that we need to rotate about the z-axis to get the vector to
    point along the x-axis.
    
    Parameters
    ----------
    v : ndarray
        2d velocity measurements. (n_samples,2)
    noise_threshold : float,0.4
        Allowed noise in fluctuation of angle amongst given time points. If it is too 
        large, an error is thrown.

    Returns
    -------
    rotAngle : float
        Angle of initial velocity.
    """
    from misc.angle import mod_angle
    assert len(v)>min_points
    vnorm = np.linalg.norm(v,axis=1)[:,None]
    assert (vnorm>0).all(),"Zero velocities not allowed."
    v = v/vnorm
    
    # Orient velocities such that velocity vectors when hands are moving back are facing 
    # in the same direction as when moving forwards.
    angle = np.arctan2(v[:,1],v[:,0])
    ix = np.abs(mod_angle(angle[0]-angle)) > (.5*np.pi)
    assert ix.any(), "Failed to get both forward and backward directions."
    assert np.diff(ix).sum()==1, "All vectors should point in same direction before and after switch in direction."
    
    # Check noise by looking at spread of angle.
    v[ix] = rotate_xy(v[ix],np.pi)
    angle = np.arctan2(v[:,1],v[:,0])
    assert mod_angle(angle[0]-angle[1:]).std()<noise_threshold, "Angle measurements are noisy."
    
    # Calculate final value by averaging across vectors.
    v = v.mean(0)
    v /= np.linalg.norm(v)
    
    return np.arctan2(v[1],v[0])