HMM.py

#This file considers "Weekday + time of day"

import numpy as np
from Mocapy import *
from numpy import *
from scipy import *
#import statsmodels.api as sm
import pandas
from patsy import dmatrices
from decimal import *
from pandas import *
import pandas as pd
from cluster import KMeansClustering
import numpy as np
from numpy import vstack,array
from scipy.cluster.vq import kmeans,vq
from cluster import KMeansClustering
#from RndCoV import RndCov
#from ghmm import *
from sklearn import hmm
from sklearn.hmm import MultinomialHMM
from itertools import *
import operator 

import csv
import time
import re
from datetime import datetime
from pytz import timezone
import pytz
#import xlwt
import pdb
import matplotlib
import matplotlib.pyplot as plt
import pylab
from pylab import plot, show
from matplotlib.dates import date2num
from time import mktime
import matplotlib.pylab as mp
from matplotlib.dates import MinuteLocator, DateFormatter, HourLocator
from pylab import figure
from scipy.optimize import curve_fit




#read two files
with open('20130128_offices.csv','rb') as csvfile:
    offices = csv.reader(csvfile)        
    officelist = list()
    for row in offices:
        officelist.append(row)
officelist.pop(0)
        
OfficeId = list()
for item in officelist:
    OfficeId.append(item[0])
Num_Office = len(OfficeId)
print "The number of office is ",
print Num_Office
        
with open('20130323_waiting_times.csv') as csvfile:
    waiting_times = csv.reader(csvfile)
    waiting_timeslist = list()
    for row in waiting_times:
        waiting_timeslist.append(row)

waiting_timeslist.pop(0)

number = len(waiting_timeslist)
    
#define a hashfunction to map the same bucket (ignoring weekday) to a unique value
#Then put the hash value in item[6]

def Hashfunction(item):
    Hash = (str(item[4].weekday())+str(item[4].hour).zfill(2)+str((item[4].minute)/10))
    return Hash

#Define a PutInMap function : put the item with the same hash value into the same key
#the key is the has value
#the value is a list
#0:total waiting time   1: total times
def PutInMap(item,Map):
    temp = item[6]
    if(temp not in Map):
        Map[temp] = list()
        Map[temp].append(float(item[3]))               #if temp is not in hashmap yet 
        Map[temp].append(1)
    else:
        Map[temp][0] = Map[temp][0] + float(item[3])   #add the waiting time
        Map[temp][1] = Map[temp][1] + 1                #count the times

#This function deals with test data set
#The key is the hash value
#the value is a list 
#put the corresponding wo_appointment into the list
def PutInMap2(item,Map):
    temp = item[6]
    if(temp not in Map):
        
        Map[temp] = list()
        Map[temp].append(item[3])
    else:
        Map[temp].append(item[3])

# Compare the key sets of two dictionary
# If they are equal, return true
def ComKeyDic(dic1,dic2):
    return (set(dic1.keys()) == set(dic2.keys()))

# Convert String to Datetime
# each id:
# 0:d 1:office_id 2:w_appointment 3:wo_appointment 4:created_at 5:updated_at 6:Hash Value

col = 4
for i in xrange(number):
    
    #extract the time string and convert them to standard EST datetime
    
    temp = datetime.strptime(waiting_timeslist[i][col], "%Y-%m-%d %H:%M:%S")   
    utc = pytz.UTC
    ams = pytz.timezone('US/Pacific')
    waiting_timeslist[i][col] = utc.localize(temp)
    waiting_timeslist[i][col] = waiting_timeslist[i][col].astimezone(ams)
    
    temp = waiting_timeslist[i]
    waiting_timeslist[i].append(Hashfunction(temp))
    
   
    
#Split the whole set into train_set and test_set

train_set = waiting_timeslist[1:number/2]
test_set = waiting_timeslist[number/2:]


AllTimeList = list()
WaitTimeList = list()

#Office548 = dict()
Officeid = 632

WaitTimeArray = list()
for item in waiting_timeslist:
    if(int(item[1])==Officeid):
        if(item[4].weekday()==2):
            if(item[4].hour>=9 and item[4].hour<=17):
                if(item[4].hour!=17):
                    WaitTimeArray.append(float(item[3]))
                    AllTimeList.append(item[6])
                else:
                    if(item[4].minute==0):
                        WaitTimeArray.append(float(item[3]))
                        AllTimeList.append(item[6])
                        
        else:
            if(item[4].hour>=8 and item[4].hour<=17):
                if(item[4].hour!=17):
                    WaitTimeArray.append(float(item[3]))
                    AllTimeList.append(item[6])
                else:
                    if(item[4].minute==0):
                        WaitTimeArray.append(float(item[3]))
                        AllTimeList.append(item[6])

#Remove the day with all zeros
def FindRemoveDay(TempTime):
    Day = dict()
    for item in TempTime:
        if (item[0] not in Day):
            Day[item[0]] = 1
        else:
            TempCount = Day[item[0]] + 1
            Day[item[0]] = TempCount
    return max(Day.iterkeys() , key=(lambda key : Day[key]))
    

groups = []
uniquekeys = []
for k, g in groupby(enumerate(zip(AllTimeList,WaitTimeArray)), lambda(i,(x,y)): y==0.0):
    groups.append(list(g))
    uniquekeys.append(k)

RemoveIndex = list()

for i in range(len(uniquekeys)):
    if(uniquekeys[i]==True):
        if(len(groups[i])>=50):
            (Index,TempValue) = zip(*groups[i])
            (TempTime,TempWaitTime) = zip(*TempValue)
            RemoveDay = FindRemoveDay(TempTime)
            for j in range(len(TempTime)):
                if(TempTime[j][0]==RemoveDay):
                    RemoveIndex.append(Index[j])
tempWaitTime = [item for i, item in enumerate(WaitTimeArray) if i not in RemoveIndex]
WaitTimeArray = tempWaitTime
tempTime = [item for i , item in enumerate(AllTimeList) if i not in RemoveIndex]
AllTimeList = tempTime


                    
'''
#Compute the centroids for this office 
WaitTimeArray = list()
for item in waiting_timeslist:
    if(int(item[1])==Officeid):
        WaitTimeArray.append(float(item[3]))
        AllTimeList.append(item[6])
'''
'''
WaitTimeArray = np.array(WaitTimeArray)
np.vstack(WaitTimeArray)
###############################################################change the number of clusters here 
centroids,_ = kmeans(WaitTimeArray,5, iter=30,thresh=1e-6)
clusters,_ = vq(WaitTimeArray,centroids)

print "The centroids are: " + repr(centroids)
print "The result of clustering is :" + repr(clusters)
NumCluster = len(centroids)
print "The number of centroids is :" + repr(NumCluster)

#Construct the training data
TimeList = list()
for i in xrange(len(WaitTimeArray)/2):
    TimeList.append(AllTimeList[i])
    WaitTimeList.append(clusters[i])
    
#Construct the testing data
TestTimeList = list()
TestWaitTimeList = list()
for i in xrange(len(WaitTimeArray)/2,len(WaitTimeArray)):
    TestTimeList.append(AllTimeList[i])
    TestWaitTimeList.append(clusters[i])
'''
#################################################################equi-height histogram

NumElement = len(WaitTimeArray)
SortedWaitTime = sorted(WaitTimeArray)


'''
ZeroWaitTime = list()
NonZeroWaitTime = list()
for i in range(len(SortedWaitTime)):
    if(SortedWaitTime[i]==0):
        ZeroWaitTime.append(0)
    else:
        NonZeroWaitTime.append(SortedWaitTime[i])
'''
###############################################change the number of group here 
NumGroup = 20
#NumNonZero = len(SortedWaitTime)
Count = math.ceil(float(NumElement)/NumGroup)
GroupWaitTime = dict()
GroupIndex = 0
GroupWaitTime[GroupIndex] = list()
for i in range(NumElement):
    GroupWaitTime[GroupIndex].append(SortedWaitTime[i])                 
    if((i+1) % Count == 0):
        GroupIndex = GroupIndex + 1
        GroupWaitTime[GroupIndex] = list()

centroids= list()
#centroids.append(0.0)        #add zeros to centroids
for key in GroupWaitTime:
    centroids.append(sum(GroupWaitTime[key])/float(len(GroupWaitTime[key])))
centroids = sorted(centroids)

# change the continous waiting time into discrete waiting time 
DisWaitingTime = list()
for i in range(len(WaitTimeArray)):
    '''
    if(WaitTimeArray[i]==0):
        DisWaitingTime.append(WaitTimeArray[i])
    '''   
    
    DisWaitingTime.append(min(centroids,key=lambda x:abs(x-WaitTimeArray[i])))
        
centroids = sorted(centroids)

# draw the histogram for all the data for each bucket
TotalMap = dict()
for i in range(len(WaitTimeArray)):
    if(TotalMap.has_key(AllTimeList[i])):
        TotalMap[AllTimeList[i]].append(WaitTimeArray[i])
    else:
        TotalMap[AllTimeList[i]] = list()
        TotalMap[AllTimeList[i]].append(WaitTimeArray[i])
'''
keyList = list()
for key in sorted(TotalMap.iterkeys()):
    keyList.append(key)


for k in range(1,13):
    plt.figure(k)
    j = 1
    for i in range((k-1)*22,(k-1)*22+22):
        plt.subplot(4,6,j)
        TempList = list(TotalMap[keyList[i]])
        plt.hist(TempList)
        #plt.title(keyList[i])
        plt.savefig(str(k) + '.jpg')
        j = j + 1

#plt.autoscale_view(True,True,True)
plt.show()
'''

'''

TotalMapList = list()
#for 
    #frequency, bin_edges = np.histogram(np.array(TotalMap[key]))
    #plt.subplot(12,22,i)
    #TempList = list(TotalMap[key])
    #plt.hist(TempList)
    #plt.title(key)
    #i = i + 1
    #TotalMapList.extend(list(frequency))
plt.hist(TotalMapList)
plt.show()
'''
    
#draw the histogram for all the data for each hour
'''
HourMap = dict()
for i in range(len(WaitTimeArray)):
    Hour = AllTimeList[i][0] + AllTimeList[i][1] + AllTimeList[i][2]
    if(HourMap.has_key(Hour)):
        HourMap[Hour].append(WaitTimeArray[i])
    else:
        HourMap[Hour] = list()
        HourMap[Hour].append(WaitTimeArray[i])

HourKeyList = list()
for key in sorted(HourMap.iterkeys()):
    HourKeyList.append(key)
'''
    
#draw the histogram for the training data for each hour
TrainWaitTimeArray = WaitTimeArray[:len(WaitTimeArray)/2]
TrainTimeList = AllTimeList[:len(AllTimeList)/2]
HourMap = dict()
for i in range(len(TrainWaitTimeArray)):
    Hour = TrainTimeList[i][1] + TrainTimeList[i][2] + TrainTimeList[i][3]
    if(HourMap.has_key(Hour)):
        HourMap[Hour].append(TrainWaitTimeArray[i])
    else:
        HourMap[Hour] = list()
        HourMap[Hour].append(TrainWaitTimeArray[i])

HourKeyList = list()
for key in sorted(HourMap.iterkeys()):
    HourKeyList.append(key)
    
#Define model function to be used to fit to the data
def gauss(x, *p):
    A, mu, sigma = p
    return A*np.exp(-(x-mu)**2/(2.*sigma**2))

GaussianMap = dict()
for key in HourKeyList:
    data = HourMap[key]
    hist, bin_edges = np.histogram(data,density=True)
    bin_centres = (bin_edges[:-1] + bin_edges[1:])/2
    #p0 = [1.,0.,1.]
    x = sum(bin_centres*hist)/sum(hist)
        #coeff, var_matrix = curve_fit(gauss, bin_centres, hist,p0=p0)
        #hist_fit = gauss(bin_centres, *coeff)
    width = sqrt(abs(sum((bin_centres-x)**2*hist)/sum(hist)))
    GaussianMap[key] = list()
    GaussianMap[key].append(x)
    GaussianMap[key].append(width)
    
    
#fit Gaussian distribution to each hour's waiting time


'''
for k in range(13,24):
    plt.figure(k)
    j = 1
    for i in range((k-13)*4, (k-12)*4):
        plt.subplot(2,2,j)
        TempList = list(HourMap[HourKeyList[i]])
        plt.hist(TempList)
        plt.title(str(int(HourKeyList[i][0])+1) + HourKeyList[i][1] + HourKeyList[i][2])
        plt.savefig('Hour'+str(k) + '.jpg')
        j = j + 1
plt.show()
'''

    


######################################Convert the discrete waiting time into the index of state number, DisWaitingTime
for i in range(len(WaitTimeArray)): 
    temp = centroids.index(DisWaitingTime[i])
    DisWaitingTime[i] = temp

        
#Construct the training data, WaitTimeList and TimeList is training data
TimeList = list()
for i in xrange(len(WaitTimeArray)/2):
    TimeList.append(AllTimeList[i])
    WaitTimeList.append(DisWaitingTime[i])
    
#Construct the testing data
TestTimeList = list()
TestWaitTimeList = list()
for i in xrange(len(WaitTimeArray)/2,len(WaitTimeArray)):
    TestTimeList.append(AllTimeList[i])
    TestWaitTimeList.append(DisWaitingTime[i])

DisWaitingTime = np.array(DisWaitingTime)
WaitingTimeValue = np.unique(DisWaitingTime)
NumWaitingTime = WaitingTimeValue.size

#convert the time value to integer 
TimeList = [int(x) for x in TimeList]
TimeArray = np.array(TimeList)
TimeValue = np.unique(TimeArray)
NumTime = len(TimeValue)

#Calculate the probability P(wt|time)
#Convert timevalue to string
TimeString = [str(x).zfill(4) for x in TimeList]
TimeString = [(str(int(x[0])+1) + x[1:]) for x in TimeString]
TimeValueString = [str(x).zfill(4) for x in list(TimeValue)]
TimeValueString = [(str(int(x[0])+1) + x[1:]) for x in TimeValueString]
PredictWaiting = DataFrame(zeros(NumTime,NumWaitingTime),index=TimeValueString,columns=WaitingTimeValue)
for i in range(len(WaitTimeList)):
    tempindex = TimeString[i]
    tempwaitingtime = WaitTimeList[i]
    count = PredictWaiting[tempindex][i]
    count = count + 1
    PredictWaiting[tempindex][i] = count

for row in TimeValueString:
    PredictWaiting[row] = PredictWaiting[row]/sum(PredictWaiting[row])

#Construct the training sequences 
index = 0
WaitingTimeSeq = dict()
TimeSeq = dict()
OriginalWaitingTimeSeq = dict()

key = 0
while(index<len(TimeList)):
    WaitingTimeSeq[key] = list()
    TimeSeq[key] = list()
    OriginalWaitingTimeSeq[key] = list()
    print key
    while (True):    
        if((index+2)<=len(TimeList) and TimeList[index+1] >= TimeList[index]):
            WaitingTimeSeq[key].append(DisWaitingTime[index])
            TimeSeq[key].append(TimeList[index])
            OriginalWaitingTimeSeq[key].append(WaitTimeArray[index])
            index = index + 1
            print index
            continue
        else:
            WaitingTimeSeq[key].append(DisWaitingTime[index])
            TimeSeq[key].append(TimeList[index])
            OriginalWaitingTimeSeq[key].append(WaitTimeArray[index])
            key = key + 1
            index = index + 1
            break
NumSeq = key


'''
# Convert the sequence to a new sequence based on hour series
HourSeq = dict()
for key in WaitingTimeSeq:
    HourSeq[key] = dict()
    for i in range(len(WaitingTimeSeq[key])):
        HourKey = TimeSeq[key][i][0] + TimeSeq[key][i][1] + TimeSeq[key][i][2]
        if(HourSeq[key].has_key(HourKey)):
            HourMap[key][HourKey].append(WaitingTimeSeq[key][i])
        else:
            HourMap[key][HourKey] = list()
            HourMap[key][HourKey].append(WaitingTimeSeq[key][i])
'''

for i in range(NumSeq):
    print 'The {index}th WaitingTimesequence is: '.format(index=i+1) + repr(WaitingTimeSeq[i])

#index of minimum waiting_time 
Index0 = np.argmin(centroids)

#Construct the transition and emission matrix

Transition = np.zeros((NumWaitingTime,NumWaitingTime),'d')
Transition = DataFrame(Transition)
index = WaitingTimeValue
columns = WaitingTimeValue
Transition = DataFrame(Transition,index=index, columns=columns)
Count = Series(np.zeros(NumWaitingTime), index=WaitingTimeValue)
for key in WaitingTimeSeq:
    print 'This is the {index}st sequence: '.format(index = key+1) + repr(WaitingTimeSeq[key])
    for i in xrange(1,len(WaitingTimeSeq[key])):
        Current = Transition.get_value(WaitingTimeSeq[key][i],WaitingTimeSeq[key][i-1])
        Current = Current + 1
        Count[WaitingTimeSeq[key][i]] = Count[WaitingTimeSeq[key][i]] + 1
        Transition.set_value(WaitingTimeSeq[key][i],WaitingTimeSeq[key][i-1],Current)

#Decrease the number of transition from 0 to 0

#getcontext().prec = 3

CurrentCount = Transition.get_value(Index0,Index0)
##########################################################################  change the decreasing rate here 
CurrentCount = float(CurrentCount) * 1 
Transition.set_value(Index0,Index0,CurrentCount)
for row in WaitingTimeValue:
    SumRow = sum(Transition.ix[row])
    for column in WaitingTimeValue:
        Current = Transition.get_value(row,column)
        Current = Decimal((Current+1))/Decimal((SumRow+NumWaitingTime))
        Transition.set_value(row,column,Current)

# Compute the baseline prediction p(wt|wt-1) = p(wt-1|wt)(wt)/p(wt-1)
# wt,h = argmax(wt) p(wt,h|wt-1) = argmax(wt) p(wt-1|wt)*(wt,h)
BaseLinePre = Series(zeros(NumTime),index=TimeValueString)



'''
#Construct the transition matrix to compute P(wt|wt-1,h)
NewTransition = Panel(np.zeros(Num))
'''


#EMission Matrix    
Emission = np.zeros((NumWaitingTime,NumTime),'d')
index = WaitingTimeValue
columns = TimeValue
Emission = DataFrame(Emission,index=index,columns=columns)
Count1 = Series(np.zeros(NumWaitingTime), index=WaitingTimeValue)

for key in WaitingTimeSeq:
    for i in xrange(len(WaitingTimeSeq[key])):
        Current = Emission.get_value(WaitingTimeSeq[key][i],TimeSeq[key][i])
        Current = Current + 1
        Count1[WaitingTimeSeq[key][i]] = Count1[WaitingTimeSeq[key][i]] + 1
        Emission.set_value(WaitingTimeSeq[key][i],TimeSeq[key][i],Current)
for row in WaitingTimeValue:
    for column in TimeValue:
        Current = Emission.get_value(row,column)
        Current = (Current+1)/(Count1[row]+NumTime)
        Emission.set_value(row,column,Current)

#Initial Probability
Initial = Series(np.zeros(NumWaitingTime), index = WaitingTimeValue)
for key in WaitingTimeSeq:
    if (key>0) :
        Current = Initial[WaitingTimeSeq[key][0]]
        Current = Current + 1
        Initial[WaitingTimeSeq[key][0]] = Current
Initial = (Initial+1)/(NumSeq-1+NumWaitingTime)

#Convert pandas dataframe to numpy array
subset = Transition[WaitingTimeValue]
tuples = [tuple(x) for x in subset.values]
Transition = np.array(tuples)
subset = Emission[TimeValue]
tuples = [tuple(x) for x in subset.values]
Emission = np.array(tuples)
Initial = np.array(Initial)

#Use HMM to predict the waiting time
n_components = NumWaitingTime
n_symbols = NumTime
startprob = Initial
transmat = Transition
emissionprob_ = Emission
emitter = hmm.MultinomialHMM(n_components, startprob, transmat)
emitter.emissionprob_ = emissionprob_

#Construct testing sequence 
key = 0
index = 0
TestWaitTimeSeq = dict()
TestTimeSeq = dict()
#TestOriginWaitTimeSeq = dict()

while(index<len(TestTimeList)):
    TestWaitTimeSeq[key] = list()
    TestTimeSeq[key] = list()
    print key
    while (True):    
        if((index+2)<=len(TestTimeList) and TestTimeList[index+1] >= TestTimeList[index]):
            TestTimeSeq[key].append(TestTimeList[index])
            TestWaitTimeSeq[key].append(TestWaitTimeList[index])
            index = index + 1
            print index
            continue
        else:
            TestWaitTimeSeq[key].append(TestWaitTimeList[index])
            TestTimeSeq[key].append(TestTimeList[index])
           # TestOriginWaitTimeSeq[key].append(TestOriginWaitTimeList[index])
            #TestWeekSeq[key].append(TestWeekList[index])
           # TestWeekTimeSeq[key].append(TestWeekTimeList[index])
            key = key + 1
            index = index + 1
            break
#observations = arange(NumTime)
#observations = observations.reshape(len(observations),)
totalError = 0.0
totalCount = 0.0
for key in TestWaitTime:
    observations = TestWaitTime[key]
    Prediction = emitter.predict(observations)
    errors = sum(map(abs,Prediction - observations))
    totalCount += len(TestWaitTime[key])
averageError = totalError/totalCount
print "Average Error is: " + str(averageError) 
'''
#"Prediction" is the discrete waiting time state prediction
#"FinalPrediction" is the continous waiting time prediction 
for i in xrange(len(Prediction)):
    FinalPrediction.append(centroids[Prediction[i]])
print "The prediction is: " + repr(Prediction)
print "The FinalPrediction[i]" + repr(FinalPrediction)


plt.plot(FinalPrediction)
plt.ylabel('Waiting Hours')
plt.xlabel('Time(Week+Day)')
FinalPrediction = Series(FinalPrediction,index=TimeValue)

###############################################################################change between continous and discrete waiting time here 

#convert the discrete waiting time in testing data to continous value
for i in xrange(len(TestWaitTimeList)):
    TestWaitTimeList[i] = centroids[TestWaitTimeList[i]]


#TestMap is the dictionary whose key is "Time", and value is "Waiting time list" 
TestMap = dict()
for i in xrange(len(TestTimeList)):
    if(TestMap.has_key(int(TestTimeList[i]))):
        TestMap[int(TestTimeList[i])].append(TestWaitTimeList[i])
    else:
        TestMap[int(TestTimeList[i])] = list()
        TestMap[int(TestTimeList[i])].append(TestWaitTimeList[i])

Prediction = Series(Prediction,index=TimeValue)

#Compute the percentage of prediction error using "discrete waiting time state" 
ErrorCount = 0 
for key in Prediction.index:
    PredictionState = Prediction[key]
    for value in TestMap[key]:
        if(value!=PredictionState):
            ErrorCount = ErrorCount+1
TotalCount = len(TestTimeList)
ErrorPercent = float(ErrorCount)/TotalCount
print "The percentage of error is: " + repr(ErrorPercent)
    
#Compute the average prediction error using "continous waiting time" 
DifferMap = {}
for key in FinalPrediction.index:
    a = map(float, TestMap[key])
    b = [FinalPrediction[key]]*len(TestMap[key])    
    DifferList = [a-b for a,b in zip(a,b)]
    Average_Error = sum(map(abs,DifferList))/len(DifferList)
    
    DifferMap[key] = list()
    #DifferMap[key].append(CountTime[key][2])
    DifferMap[key].append(Average_Error)
error = (sum(DifferMap.values()))/float(len(DifferMap))
print "The average error is: " + repr(error) +" hours"

#plt.title('No Decrease, Error is: '+ repr(error)+ '(hour)')
#plt.title('Decrease by 5/6, Error is: '+repr(error) + '(hour)')
plt.title('3 State Error Percentage: '+ repr(ErrorPercent))
plt.show()
'''