def lcmnl_simulate(cmdata,numclasses,csdata,numalts,betas):
cmfnames = ['cm%d'%i for i in range(cmdata.shape[1])]
cmdata,newcmfnames = prep_cm_data(cmdata,numclasses,cmfnames)
cmchoices = mnl.mnl_simulate(cmdata,betas['cm'],numclasses,GPU=GPU,returnprobs=0)
cspdf = {}
for cls in range(numclasses):
cspdf['cs%d'%cls] = mnl.mnl_simulate(csdata,betas['cs%d'%cls],numalts,GPU=GPU,returnprobs=1).flatten()
cschoices = None
return cmchoices, cschoices, cspdf
def lcmnl_estimate(cmdata,numclasses,csdata,numalts,chosen,maxiter=MAXITER,emtol=EMTOL,\
skipprep=False,csbeta=None,cmbeta=None):
loglik = -999999
if csbeta is None: csbeta = [np.random.rand(csdata.shape[1]) for i in range(numclasses)]
if not skipprep: cmdata = prep_cm_data(cmdata,numclasses)
if cmbeta is None: cmbeta = np.zeros(cmdata.shape[1])
for i in range(maxiter):
print "Running iteration %d" % (i+1)
print time.ctime()
# EXPECTATION
print "Running class membership model"
cmprobs = mnl.mnl_simulate(cmdata,cmbeta,numclasses,GPU=GPU,returnprobs=1)
csprobs = []
for cno in range(numclasses):
tmp = mnl.mnl_simulate(csdata,csbeta[cno],numalts,GPU=GPU,returnprobs=1)
tmp = np.sum(tmp*chosen,axis=1) # keep only chosen probs
csprobs.append(np.reshape(tmp,(-1,1)))
csprobs = np.concatenate(csprobs,axis=1)
h = csprobs * cmprobs
oldloglik = loglik
loglik = np.sum(np.log(np.sum(h,axis=1)))
print "current csbeta", csbeta
print "current cmbeta", cmbeta
print "current loglik", loglik, i+1, "\n\n"
if abs(loglik-oldloglik) < emtol: break
wts = h / np.reshape(np.sum(h,axis=1),(-1,1))
# MAXIMIZATION
for cno in range(numclasses):
print "Estimating class specific model for class %d" % (cno+1)
t1 = time.time()
weights=np.reshape(wts[:,cno],(-1,1))
print weights.shape
fit, results = mnl.mnl_estimate(csdata,chosen,numalts,GPU=GPU,weights=weights,beta=csbeta[cno])
print "Finished in %fs" % (time.time()-t1)
csbeta[cno] = zip(*results)[0]
print "Estimating class membership model"
t1 = time.time()
fit, results = mnl.mnl_estimate(cmdata,None,numclasses,GPU=GPU,weights=wts,lcgrad=True, \
beta=cmbeta,coeffrange=(-1000,1000))
print "Finished in %fs" % (time.time()-t1)
cmbeta = zip(*results)[0]
def expectation(cmbeta,csbeta):
print "Running class membership model"
cmprobs = mnl.mnl_simulate(cmdata,cmbeta,numclasses,GPU=GPU,returnprobs=1)
csprobs = []
for cno in range(numclasses):
tmp = mnl.mnl_simulate(csdata,csbeta[cno],numalts,GPU=GPU,returnprobs=1)
tmp = np.sum(tmp*chosen,axis=1) # keep only chosen probs
csprobs.append(np.reshape(tmp,(-1,1)))
csprobs = np.concatenate(csprobs,axis=1)
h = csprobs * cmprobs
loglik = np.sum(np.log(np.sum(h,axis=1)))
wts = h / np.reshape(np.sum(h,axis=1),(-1,1))
return loglik, wts
def lcmnl_simulate(cmdata, numclasses, csdata, numalts, betas):
cmfnames = ['cm%d' % i for i in range(cmdata.shape[1])]
cmdata, newcmfnames = prep_cm_data(cmdata, numclasses, cmfnames)
cmchoices = mnl.mnl_simulate(cmdata,
betas['cm'],
numclasses,
GPU=GPU,
returnprobs=0)
cspdf = {}
for cls in range(numclasses):
cspdf['cs%d' % cls] = mnl.mnl_simulate(csdata,
betas['cs%d' % cls],
numalts,
GPU=GPU,
returnprobs=1).flatten()
cschoices = None
return cmchoices, cschoices, cspdf
def lcmnl_estimate(cmdata,numclasses,csdata,numalts,chosen,maxiter=MAXITER,emtol=EMTOL):
loglik = -999999
beta = [np.ones(csdata.shape[1]) for i in range(numclasses)]
#cmbeta = np.ones((cmdata.shape[1]+1)*(numclasses-1)) # +1 if for asc, -1 is base alt
cmbeta = np.ones(cmdata.shape[1])
for i in range(maxiter):
print "Running iteration %d" % (i+1)
# EXPECTATION
print "Running class membership model"
print "cmbeta", cmbeta
cmprobs = mnl.mnl_simulate(cmdata,cmbeta,numclasses,GPU=GPU,returnprobs=1)
print "cmprobs", cmprobs
csprobs = []
for cno in range(numclasses):
print "Running class specific model for class %d" % (cno+1)
print "csbeta", beta[cno]
tmp = mnl.mnl_simulate(csdata,beta[cno],numalts,GPU=GPU,returnprobs=1)
print "csprobs", tmp
tmp = np.sum(tmp*chosen,axis=1) # keep only chosen probs
csprobs.append(np.reshape(tmp,(-1,1)))
csprobs = np.concatenate(csprobs,axis=1)
h = csprobs * cmprobs
oldloglik = loglik
loglik = np.sum(np.log(np.sum(h,axis=1)))
print "current loglik", loglik, i+1
print "current beta", beta
if abs(loglik-oldloglik) < emtol: break
wts = h / np.reshape(np.sum(h,axis=1),(-1,1))
for i in range(wts.shape[1]): print i, pd.Series(wts[:,i]).describe()
# MAXIMIZATION
print "Estimating class membership model"
fit, results = mnl.mnl_estimate(cmdata,None,numclasses,GPU=GPU,weights=wts,lcgrad=True)
cmbeta = zip(*results)[0]
for cno in range(numclasses):
print "Estimating class specific model for class %d" % (cno+1)
fit, results = mnl.mnl_estimate(csdata,chosen,numalts,GPU=GPU,weights=wts[:,cno])
beta[cno] = zip(*results)[0]
def expectation(cmbeta, csbeta):
print "Running class membership model"
cmprobs = mnl.mnl_simulate(cmdata,
cmbeta,
numclasses,
GPU=GPU,
returnprobs=1)
csprobs = []
for cno in range(numclasses):
tmp = mnl.mnl_simulate(csdata,
csbeta[cno],
numalts,
GPU=GPU,
returnprobs=1)
tmp = np.sum(tmp * chosen, axis=1) # keep only chosen probs
csprobs.append(np.reshape(tmp, (-1, 1)))
csprobs = np.concatenate(csprobs, axis=1)
h = csprobs * cmprobs
loglik = np.sum(np.log(np.sum(h, axis=1)))
wts = h / np.reshape(np.sum(h, axis=1), (-1, 1))
return loglik, wts
def lcmnl_estimate(cmdata,numclasses,csdata,numalts,chosen,maxiter=MAXITER,emtol=EMTOL,\
skipprep=False,csbeta=None,cmbeta=None):
loglik = -999999
if csbeta is None:
csbeta = [np.random.rand(csdata.shape[1]) for i in range(numclasses)]
if not skipprep: cmdata = prep_cm_data(cmdata, numclasses)
if cmbeta is None: cmbeta = np.zeros(cmdata.shape[1])
for i in range(maxiter):
print "Running iteration %d" % (i + 1)
print time.ctime()
# EXPECTATION
print "Running class membership model"
cmprobs = mnl.mnl_simulate(cmdata,
cmbeta,
numclasses,
GPU=GPU,
returnprobs=1)
csprobs = []
for cno in range(numclasses):
tmp = mnl.mnl_simulate(csdata,
csbeta[cno],
numalts,
GPU=GPU,
returnprobs=1)
tmp = np.sum(tmp * chosen, axis=1) # keep only chosen probs
csprobs.append(np.reshape(tmp, (-1, 1)))
csprobs = np.concatenate(csprobs, axis=1)
h = csprobs * cmprobs
oldloglik = loglik
loglik = np.sum(np.log(np.sum(h, axis=1)))
print "current csbeta", csbeta
print "current cmbeta", cmbeta
print "current loglik", loglik, i + 1, "\n\n"
if abs(loglik - oldloglik) < emtol: break
wts = h / np.reshape(np.sum(h, axis=1), (-1, 1))
# MAXIMIZATION
for cno in range(numclasses):
print "Estimating class specific model for class %d" % (cno + 1)
t1 = time.time()
weights = np.reshape(wts[:, cno], (-1, 1))
print weights.shape
fit, results = mnl.mnl_estimate(csdata,
chosen,
numalts,
GPU=GPU,
weights=weights,
beta=csbeta[cno])
print "Finished in %fs" % (time.time() - t1)
csbeta[cno] = zip(*results)[0]
print "Estimating class membership model"
t1 = time.time()
fit, results = mnl.mnl_estimate(cmdata,None,numclasses,GPU=GPU,weights=wts,lcgrad=True, \
beta=cmbeta,coeffrange=(-1000,1000))
print "Finished in %fs" % (time.time() - t1)
cmbeta = zip(*results)[0]
