def hmm_tester(x, start, trans, emit):
"""
This is a function to test the forward-backward algorithm in hmm.py. Splits
by student and runs f-b on all steps up to n-1, then compares prediction to
nth step
Inputs
------
x : ndarray
training observation data, nx1
start : ndarray
starting probabilities, kx1
trans : ndarray
transition probabilities, kxk
emit : ndarray
emission probabilities, kxd
Returns
-------
rmse : ndarray
array of root-mean-square-error on prediction of first correct on next
question compared to actual next data point result. This is currently
not using the test data.
"""
#Initialize array of predictions, probability of correct on next question
predicts = np.zeros(numStud)
#Initialize array for rmse to compare to actual test data
rmse = np.zeros(numStud)
#Run forward-backward on first student
f,b,probF,probB,post = hmm.frwd_bkwd(observations[:idSplit[0]-1],
startP,transP,emitP)
#Predict and compute error on first student
predicts[0] = np.dot(emitP[:,2],np.dot(transP,post[-1]))
rmse[0] = np.sqrt((data[idSplit[0]-1,1]-predicts[0])**2)
#Run forward-backward on last student
f,b,probF,probB,post = hmm.frwd_bkwd(observations[idSplit[-1]:-1],
startP,transP,emitP)
#Predict and compute error on last student
predicts[-1] = np.dot(emitP[:,2],np.dot(transP,post[-1]))
rmse[-1] = np.sqrt((data[-1,1]-predicts[-1])**2)
#Run fwd-bkwd, predict, and compute error on remaining students
for k in range(numStud-2):
f,b,probF,probB,post = hmm.frwd_bkwd(
observations[idSplit[k]:idSplit[k+1]-1],
startP,transP,emitP)
predicts[k] = np.dot(emitP[:,2],np.dot(transP,post[-1]))
rmse[k] = np.sqrt((data[idSplit[k]-1,1]-predicts[k])**2)
return rmse
def test_MCSGDBW(train, test, splitIds, outPairs, sp, tp, ep,index, binList):
testLen = len(test)
#binList = [0,.1,1.1,6,15,30,100]
bins = np.array(binList)
binVec = bins#(bins[1:]-bins[:-1])/2.+bins[:-1]
observations = np.digitize(train[:,index],binList)
observations -= 1.
print np.max(observations),np.min(observations)
for k in range(testLen):
if outPairs[k,1] != 0:
print k
index0 = outPairs[k,0]
index1 = outPairs[k,1]+outPairs[k,0]
f,b,pF,pB,post = hmm.frwd_bkwd(observations[index0:index1],
sp, tp, ep)
state = np.dot(ep.T,np.dot(tp.T,post[-1]))
test[k,index] = np.dot(binVec,state)
print test[k,index]
else:
test[k,index] = 0
test[:,13] = np.digitize(test[:,index],binList)
return test