def predict(total_number):
hit = 0
for n in range(total_number):
target = APIREC.target_context[n][0]
change_context = APIREC.target_context[n][1]
code_context = APIREC.target_context[n][2]
h = []
for candidate in APIREC.P_list:
candidate_target = dict(target)
candidate_target['label'] = candidate[2]
_score = score(candidate_target, change_context, code_context)
q = heapq.heappush(h, (_score, candidate))
top10 = heapq.nlargest(10, h)
# print("Top 10: " + str(top10))
prediction = Mining.atmoic_change(target)
# print("Target: " + str(prediction))
for i in top10:
candidate = i[1]
if candidate == prediction:
hit += 1
# print("hit")
break
print("Final prediction accuracy: " + str(1.0 * hit / total_number))
def _predict(n):
# print("{} with {}".format(n, training))
global top
global weights
hit = 0
target = APIREC.target_context[n][0]
change_context = APIREC.target_context[n][1]
code_context = APIREC.target_context[n][2]
h = []
for candidate in APIREC.P_list:
candidate_target = dict(target)
candidate_target['label'] = candidate[2]
change_score = APIREC.change_score_transaction(candidate_target,
change_context)
code_score = APIREC.code_score_transaction(candidate_target,
code_context)
w_change = weights[0]
w_code = 1 - w_change
_score = w_change * change_score + w_code * code_score
heapq.heappush(h, (_score, candidate))
hit = [0] * 5
for i in range(5):
topX = heapq.nlargest(top[i], h)
# print("Top 10: " + str(top10))
prediction = Mining.atmoic_change(target)
# print("Target: " + str(prediction))
for item in topX:
candidate = item[1]
if candidate == prediction:
hit[i] += 1
# print("{} hit".format(i))
break
# print("return {}".format(hit))
return hit
def _predict(n, training=[0.5]):
# print("{} with {}".format(n, training))
hit = 0
target = APIREC.target_context[n][0]
change_context = APIREC.target_context[n][1]
code_context = APIREC.target_context[n][2]
h = []
for i in training:
h.append([])
for candidate in APIREC.P_list:
candidate_target = dict(target)
candidate_target['label'] = candidate[2]
change_score = APIREC.change_score_transaction(candidate_target,
change_context)
code_score = APIREC.code_score_transaction(candidate_target,
code_context)
for i in range(len(training)):
w_change = training[i]
w_code = 1 - w_change
_score = w_change * change_score + w_code * code_score
heapq.heappush(h[i], (_score, candidate))
hit = [0] * len(training)
for i in range(len(training)):
top10 = heapq.nlargest(10, h[i])
# print("Top 10: " + str(top10))
prediction = Mining.atmoic_change(target)
# print("Target: " + str(prediction))
for item in top10:
candidate = item[1]
if candidate == prediction:
hit[i] += 1
# print("{} hit".format(i))
break
# print("return {}".format(hit))
return hit
def test_doctorAI( modelFile='model.txt', seqFile='seq.txt', inputDimSize=20000, labelFile='label.txt', numClass=500, timeFile='', predictTime=False, useLogTime=True, hiddenDimSize=[200,200], batchSize=100, logEps=1e-8, mean_duration=20.0, verbose=False ): options = locals().copy() if len(timeFile) > 0: useTime = True else: useTime = False options['useTime'] = useTime models = np.load(modelFile) tparams = init_tparams(models) print 'build model ... ', if predictTime: x, t, mask, codePred, timePred = build_model(tparams, options) predict_code = theano.function(inputs=[x,t,mask], outputs=codePred, name='predict_code') predict_time = theano.function(inputs=[x,t,mask], outputs=timePred, name='predict_time') elif useTime: x, t, mask, codePred = build_model(tparams, options) predict_code = theano.function(inputs=[x,t,mask], outputs=codePred, name='predict_code') else: x, mask, codePred = build_model(tparams, options) predict_code = theano.function(inputs=[x,mask], outputs=codePred, name='predict_code') options['inputDimSize']=models['W_emb'].shape[0] options['numClass']=models['b_output'].shape[0] print 'load data ... ', testSet = load_data(seqFile, labelFile, timeFile) n_batches = int(np.ceil(float(len(testSet[0])) / float(batchSize))) print 'done' predVec = [] trueVec = [] predTimeVec = [] trueTimeVec = [] iteration = 0 for batchIndex in range(n_batches): tempX = testSet[0][batchIndex*batchSize: (batchIndex+1)*batchSize] tempY = testSet[1][batchIndex*batchSize: (batchIndex+1)*batchSize] if predictTime: tempT = testSet[2][batchIndex*batchSize: (batchIndex+1)*batchSize] x, t, mask, lengths = padMatrixWithTime(tempX, tempT, options) codeResults = predict_code(x, t, mask) timeResults = predict_time(x, t, mask) elif useTime: tempT = testSet[2][batchIndex*batchSize: (batchIndex+1)*batchSize] x, t, mask, lengths = padMatrixWithTime(tempX, tempT, options) codeResults = predict_code(x, t, mask) else: x, mask, lengths = padMatrixWithoutTime(tempX, options) codeResults = predict_code(x, mask) for i in range(codeResults.shape[1]): tensorMatrix = codeResults[:,i,:] thisY = tempY[i][1:] for timeIndex in range(lengths[i]): if len(thisY[timeIndex]) == 0: continue trueVec.append(thisY[timeIndex]) output = tensorMatrix[timeIndex] predVec.append(zip(*heapq.nlargest(30, enumerate(output), key=operator.itemgetter(1)))[0]) if predictTime: for i in range(timeResults.shape[1]): timeVec = timeResults[:,i] trueTimeVec.extend(tempT[i][1:]) for timeIndex in range(lengths[i]): predTimeVec.append(timeVec[timeIndex]) if (iteration % 10 == 0) and verbose: print 'iteration:%d/%d' % (iteration, n_batches) iteration += 1 if iteration == 10: break recall = recallTop(trueVec, predVec) print 'recall@10:%f, recall@20:%f, recall@30:%f' % (recall[0], recall[1], recall[2]) if predictTime: r_squared = calculate_r_squared(trueTimeVec, predTimeVec, options) print 'R2:%f' % r_squared
def test_doctorAI(modelFile='model.txt',
seqFile='seq.txt',
inputDimSize=20000,
labelFile='label.txt',
numClass=500,
timeFile='',
predictTime=False,
useLogTime=True,
hiddenDimSize=[200, 200],
batchSize=100,
logEps=1e-8,
mean_duration=20.0,
verbose=False,
embFile='embFile.txt',
topicSize=20,
topicFile='',
experiment='1'):
options = locals().copy()
if len(timeFile) > 0: useTime = True
else: useTime = False
options['useTime'] = useTime
if len(topicFile) > 0: useTopics = True
else: useTopics = False
options['useTopics'] = useTopics
models = np.load(modelFile)
tparams = init_tparams(models)
# print tparams.keys()
def load_embedding(infile):
Wemb = np.array(pickle.load(open(infile, 'rb'))).astype(config.floatX)
return Wemb
if len(embFile) > 0:
W_emb = load_embedding(embFile)
else:
W_emb = None
print 'build model ... ',
if useTime and useTopics:
x, t, topics, mask, codePred = build_model(tparams, options, W_emb)
predict_code = theano.function(inputs=[x, t, topics, mask],
outputs=codePred,
name='predict_code')
elif useTime:
x, t, topics, mask, codePred = build_model(tparams, options, W_emb)
predict_code = theano.function(inputs=[x, t, mask],
outputs=codePred,
name='predict_code')
elif useTopics:
x, topics, mask, codePred = build_model(tparams, options, W_emb)
predict_code = theano.function(inputs=[x, topics, mask],
outputs=codePred,
name='predict_code')
else:
x, topics, mask, codePred = build_model(tparams, options, W_emb)
predict_code = theano.function(inputs=[x, mask],
outputs=codePred,
name='predict_code')
if 'W_emb' in models.keys():
options['inputDimSize'] = models['W_emb'].shape[0]
elif W_emb is not None:
options['inputDimSize'] = W_emb.shape[0]
options['numClass'] = models['b_output'].shape[0]
print 'load data ... ',
testSet = load_data(seqFile, labelFile, timeFile, topicFile)
n_batches = int(np.ceil(float(len(testSet[0])) / float(batchSize)))
print 'done'
predVec = []
trueVec = []
predTimeVec = []
trueTimeVec = []
iteration = 0
for batchIndex in range(n_batches):
tempX = testSet[0][batchIndex * batchSize:(batchIndex + 1) * batchSize]
tempY = testSet[1][batchIndex * batchSize:(batchIndex + 1) * batchSize]
if useTime and useTopics:
tempT = testSet[2][batchIndex * batchSize:(batchIndex + 1) *
batchSize]
tempTopics = testSet[3][batchIndex * batchSize:(batchIndex + 1) *
batchSize]
x, t, topics, mask, lengths = padMatrixWithTime(
tempX, tempT, options, tempTopics)
codeResults = predict_code(x, t, topics, mask)
elif useTime:
tempT = testSet[2][batchIndex * batchSize:(batchIndex + 1) *
batchSize]
x, t, mask, lengths = padMatrixWithTime(tempX, tempT, options)
codeResults = predict_code(x, t, mask)
elif useTopics:
tempTopics = testSet[3][batchIndex * batchSize:(batchIndex + 1) *
batchSize]
x, topics, mask, lengths = padMatrixWithoutTime(
tempX, options, tempTopics) # todo
codeResults = predict_code(x, topics, mask)
else:
x, mask, lengths = padMatrixWithoutTime(tempX, options)
codeResults = predict_code(x, mask)
for i in range(codeResults.shape[1]):
tensorMatrix = codeResults[:, i, :]
thisY = tempY[i][1:]
for timeIndex in range(lengths[i]):
if len(thisY[timeIndex]) == 0: continue
trueVec.append(thisY[timeIndex])
output = tensorMatrix[timeIndex]
predVec.append(
zip(*heapq.nlargest(
30, enumerate(output), key=operator.itemgetter(1)))[0])
if (iteration % 10 == 0) and verbose:
print 'iteration:%d/%d' % (iteration, n_batches)
iteration += 1
if iteration == 10: break
# pickle.dump(trueVec,open('experiments/trueVec_exp' + str(experiment),'wr'))
# pickle.dump(predVec,open('experiments/predVec_exp' + str(experiment),'wr'))
recall = recallTop(trueVec, predVec)
# for cross validation
# pickle.dump(recall,open('experiments-cv/recall' + str(experiment),'wr'))
# for experiments
pickle.dump(recall, open('experiments/recall' + str(experiment), 'wr'))
print 'recall@10:%f, recall@20:%f, recall@30:%f' % (recall[0], recall[1],
recall[2])