ffn.py 文件源码

def createTrainingList(reviewLst):
    sds = SupervisedDataSet(100,1)
    for review in reviewLst:
        revString = unicode(review[1], errors='ignore')
        revSentences = nltk.word_tokenize(revString.strip())
        revWords = []
        for i in revSentences:
            revWords += i.lower().split()
        vec = 0
        for i in revWords:
            try:
                vec+=model[i]/2
            except:
                pass
        vec=vec/len(revWords)
        sds.addSample(vec,review[0])
    net = buildNetwork(100, 20, 1, hiddenclass=TanhLayer, outclass=SoftmaxLayer,bias=True)
    trainer = BackpropTrainer(net, sds)
    print "Error score:",trainer.train()
    print trainer.trainUntilConvergence(verbose=True,maxEpochs=100)