def cross_validate_one(idx):
    global gdata
    (M,b,trainlist,threads)=gdata
    if idx==0:
        (M2,b2)=testGrad(M,b,trainlist,idx)
        res=0
    else:
        random.shuffle(trainlist)
        trainvalborder=len(trainlist)*(threads-2)/(threads-1)
        (M2,b2)=testGrad(M,b,trainlist[:trainvalborder],idx)
        print 'MMR after unigram learning train(idx=',idx,'):',mrr(M2,b2,trainlist)
        res=mrr(M2,b2,trainlist[trainvalborder:])
        print 'MMR after unigram learning val(idx=',idx,'):',res
    return (res,M2,b2)
Exemple #2
0
def cross_validate_one(idx):
    global gdata
    (M, b, trainlist, threads) = gdata
    if idx == 0:
        (M2, b2) = testGrad(M, b, trainlist, idx)
        res = 0
    else:
        random.shuffle(trainlist)
        trainvalborder = len(trainlist) * (threads - 2) / (threads - 1)
        (M2, b2) = testGrad(M, b, trainlist[:trainvalborder], idx)
        print 'MMR after unigram learning train(idx=', idx, '):', mrr(
            M2, b2, trainlist)
        res = mrr(M2, b2, trainlist[trainvalborder:])
        print 'MMR after unigram learning val(idx=', idx, '):', res
    return (res, M2, b2)
Exemple #3
0
def trainMb(trainlist, ans1, ans0):
    """Unigram training from saved Qlist files, returns Mb weights.
    You can play with the learning constants in trainConsts() of basicgrad.py"""
    t0 = time.time()
    M = np.random.normal(0, 0.01, (GLOVELEN, GLOVELEN))
    b = -0.0001
    #    M=np.loadtxt('data/M58prop')
    #    b=np.loadtxt('data/b58prop')
    mrrs, crossmrr, M, b = cross_validate_all(M, b, trainlist)
    t1 = time.time()
    print "time spent training =", t1 - t0
    print "MRR after crossvalidation=", crossmrr

    # XXX: This has a sideeffect, setting resolutions in trainlist
    trainmrr = mrr(M, b, trainlist)
    print 'Mb MRR on train:', trainmrr
    l, alpha = trainConsts()
    results = [crossmrr, mrrs, l, alpha, trainmrr]
    return (M, b, results)
def trainMb(trainlist,ans1,ans0):
    """Unigram training from saved Qlist files, returns Mb weights.
    You can play with the learning constants in trainConsts() of basicgrad.py"""
    t0=time.time()
    M=np.random.normal(0,0.01,(GLOVELEN,GLOVELEN))
    b=-0.0001
#    M=np.loadtxt('data/M58prop')
#    b=np.loadtxt('data/b58prop')
    mrrs,crossmrr,M,b=cross_validate_all(M,b,trainlist)
    t1=time.time()
    print "time spent training =",t1-t0
    print "MRR after crossvalidation=",crossmrr

    # XXX: This has a sideeffect, setting resolutions in trainlist
    trainmrr=mrr(M,b,trainlist)
    print 'Mb MRR on train:', trainmrr
    l,alpha=trainConsts()
    results=[crossmrr,mrrs,l,alpha,trainmrr]
    return (M,b,results)