e822a02daa1cd1c42f6d2bf309d5996a21979ef1,qucumber/tests/test_grads_positive.py,,algorithmic_gradKL,#Any#Any#Any#,75

Before Change

    return NLL

def algorithmic_gradKL(nn_state,target_psi,vis):
    grad_KL={}
    for rbmType in nn_state.gradient(vis[0]):
        grad_KL[rbmType] = {}
        for pars in nn_state.gradient(vis[0])[rbmType]:
            grad_KL[rbmType][pars]=0
    Z = partition(nn_state,vis)
    for i in range(len(vis)):
        for rbmType in nn_state.gradient(vis[i]):
            for pars in nn_state.gradient(vis[i])[rbmType]:

After Change

    //    for pars in nn_state.gradient(vis[0])[rbmType]:
    //        grad_KL[rbmType][pars]=0
    Z = partition(nn_state,vis)
    grad_KL = torch.zeros(nn_state.rbm_am.num_pars,dtype=torch.double)
    for i in range(len(vis)):
        grad_KL += ((target_psi[i,0])**2)*nn_state.gradient(vis[i]) 
        grad_KL -=probability(nn_state,vis[i], Z)*nn_state.gradient(vis[i])
        //for rbmType in nn_state.gradient(vis[i]):

Instances