Is MonteCarlo elbo and Importance Weighted elbo computation mismatched?

[pkulwj1994]

I notice the implement of MonteCarlo elbo and Importance Weighted elbo computation and find out things seems to be mismatched among two methods.

In elbo implementation from montecarlo.py I find code blocks below:

from numbers import Number
from torch.nn.functional import softmax

def elbo(q, p, sample_dim=None, batch_dim=None, alpha=0.1, beta=1.0, size_average=True, reduce=True):
    log_weights = q.log_joint(sample_dim, batch_dim, q.conditioned())    
    return (log_like(q, p, sample_dim, batch_dim, log_weights,size_average=size_average, reduce=reduce) -beta * kl(q, p, sample_dim, batch_dim, log_weights,size_average=size_average, reduce=reduce) +(beta + alpha) * ml(q, sample_dim, batch_dim, log_weights,size_average=size_average, reduce=reduce))

which seems to compute IW elbo involving computation of Importance Weight;

While in importance.py, I find shorter implementation below

from numbers import Number
from probtorch.util import log_mean_exp

def elbo(q, p, sample_dim=None, batch_dim=None, alpha=0.1,size_average=True, reduce=True):
    z = [n for n in q.sampled() if n in p]    log_pxyz = p.log_joint(sample_dim, batch_dim)    
    log_qz = q.log_joint(sample_dim, batch_dim, z)    
    log_qy = q.log_joint(sample_dim, batch_dim, q.conditioned())    
    log_pq = (log_pxyz - log_qz)    
    if sample_dim is None:        
        objective = log_pq + alpha * log_qy    
    else:        
        objective = log_mean_exp(log_pq, 0)        
        if not isinstance(log_qy, Number):            
            objective = objective + alpha * log_mean_exp(log_qy, 0)    
    if reduce:        
        objective = objective.mean() if size_average else objective.sum()    
    return objective

which seems like direct estimate of elbo based on samples from variational distribution q(z|x). So am I wrong or these two implenemtation are truely mismatched?