From numpy to pyro (distribution shapes)

[ekb]

I am trying to recreate this log likelihood in pyro following the tutorial but I am not sure if this is correct/optimal:

This is the code in python using scipy:

def log_likelihoods(data, input, mus, sigmas):

    # data T x D
    # input T x D
    # mus K x D
    # sigmas K x D

    log_rates = norm.logpdf(input[:, None], mus[None, :], sigmas[None, :])  # T x K x D
    log_rates = log_rates + log_lambdas  # T x K x D

    lls = -gammaln(data[:, None, :] + 1) - np.exp(log_rates) + data[:, None, :] * log_rates  # T x K x D
    return np.sum(lls, axis=2)  # T x K

pyro version:

def model(data, input):

 alpha = pyro.sample("alpha", dist.Normal(0, 1).expand([K, D]))
 mu = pyro.sample("mu", dist.Normal(0, 1).expand([K, D]))
 log_sigma = pyro.sample("log_sigma", dist.Normal(0, 1).expand([K, D]))
 pdist = dist.Normal(mu, torch.exp(log_sigma)) # K x D
 pdist.log_prob(input) # where input is T x D

The last line gives me an error(see below), any suggestions?

Value is not broadcastable with batch_shape+event_shape: torch.Size([T, D]) vs torch.Size([K, D]).

thanks!

[fehiepsi]

You might try input = input.unsqueeze(1) before the last call.