gibbs_stuff.py

def run_gibbs(nsweeps=1):
    xs = get_dataset()
    xs, ys = xs[:1,:], xs[2,:]
    # xs is Shape([1000,2])
    # ys is Shape([1000,1]) of cluster labels
    prior = [Normal(-1, 0.5), Normal(1, 0.5)]
    for sweep in range(0, nsweeps):
        postieror = gibbs(prior, xs)
        prior = posterior # do something with that posterior?

def gibbs(prior, data):
    chain = [] # used for updating later?
    assert len(prior) == 2, "just assume 2 features in the data with 2 variables that correspond."
    # ...do a more general case later
    for xy in data:
        old_x, old_y = xy
        x_prior = prior[0]
        new_x = x_prior.sample()
        new_cond_old = None # find conditional of x|new_x?
        chain.append(("x", old_x, new_x, new_cond_old)) # for an update later?
        # repeat in y

    posterior = prior
    return posterior

make_some_data.py

def get_datapoint(sample_shape=[2]):
    """return datapoint sampled from one of two 2-dimensional gaussians"""
    cluster = Bernoulli(torch.tensor([0.5])).sample()
    mu = 5 if cluster.item() == 1 else 2
    sample = Normal(mu, 1.0).sample(sample_shape=sample_shape)
    return torch.cat((sample, cluster))


def get_dataset(n=1000):
    """get dataset of two clusters of 2-dimensional gaussians"""
    mkdata = lambda x: get_datapoint()
    return torch.stack(list(map(mkdata, range(0, n))))