Online updates of guide

Misc.
1

Is it possible to do updates of the guide with new data without using the previous data?
Using the code from the Bayesian Regression Example:

N = 100  # size of toy data

def build_linear_dataset(N, w=3, p=1, noise_std=0.01):
    X = np.random.rand(N, p)
    # w = 3
    w = w * np.ones(p)
    # b = 1
    y = np.matmul(X, w) + np.repeat(1, N) + np.random.normal(0, noise_std, size=N)
    y = y.reshape(N, 1)
    X, y = torch.tensor(X).type(torch.Tensor), torch.tensor(y).type(torch.Tensor)
    data = torch.cat((X, y), 1)
    assert data.shape == (N, p + 1)
    return data

data = build_linear_dataset(N, w=3)

pyro.clear_param_store()
for j in range(num_iterations):
    # calculate the loss and take a gradient step
    loss = svi.step(data)
    if j % 100 == 0:
        print("[iteration %04d] loss: %.4f" % (j + 1, loss / float(N)))
for name in pyro.get_param_store().get_all_param_names():
    print("[%s]: %.3f" % (name, pyro.param(name).data.numpy()))

data = build_linear_dataset(N, w=1)

for j in range(num_iterations):
    # calculate the loss and take a gradient step
    loss = svi.step(data)
    if j % 100 == 0:
        print("[iteration %04d] loss: %.4f" % (j + 1, loss / float(N)))

for name in pyro.get_param_store().get_all_param_names():
    print("[%s]: %.3f" % (name, pyro.param(name).data.numpy()))

Output:

[guide_mean_weight]: 2.994
[guide_log_scale_weight]: -3.848
[guide_mean_bias]: 0.994
[guide_log_scale_bias]: -4.169

[guide_mean_weight]: 1.000
[guide_log_scale_weight]: -4.037
[guide_mean_bias]: 0.998
[guide_log_scale_bias]: -4.521

I would like to “continue” training s.t. the final weight is 2 (so the mean between 3 and 1) and not 1.

Thanks a lot for any hints.

2

it sounds like you want to do something like online bayesian updating. if so what you need to do is the following. after you’ve learned an approximate posterior q for the first dataset, you create a second model in which q becomes the prior. then, when you proceed to learn q2 on the second dataset, you will in fact be conditioning your model on both datasets, since the first dataset will be accounted for by the new prior.

2 Likes
3

Perfect, thanks a lot!