def main():
# generate data
(train_x, train_y), (test_x, test_y) = iris()
# clear parameter store
pyro.PARAM_STORE.clear()
# learning rate
lr = 0.005
# training steps
num_steps = 2000
# SVI for inference
svi = pyro.SVI(model,
guide,
optimizer=pyro.Adam({'lr': lr}),
loss_fn=pyro.elbo)
losses, w, b = [], [], []
for step in tqdm(range(num_steps)):
loss = svi.step(train_x, train_y)
# if step % 100 == 0:
losses.append(loss)
w.append(pyro.param('w_loc').data.numpy())
b.append(pyro.param('b_loc').item())
# if step % 100 == 0:
# print('[{}] loss : {}'.format(step, loss))
w_ps = pyro.param('w_loc')
b_ps = pyro.param('b_loc')
print('w : {}; b : {}'.format(w_ps, b_ps))
def predict(x):
x = torch.tensor(x)
return torch.sigmoid((w_ps * x).sum() + b_ps)
correcto = 0
for xi, yi in zip(test_x, test_y):
num_samples = 1000
samples = sample_from_posterior(
torch.tensor(xi).view(1, -1), prob_forward, num_samples)
correcto = (((samples > 0.5) == yi.item()).sum())
print('[{}] {}/{} with {}% certainty'.format(
yi, correcto, num_samples, 100. * correcto / num_samples))
if __name__ == '__main__':
# generate data
(train_x, train_y), (test_x, test_y) = iris()
# clear parameter store
pyro.PARAM_STORE.clear()
# learning rate
lr = 0.005
# training steps
num_steps = 1000
# SVI for inference
svi = pyro.SVI(model,
guide,
optimizer=pyro.Adam({'lr': lr}),
loss_fn=pyro.elbo)
losses, w, b = [], [], []
for step in range(num_steps):
loss = svi.step(train_x, train_y)
# if step % 100 == 0:
losses.append(loss)
w.append(pyro.param('w_loc').data.numpy())
b.append(pyro.param('b_loc').item())
if step % 100 == 0:
print('[{}] loss : {}'.format(step, loss))
w_ps = pyro.param('w_loc')
b_ps = pyro.param('b_loc')