def train_bayes(training_generator):
svi, model, guide = get_pyro_model(return_all=True)
loss_hist = []
for e in range(EPOCHS):
losses = []
for x_data, y_data in tqdm(training_generator):
losses.append(svi.step(x_data, y_data))
#test_model(model(x_data,y_data), model(x_data,y_data), Trace_ELBO())
loss_hist.append(np.mean(losses))
print(f"epoch {e}/{EPOCHS} :", loss_hist[-1])
trace = poutine.trace(model).get_trace(x_data, y_data)
trace.compute_log_prob(
) # optional, but allows printing of log_prob shapes
print(trace.format_shapes())
plt.plot(loss_hist)
plt.yscale('log')
plt.title("ELBO")
plt.xlabel("step")
plt.ylabel("Epoch loss")
trace_summary(svi, model, x_data, y_data)
for name, value in pyro.get_param_store().items():
print(name, pyro.param(name))
#test_model(model(x_data,y_data), model(x_data,y_data), Trace_ELBO())
loss_hist.append(np.mean(losses))
print(f"epoch {e}/{EPOCHS} :", loss_hist[-1])
trace = poutine.trace(model).get_trace(x_data, y_data)
trace.compute_log_prob() # optional, but allows printing of log_prob shapes
print(trace.format_shapes())
plt.plot(loss_hist)
plt.yscale('log')
plt.title("ELBO")
plt.xlabel("step")
plt.ylabel("Epoch loss")
df = trace_summary(svi, model, x_data, y_data)
for name, value in pyro.get_param_store().items():
print(name, pyro.param(name))
# plot the gaussian data
# g = guide(None, None)
# p = g(x_data)
# plt.figure()
# pp = p.cpu().detach().numpy()[:,0]
# yy = y_data.cpu().detach().numpy()[:,0]
# ss = y_data.cpu().detach().numpy()[:,1]
# i = np.linspace(.1, 1, 30)
# for x in x_data.cpu().detach().numpy():
# plt.plot(i, x)
# plt.plot(pp, .25*np.ones_like(pp),'o')