def monitored_training_loop(model: gpflow.models.SVGP,
logdir: str,
epochs: int = 1,
logging_epoch_freq: int = 10,
num_samples: int = 10):
summary_writer = tf.summary.create_file_writer(logdir)
tf_optimization_step = tf.function(optimization_step)
batches = iter(train_dataset)
with summary_writer.as_default():
for epoch in range(epochs):
for _ in range(num_batches_per_epoch):
tf_optimization_step(model, next(batches))
epoch_id = epoch + 1
if epoch_id % logging_epoch_freq == 0:
tf.print(f"Epoch {epoch_id}: ELBO (train) {model.elbo(data)}")
mean, var = model.predict_f(samples_input)
samples = model.predict_f_samples(samples_input, num_samples)
fig = plotting_regression(X, Y, samples_input, mean, var,
samples)
summary_matplotlib_image(dict(model_samples=fig), step=epoch)
tf.summary.scalar('elbo', data=model.elbo(data), step=epoch)
tf.summary.scalar('likelihood/variance',
data=model.likelihood.variance,
step=epoch)
tf.summary.scalar('kernel/lengthscale',
data=model.kernel.lengthscale,
step=epoch)
tf.summary.scalar('kernel/variance',
data=model.kernel.variance,
step=epoch)
def optimization_step(optimizer, model: gpflow.models.SVGP, batch):
with tf.GradientTape(watch_accessed_variables=False) as tape:
tape.watch(model.trainable_variables)
objective = -model.elbo(*batch)
grads = tape.gradient(objective, model.trainable_variables)
optimizer.apply_gradients(zip(grads, model.trainable_variables))
return objective
def step(model: gpf.models.SVGP, batch: Tuple[tf.Tensor, tf.Tensor]):
with tf.GradientTape(watch_accessed_variables=False) as tape:
tape.watch(gp.trainable_variables)
loss = model.training_loss(batch)
grads = tape.gradient(loss, gp.trainable_variables)
optimizer.apply_gradients(zip(grads, gp.trainable_variables))
return loss
def optimization_step(model: gpflow.models.SVGP, batch: Tuple[tf.Tensor,
tf.Tensor]):
with tf.GradientTape(watch_accessed_variables=False) as tape:
tape.watch(model.trainable_variables)
obj = -model.elbo(batch)
grads = tape.gradient(obj, model.trainable_variables)
optimizer.apply_gradients(zip(grads, model.trainable_variables))