dataset_iter = tfe.Iterator(train_dataset)
x_batch = dataset_iter.next()
if opt_type is not 'adam':
lr.assign(
tf.train.polynomial_decay(lr_rbm,
step,
decay_steps,
lr_end,
power=0.5))
# update the variables following gradients info
with tf.GradientTape() as rbm_tape:
batch_loss_rbm = loss_rbm(x_batch)
average_loss_rbm += batch_loss_rbm
grad_rbm = rbm_tape.gradient(batch_loss_rbm, rbm.params())
optimizer_rbm.apply_gradients(zip(grad_rbm, rbm.params()))
for j in xrange(dec_per_rbm):
with tf.GradientTape() as dec_tape:
batch_loss_dec, batch_energy_neg, batch_ent_dec = loss_dec()
average_loss_dec += batch_loss_dec
average_energy_neg += batch_energy_neg
average_ent_dec += batch_ent_dec
grad_dec = dec_tape.gradient(batch_loss_dec,
dec.params() + enc_h2z.params())
optimizer_dec.apply_gradients(
zip(grad_dec,
dec.params() + enc_h2z.params()))
for j in xrange(enc_per_rbm):
def evaluate_ais_ll(ds, data_train):
_, ais_log_z, _ = estimate_log_partition_function(data_train, rbm)
ds = ds.batch(evaluation_batch_size)
ds_iter = tfe.Iterator(ds)
ais_ll = []
for batch in ds_iter:
ais_ll.append(
tf.reduce_mean(likelihood_ais(batch, rbm, ais_log_z)).numpy())
return ais_log_z.numpy(), np.mean(ais_ll)
'''
train loop
'''
saver = tf.contrib.eager.Saver(rbm.params())
average_loss = 0.
start = time.time()
ais_ll_list, ais_log_z_list = [], []
best_valid_ll, best_test_ll = -np.inf, -np.inf
for step in range(1, num_steps + 1):
# after one epoch, shuffle and refill the quene
try:
x_batch = dataset_iter.next()
except StopIteration:
dataset_iter = tfe.Iterator(train_dataset)
x_batch = dataset_iter.next()
# anneal learning rate if necessary
if step % anneal_steps == 0:
