def testWeightedMovingAverage(self):
with self.test_session() as sess:
decay = 0.5
weight = tf.placeholder(tf.float32, [])
val = tf.placeholder(tf.float32, [])
wma = moving_averages.weighted_moving_average(val, decay, weight)
tf.initialize_all_variables().run()
# Get the first weighted moving average.
val_1 = 3.0
weight_1 = 4.0
wma_array = sess.run(wma, feed_dict={val: val_1, weight: weight_1})
numerator_1 = val_1 * weight_1 * (1.0 - decay)
denominator_1 = weight_1 * (1.0 - decay)
self.assertAllClose(numerator_1 / denominator_1, wma_array)
# Get the second weighted moving average.
val_2 = 11.0
weight_2 = 22.0
wma_array = sess.run(wma, feed_dict={val: val_2, weight: weight_2})
numerator_2 = numerator_1 * decay + val_2 * weight_2 * (1.0 -
decay)
denominator_2 = denominator_1 * decay + weight_2 * (1.0 - decay)
self.assertAllClose(numerator_2 / denominator_2, wma_array)
def testWeightedMovingAverageBfloat16(self):
bfloat16 = pywrap_tensorflow.TF_bfloat16_type()
with self.cached_session() as sess:
decay = 0.5
weight = array_ops.placeholder(dtypes.bfloat16, [])
val = array_ops.placeholder(dtypes.bfloat16, [])
wma = moving_averages.weighted_moving_average(val, decay, weight)
variables.global_variables_initializer().run()
# Get the first weighted moving average.
val_1 = 3.0
weight_1 = 4.0
wma_array = sess.run(wma, feed_dict={val: val_1, weight: weight_1})
numerator_1 = val_1 * weight_1 * (1.0 - decay)
denominator_1 = weight_1 * (1.0 - decay)
self.assertAllClose(numerator_1 / denominator_1, wma_array)
# Get the second weighted moving average.
val_2 = 11.0
weight_2 = 22.0
wma_array = sess.run(wma, feed_dict={val: val_2, weight: weight_2})
numerator_2 = numerator_1 * decay + val_2 * weight_2 * (1.0 - decay)
denominator_2 = denominator_1 * decay + weight_2 * (1.0 - decay)
self.assertAllClose(bfloat16(numerator_2 / denominator_2), wma_array)
def testWeightedMovingAverage(self):
with self.test_session() as sess:
decay = 0.5
weight = tf.placeholder(tf.float32, [])
val = tf.placeholder(tf.float32, [])
wma = moving_averages.weighted_moving_average(val, decay, weight)
tf.initialize_all_variables().run()
# Get the first weighted moving average.
val_1 = 3.0
weight_1 = 4.0
wma_array = sess.run(
wma, feed_dict={val: val_1, weight: weight_1})
numerator_1 = val_1 * weight_1 * (1.0 - decay)
denominator_1 = weight_1 * (1.0 - decay)
self.assertAllClose(numerator_1 / denominator_1, wma_array)
# Get the second weighted moving average.
val_2 = 11.0
weight_2 = 22.0
wma_array = sess.run(
wma, feed_dict={val: val_2, weight: weight_2})
numerator_2 = numerator_1 * decay + val_2 * weight_2 * (1.0 - decay)
denominator_2 = denominator_1 * decay + weight_2 * (1.0 - decay)
self.assertAllClose(numerator_2 / denominator_2, wma_array)
def compute_losses_multi_or(logits,
actions_one_hot,
if_reach_goal,
weights=None,
num_actions=-1,
data_loss_wt=1.,
reg_loss_wt=1.,
ewma_decay=0.99,
reg_loss_op=None,
batch_size=4):
assert (num_actions > 0), 'num_actions must be specified and must be > 0.'
with tf.name_scope('loss'):
if_reach_goal = tf.squeeze(if_reach_goal, [-1])
batch_total = tf.reduce_sum(if_reach_goal, reduction_indices=1)
total = tf.reduce_sum(if_reach_goal)
if weights is None:
weights = tf.ones_like(actions_one_hot,
dtype=tf.float32,
name='weight')
action_prob = tf.nn.softmax(logits) # (Bx40) x 6
action_prob = tf.reshape(action_prob,
shape=[batch_size, -1,
num_actions]) # B x 40 x 6
multip = tf.multiply(action_prob, actions_one_hot) # B x 40 x 6
action_prob = tf.reduce_sum(multip, reduction_indices=2) # B x 40
example_loss = -tf.math.log(tf.maximum(tf.constant(1e-4),
action_prob)) # B x 40
example_loss = tf.multiply(example_loss, if_reach_goal) # B x 40
batch_loss = tf.reduce_sum(example_loss, reduction_indices=1) / \
tf.maximum(batch_total, tf.constant(1.)) # B. .If batch_total is zero all example_loss in that batch is all zeros.
data_loss_op = tf.reduce_sum(batch_loss) / batch_size
if reg_loss_op is None:
if reg_loss_wt > 0:
reg_loss_op = tf.add_n(tf.losses.get_regularization_losses())
else:
reg_loss_op = tf.constant(0.)
if reg_loss_wt > 0:
total_loss_op = data_loss_wt * data_loss_op + reg_loss_wt * reg_loss_op
else:
total_loss_op = data_loss_wt * data_loss_op
is_correct = tf.cast(tf.greater(action_prob, 0.5, name='pred_class'),
tf.float32)
is_correct = tf.multiply(is_correct, if_reach_goal)
acc_op = tf.reduce_sum(is_correct) / total
ewma_acc_op = moving_averages.weighted_moving_average(acc_op,
ewma_decay,
weight=total,
name='ewma_acc')
acc_ops = [ewma_acc_op]
return reg_loss_op, data_loss_op, total_loss_op, acc_ops
def compute_losses_multi_or(logits,
actions_one_hot,
weights=None,
num_actions=-1,
data_loss_wt=1.,
reg_loss_wt=1.,
ewma_decay=0.99,
reg_loss_op=None):
assert (num_actions > 0), 'num_actions must be specified and must be > 0.'
with tf.name_scope('loss'):
if weights is None:
weight = tf.ones_like(actions_one_hot,
dtype=tf.float32,
name='weight')
actions_one_hot = tf.cast(
tf.reshape(actions_one_hot, [-1, num_actions],
're_actions_one_hot'), tf.float32)
weights = tf.reduce_sum(tf.reshape(weights, [-1, num_actions],
're_weight'),
reduction_indices=1)
total = tf.reduce_sum(weights)
action_prob = tf.nn.softmax(logits)
action_prob = tf.reduce_sum(tf.multiply(action_prob, actions_one_hot),
reduction_indices=1)
example_loss = -tf.log(tf.maximum(tf.constant(1e-4), action_prob))
data_loss_op = tf.reduce_sum(example_loss * weights) / total
if reg_loss_op is None:
if reg_loss_wt > 0:
reg_loss_op = tf.add_n(tf.losses.get_regularization_losses())
else:
reg_loss_op = tf.constant(0.)
if reg_loss_wt > 0:
total_loss_op = data_loss_wt * data_loss_op + reg_loss_wt * reg_loss_op
else:
total_loss_op = data_loss_wt * data_loss_op
is_correct = tf.cast(tf.greater(action_prob, 0.5, name='pred_class'),
tf.float32)
acc_op = tf.reduce_sum(is_correct * weights) / total
ewma_acc_op = moving_averages.weighted_moving_average(acc_op,
ewma_decay,
weight=total,
name='ewma_acc')
acc_ops = [ewma_acc_op]
return reg_loss_op, data_loss_op, total_loss_op, acc_ops
def compute_losses_multi_or(logits, actions_one_hot, weights=None,
num_actions=-1, data_loss_wt=1., reg_loss_wt=1.,
ewma_decay=0.99, reg_loss_op=None):
assert(num_actions > 0), 'num_actions must be specified and must be > 0.'
with tf.name_scope('loss'):
if weights is None:
weight = tf.ones_like(actions_one_hot, dtype=tf.float32, name='weight')
actions_one_hot = tf.cast(tf.reshape(actions_one_hot, [-1, num_actions],
're_actions_one_hot'), tf.float32)
weights = tf.reduce_sum(tf.reshape(weights, [-1, num_actions], 're_weight'),
reduction_indices=1)
total = tf.reduce_sum(weights)
action_prob = tf.nn.softmax(logits)
action_prob = tf.reduce_sum(tf.multiply(action_prob, actions_one_hot),
reduction_indices=1)
example_loss = -tf.log(tf.maximum(tf.constant(1e-4), action_prob))
data_loss_op = tf.reduce_sum(example_loss * weights) / total
if reg_loss_op is None:
if reg_loss_wt > 0:
reg_loss_op = tf.add_n(tf.losses.get_regularization_losses())
else:
reg_loss_op = tf.constant(0.)
if reg_loss_wt > 0:
total_loss_op = data_loss_wt*data_loss_op + reg_loss_wt*reg_loss_op
else:
total_loss_op = data_loss_wt*data_loss_op
is_correct = tf.cast(tf.greater(action_prob, 0.5, name='pred_class'), tf.float32)
acc_op = tf.reduce_sum(is_correct*weights) / total
ewma_acc_op = moving_averages.weighted_moving_average(
acc_op, ewma_decay, weight=total, name='ewma_acc')
acc_ops = [ewma_acc_op]
return reg_loss_op, data_loss_op, total_loss_op, acc_ops
