def build_gradient_merge_and_update(self, global_step, losses,
device_grads):
fetches = {}
apply_gradient_devices = self.devices
gradient_state = device_grads
training_ops = []
# gradient_state is the merged gradient.
if self.mode == tf.contrib.learn.ModeKeys.TRAIN:
apply_gradient_devices, gradient_state = self.variable_mgr.preprocess_device_grads(
device_grads, self.params.independent_replica)
for d, device in enumerate(apply_gradient_devices):
with tf.device(device):
if self.mode != tf.contrib.learn.ModeKeys.INFER:
average_loss = (losses[d]
if self.params.independent_replica else
tf.reduce_sum(losses))
if self.mode == tf.contrib.learn.ModeKeys.TRAIN:
avg_grads = self.variable_mgr.get_gradients_to_apply(
d, gradient_state)
if self.mode == tf.contrib.learn.ModeKeys.TRAIN:
# add gradient clipping moved to add_forward_pass_and_gradients
self.learning_rate = tf.constant(self.params.learning_rate)
opt = tf.train.AdamOptimizer(self.learning_rate)
loss_scale_params = variable_mgr_util.AutoLossScaleParams(
enable_auto_loss_scale=False,
loss_scale=None,
loss_scale_normal_steps=None,
inc_loss_scale_every_n=1000,
is_chief=True,
)
# append optimizer operators into the graph
self.variable_mgr.append_apply_gradients_ops(
gradient_state, opt, avg_grads, training_ops,
loss_scale_params)
if self.mode == tf.contrib.learn.ModeKeys.TRAIN:
fetches["train_op"] = tf.group(training_ops)
if self.mode != tf.contrib.learn.ModeKeys.INFER:
fetches["average_loss"] = (average_loss
if self.params.independent_replica else
average_loss /
tf.to_float(self.batch_size))
return fetches
def testAppendGradientsWithLossScaleForNonChiefWorker(self):
v = tf.Variable(0)
training_ops = []
get_apply_gradients_ops_func = lambda: [tf.assign(v, v + 1)]
loss_scale_params = variable_mgr_util.AutoLossScaleParams(
enable_auto_loss_scale=True,
loss_scale=tf.Variable(4),
loss_scale_normal_steps=tf.Variable(10),
inc_loss_scale_every_n=10,
is_chief=False) # Non-chief
variable_mgr_util.append_gradients_with_loss_scale(
training_ops,
get_apply_gradients_ops_func,
loss_scale_params,
grad_has_inf_nan=False)
with self.test_session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(training_ops)
self.assertEqual(sess.run(v), 1)
self.assertEqual(sess.run(loss_scale_params.loss_scale), 4)
self.assertEqual(
sess.run(loss_scale_params.loss_scale_normal_steps), 10)
def testAppendGradientsWithLossScaleWithoutNan(self):
v = tf.Variable(0)
training_ops = []
get_apply_gradients_ops_func = lambda: [tf.assign(v, v + 1)]
loss_scale_params = variable_mgr_util.AutoLossScaleParams(
enable_auto_loss_scale=True,
loss_scale=tf.Variable(4, dtype=tf.float32),
loss_scale_normal_steps=tf.Variable(10),
inc_loss_scale_every_n=10,
is_chief=True)
variable_mgr_util.append_gradients_with_loss_scale(
training_ops,
get_apply_gradients_ops_func,
loss_scale_params,
grad_has_inf_nan=tf.constant(False))
with self.test_session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(training_ops)
self.assertEqual(sess.run(v), 1)
self.assertEqual(sess.run(loss_scale_params.loss_scale), 8)
self.assertEqual(
sess.run(loss_scale_params.loss_scale_normal_steps), 0)