def _configure_optimizer(learning_rate):
"""Configures the optimizer used for training.
Args:
learning_rate: A scalar or `Tensor` learning rate.
Returns:
An instance of an optimizer.
Raises:
ValueError: if FLAGS.optimizer is not recognized.
"""
if FLAGS.optimizer == 'adadelta':
optimizer = tf.train.AdadeltaOptimizer(
learning_rate,
rho=FLAGS.adadelta_rho,
epsilon=FLAGS.opt_epsilon)
elif FLAGS.optimizer == 'adagrad':
optimizer = tf.train.AdagradOptimizer(
learning_rate,
initial_accumulator_value=FLAGS.adagrad_initial_accumulator_value)
elif FLAGS.optimizer == 'adam':
optimizer = tf.train.AdamOptimizer(
learning_rate,
beta1=FLAGS.adam_beta1,
beta2=FLAGS.adam_beta2,
epsilon=FLAGS.opt_epsilon)
elif FLAGS.optimizer == 'ftrl':
optimizer = tf.train.FtrlOptimizer(
learning_rate,
learning_rate_power=FLAGS.ftrl_learning_rate_power,
initial_accumulator_value=FLAGS.ftrl_initial_accumulator_value,
l1_regularization_strength=FLAGS.ftrl_l1,
l2_regularization_strength=FLAGS.ftrl_l2)
elif FLAGS.optimizer == 'momentum':
optimizer = tf.train.MomentumOptimizer(
learning_rate,
momentum=FLAGS.momentum,
name='Momentum')
elif FLAGS.optimizer == 'rmsprop':
optimizer = tf.train.RMSPropOptimizer(
learning_rate,
decay=FLAGS.rmsprop_decay,
momentum=FLAGS.rmsprop_momentum,
epsilon=FLAGS.opt_epsilon)
elif FLAGS.optimizer == 'sgd':
optimizer = tf.train.GradientDescentOptimizer(learning_rate)
elif FLAGS.optimizer == 'yellowfin':
optimizer = YFOptimizer(lr=1.0, mu=0.0)
else:
raise ValueError('Optimizer [%s] was not recognized', FLAGS.optimizer)
return optimizer
def test_measurement():
opt = YFOptimizer(zero_debias=False)
w = tf.Variable(np.ones([
n_dim,
]),
dtype=tf.float32,
name="w",
trainable=True)
b = tf.Variable(np.ones([
1,
], dtype=np.float32),
dtype=tf.float32,
name="b",
trainable=True)
x = tf.constant(np.ones([
n_dim,
], dtype=np.float32), dtype=tf.float32)
loss = tf.multiply(w, x) + b
tvars = tf.trainable_variables()
w_grad_val = tf.placeholder(tf.float32, shape=(n_dim, ))
b_grad_val = tf.placeholder(tf.float32, shape=(1, ))
apply_op = opt.apply_gradients(zip([w_grad_val, b_grad_val], tvars))
init_op = tf.global_variables_initializer()
with tf.Session() as sess:
sess.run(init_op)
target_h_max = 0.0
target_h_min = 0.0
g_norm_squared_avg = 0.0
g_norm_avg = 0.0
g_avg = 0.0
target_dist = 0.0
for i in range(n_iter):
feed_dict = {
w_grad_val: (i + 1) * np.ones([
n_dim,
], dtype=np.float32),
b_grad_val: (i + 1) * np.ones([
1,
], dtype=np.float32)
}
res = sess.run([
opt._curv_win, opt._h_max, opt._h_min, opt._grad_var,
opt._dist_to_opt_avg, apply_op
],
feed_dict=feed_dict)
g_norm_squared_avg = 0.999 * g_norm_squared_avg \
+ 0.001 * np.sum(( (i + 1)*np.ones( [n_dim + 1, ] ) )**2)
g_norm_avg = 0.999 * g_norm_avg \
+ 0.001 * np.linalg.norm( (i + 1)*np.ones( [n_dim + 1, ] ) )
g_avg = 0.999 * g_avg + 0.001 * (i + 1)
target_h_max = 0.999 * target_h_max + 0.001 * (i + 1)**2 * (n_dim +
1)
target_h_min = 0.999 * target_h_min + 0.001 * max(
1, i + 2 - 20)**2 * (n_dim + 1)
target_var = g_norm_squared_avg - g_avg**2 * (n_dim + 1)
target_dist = 0.999 * target_dist + 0.001 * g_norm_avg / g_norm_squared_avg
# print "iter ", i, " h max ", res[1], target_h_max, " h min ", res[2], target_h_min, \
# " var ", res[3], target_var, " dist ", res[4], target_dist
assert np.abs(target_h_max - res[1]) < np.abs(target_h_max) * 1e-3
assert np.abs(target_h_min - res[2]) < np.abs(target_h_min) * 1e-3
assert np.abs(target_var - res[3]) < np.abs(res[3]) * 1e-3
assert np.abs(target_dist - res[4]) < np.abs(res[4]) * 1e-3
print("sync measurement test passed!")
def test_lr_mu():
opt = YFOptimizer(zero_debias=False)
w = tf.Variable(np.ones([
n_dim,
]),
dtype=tf.float32,
name="w",
trainable=True)
b = tf.Variable(np.ones([
1,
], dtype=np.float32),
dtype=tf.float32,
name="b",
trainable=True)
x = tf.constant(np.ones([
n_dim,
], dtype=np.float32), dtype=tf.float32)
loss = tf.multiply(w, x) + b
tvars = tf.trainable_variables()
w_grad_val = tf.Variable(np.zeros([
n_dim,
]),
dtype=tf.float32,
trainable=False)
b_grad_val = tf.Variable(np.zeros([
1,
]),
dtype=tf.float32,
trainable=False)
apply_op = opt.apply_gradients(zip([w_grad_val, b_grad_val], tvars))
init_op = tf.global_variables_initializer()
with tf.Session() as sess:
sess.run(init_op)
target_h_max = 0.0
target_h_min = 0.0
g_norm_squared_avg = 0.0
g_norm_avg = 0.0
g_avg = 0.0
target_dist = 0.0
target_lr = 0.1
target_mu = 0.0
for i in range(n_iter):
sess.run(
tf.assign(w_grad_val,
(i + 1) * np.ones([
n_dim,
], dtype=np.float32)))
sess.run(
tf.assign(b_grad_val,
(i + 1) * np.ones([
1,
], dtype=np.float32)))
res = sess.run([
opt._curv_win, opt._h_max, opt._h_min, opt._grad_var,
opt._dist_to_opt_avg, opt._lr_var, opt._mu_var, apply_op
])
res[5] = opt._lr_var.eval()
res[6] = opt._mu_var.eval()
g_norm_squared_avg = 0.999 * g_norm_squared_avg \
+ 0.001 * np.sum(( (i + 1)*np.ones( [n_dim + 1, ] ) )**2)
g_norm_avg = 0.999 * g_norm_avg \
+ 0.001 * np.linalg.norm( (i + 1)*np.ones( [n_dim + 1, ] ) )
g_avg = 0.999 * g_avg + 0.001 * (i + 1)
target_h_max = 0.999 * target_h_max + 0.001 * (i + 1)**2 * (n_dim +
1)
target_h_min = 0.999 * target_h_min + 0.001 * max(
1, i + 2 - 20)**2 * (n_dim + 1)
target_var = g_norm_squared_avg - g_avg**2 * (n_dim + 1)
target_dist = 0.999 * target_dist + 0.001 * g_norm_avg / g_norm_squared_avg
if i > 0:
lr, mu = tune_everything(target_dist**2, target_var, 1,
target_h_min, target_h_max)
target_lr = 0.999 * target_lr + 0.001 * lr
target_mu = 0.999 * target_mu + 0.001 * mu
# print "iter ", i, " h max ", res[1], target_h_max, " h min ", res[2], target_h_min, \
# " var ", res[3], target_var, " dist ", res[4], target_dist
# print "iter ", i, " lr ", res[5], target_lr, " mu ", res[6], target_mu
assert np.abs(target_h_max - res[1]) < np.abs(target_h_max) * 1e-3
assert np.abs(target_h_min - res[2]) < np.abs(target_h_min) * 1e-3
assert np.abs(target_var - res[3]) < np.abs(res[3]) * 1e-3
assert np.abs(target_dist - res[4]) < np.abs(res[4]) * 1e-3
assert target_lr == 0.0 or np.abs(target_lr -
res[5]) < np.abs(res[5]) * 1e-3
assert target_mu == 0.0 or np.abs(target_mu -
res[6]) < np.abs(res[6]) * 5e-3
print("lr and mu computing test passed!")