def _batch_norm(self, x, name):
"""Batch Normalization"""
with tf.variable_scope(name):
params_shape = [x.get_shape()[-1]]
beta = tf.get_variable(
'beta', params_shape, tf.float64,
norm_init(0.0, stddev=0.1, dtype=tf.float64))
gamma = tf.get_variable('gamma', params_shape, tf.float64,
unif_init(0.1, 0.5, dtype=tf.float64))
mv_mean = tf.get_variable('moving_mean',
params_shape,
tf.float64,
const_init(0.0, tf.float64),
trainable=False)
mv_var = tf.get_variable('moving_variance',
params_shape,
tf.float64,
const_init(1.0, tf.float64),
trainable=False)
# Training Ops
mean, variance = tf.nn.moments(x, [0], name='moments')
hoge = assign_moving_average(mv_mean, mean, 0.99)
piyo = assign_moving_average(mv_var, variance, 0.99)
self._extra_train_ops.extend([hoge, piyo])
mean, variance = control_flow_ops.cond(self.is_training, lambda:
(mean, variance), lambda:
(mv_mean, mv_var))
y = tf.nn.batch_normalization(x, mean, variance, beta, gamma, 1e-6)
y.set_shape(x.get_shape())
return y
def _batch_norm(self, x, name):
""" Batch normalization """
with tf.variable_scope(name):
params_shape = [x.get_shape()[-1]]
beta = tf.get_variable(
'beta', params_shape, tf.float64,
norm_init(0.0, stddev=0.1, dtype=tf.float64))
print(self.sess.run(tf.shape(beta)))
gamma = tf.get_variable('gamma', params_shape, tf.float64,
unif_init(0.1, 0.5, dtype=tf.float64))
mv_mean = tf.get_variable('moving_mean',
params_shape,
tf.float64,
const_init(0.0, tf.float64),
trainable=False)
mv_var = tf.get_variable('moving_variance',
params_shape,
tf.float64,
const_init(1.0, tf.float64),
trainable=False)
# These ops will only be preformed when training
mean, variance = tf.nn.moments(x, [0], name='moments')
self._extra_train_ops.append(
assign_moving_average(mv_mean, mean, 0.99))
self._extra_train_ops.append(
assign_moving_average(mv_var, variance, 0.99))
mean, variance = control_flow_ops.cond(self.is_training, lambda:
(mean, variance), lambda:
(mv_mean, mv_var))
y = tf.nn.batch_normalization(x, mean, variance, beta, gamma, 1e-6)
y.set_shape(x.get_shape())
return y
def _batch_norm(self, x, name):
with tf.variable_scope(name):
params_shape = [x.get_shape()[-1]]
beta = tf.get_variable(
'beta', params_shape, tf.float64,
norm_init(0.0, stddev=0.1, dtype=tf.float64))
gamma = tf.get_variable('gamma', params_shape, tf.float64,
unif_init(.1, .5, dtype=tf.float64))
mv_mean = tf.get_variable('moving_mean',
params_shape,
tf.float64,
const_init(0, tf.float64),
trainable=False)
mv_var = tf.get_variable('moving_variance',
params_shape,
tf.float64,
const_init(1, tf.float64),
trainable=False)
mean, variance = tf.nn.moments(x, [0])
self._extra_train_ops.append(
assign_moving_average(mv_mean, mean, .99))
self._extra_train_ops.append(
assign_moving_average(mv_var, variance, .99))
mean, variance = control_flow_ops.cond(self.is_training, lambda :(mean, variance),\
lambda : (mv_mean,mv_var))
y = tf.nn.batch_normalization(x, mean, variance, beta, gamma, 1e-6)
y.set_shape(x.get_shape())
return y
def _batch_norm(self, x, name):
"""Batch normalization"""
with tf.variable_scope(name):
params_shape = [x.get_shape()[-1]]
beta = tf.get_variable(name='beta',
shape=params_shape,
dtype=tf.float64,
initializer=norm_init(mean=0.0,
stddev=0.1,
dtype=tf.float64),
trainable=True)
gamma = tf.get_variable(name='gamma',
shape=params_shape,
dtype=tf.float64,
initializer=unif_init(minval=0.1,
maxval=0.5,
dtype=tf.float64),
trainable=True)
mv_mean = tf.get_variable(name='moving_mean',
shape=params_shape,
dtype=tf.float64,
initializer=const_init(value=0.0,
dtype=tf.float64),
trainable=False)
mv_var = tf.get_variable(name='moving_variance',
shape=params_shape,
dtype=tf.float64,
initializer=const_init(value=1.0,
dtype=tf.float64),
trainable=False)
# These ops will only be performed when training:
mean, variance = tf.nn.moments(x=x, axes=[0], name='moments')
self._extra_train_ops.append(
assign_moving_average(variable=mv_mean, value=mean,
decay=0.99))
self._extra_train_ops.append(
assign_moving_average(variable=mv_var,
value=variance,
decay=0.99))
mean, variance = control_flow_ops.cond(pred=self.is_training,
true_fn=lambda:
(mean, variance),
false_fn=lambda:
(mv_mean, mv_var))
y = tf.nn.batch_normalization(x=x,
mean=mean,
variance=variance,
offset=beta,
scale=gamma,
variance_epsilon=1e-6)
y.set_shape(x.get_shape())
return y
def train(self):
start_time = time.time()
# train operations
self.global_step = tf.get_variable('global_step', [],
initializer=const_init(1),
trainable=False,
dtype=tf.int32)
trainable_vars = tf.trainable_variables()
grads = tf.gradients(self.loss, trainable_vars)
optimizer = tf.train.AdamOptimizer(self.learning_rate)
apply_op = optimizer.apply_gradients(zip(grads, trainable_vars),
global_step=self.global_step)
train_ops = [apply_op] + self._extra_train_ops
self.train_op = tf.group(*train_ops)
self.loss_history = []
self.init_history = []
dW_valid, X_valid = self.sample_path(self.valid_size)
feed_dict_valid = {
self.dW: dW_valid,
self.X: X_valid,
self.is_training: False
}
step = 1
sess = self.sess
sess.run(tf.global_variables_initializer())
temp_loss = sess.run(self.loss, feed_dict=feed_dict_valid)
temp_init = self.Y0.eval()[0]
self.loss_history.append(temp_loss)
self.init_history.append(temp_init)
if self.equation == 'Burger':
print("u(0,x) = %f" %
(self.u_fn(0, tf.constant(value=np.zeros([self.d])))))
print(" step : %5u , loss : %.4e , " % (0, temp_loss) + " Y0 : %.4e , runtime : %4u " % \
(temp_init, time.time() - start_time + self.t_bd))
for i in range(self.n_maxstep + 1):
step = sess.run(self.global_step)
dW_train, X_train = self.sample_path(self.batch_size)
sess.run(self.train_op,
feed_dict={
self.dW: dW_train,
self.X: X_train,
self.is_training: True
})
if step % self.n_displaystep == 0:
temp_loss = sess.run(self.loss, feed_dict=feed_dict_valid)
temp_init = self.Y0.eval()[0]
self.loss_history.append(temp_loss)
self.init_history.append(temp_init)
print(" step : %5u , loss : %.4e , " % \
(step, temp_loss) + \
" Y0 : %.4e , runtime : %4u " % \
(temp_init, time.time() - start_time + self.t_bd))
step += 1
end_time = time.time()
print(" running time : %.3fs " % \
(end_time - start_time + self.t_bd ))
def train(self, sess):
start_time = time.time()
# train operations
self.global_step = tf.get_variable(name='global_step',
shape=[],
initializer=const_init(
value=1, dtype=tf.int32),
trainable=False,
dtype=tf.int32)
trainable_vars = tf.trainable_variables()
grads = tf.gradients(self.loss, trainable_vars)
optimizer = tf.train.AdamOptimizer(self.learning_rate)
apply_op = optimizer.apply_gradients(zip(grads, trainable_vars),
global_step=self.global_step)
train_ops = [apply_op] + self._extra_train_ops
self.train_op = tf.group(*train_ops)
self.loss_history = []
self.init_history = []
#for validation
dS_valid, S_valid = self.sample_path_multivar(self.valid_size)
# initialization
step = 1
sess.run(tf.global_variables_initializer())
self._validate_and_print(sess, dS_valid, S_valid, step, start_time)
# begin SGD iteration
for _ in range(self.n_maxstep + 1):
step = sess.run(self.global_step)
dS_train, S_train = self.sample_path_multivar(self.batch_size)
feed_dict_train = {
self.dS: dS_train,
self.S: S_train,
self.is_training: True
}
sess.run(self.train_op, feed_dict=feed_dict_train)
if (step % self.n_displaystep == 0):
print(step)
self._validate_and_print(sess, dS_valid, S_valid, step,
start_time)
step += 1 # why do we need this?
end_time = time.time()
print("running time: %.3f s" % (end_time - start_time + self.t_bd))
def train(self):
sess = self.sess
trainable_variables = tf.trainable_variables()
self.global_step = tf.get_variable('global_step', [], initializer=const_init(1), trainable=False,
dtype=tf.int32)
learning_rate = tf.train.exponential_decay(1.0, self.global_step, decay_steps=300, decay_rate=0.5,staircase=False)
grads = tf.gradients(self.loss, trainable_variables)
optimizer = tf.train.AdamOptimizer( learning_rate=learning_rate )
apply_op = optimizer.apply_gradients( zip(grads, trainable_variables), global_step=self.global_step, name='train_step')
train_ops = [apply_op] + self._extra_train_ops
train_op = tf.group(*train_ops)
multiplier = 30
sess.run(tf.global_variables_initializer())
self.num_path = 100 * multiplier
step, loss, X_hist, W_hist, pi_hist, pi_plus_hist, dX_hist, dW_hist, \
lower_bound_hist, upper_bound_hist = sess.run([self.global_step, self.loss, \
self.X_hist, self.W_hist, self.pi_hist, self.pi_plus_hist, \
self.dX_hist, self.dW_hist, self.lower_bound_hist,
self.upper_bound_hist], feed_dict={
self.X: np.ones([self.num_path, 1]).dot(self.Xinit.reshape([1, -1])),
self.W: np.ones([self.num_path, 1]) * self.Winit,
self.pi: np.zeros([self.num_path, self.d]),
self.dZ: np.random.normal(size=[self.num_path, self.d, self.N]),
self.W_pos: np.ones([self.num_path, 1], dtype=bool),
self.Play_Opt_Policy: 1.0,
self.is_back_test: False,
self.is_training: False})
print('Playing Optimal Policy: %f' %(np.mean(sess.run(self.Utility(W_hist[-1])))))
for i in range(self.n_steps):
sess.run(train_op,feed_dict={
self.X : np.ones([self.num_path,1]).dot(self.Xinit.reshape([1,-1])),
self.W : np.ones([self.num_path,1]) * self.Winit,
self.pi : np.zeros([self.num_path,self.d]),
self.dZ : np.random.normal(size=[self.num_path,self.d,self.N]),
self.W_pos: np.ones([self.num_path,1],dtype=bool),
self.Play_Opt_Policy: 0.0,
self.is_back_test: False,
self.is_training : True})
if i % 50 == 0 or i == self.n_steps - 1:
step, loss, X_hist, W_hist, pi_hist, pi_plus_hist, dX_hist, dW_hist, \
lower_bound_hist, upper_bound_hist = sess.run([self.global_step,self.loss, \
self.X_hist, self.W_hist, self.pi_hist, self.pi_plus_hist, \
self.dX_hist, self.dW_hist, self.lower_bound_hist, self.upper_bound_hist],feed_dict={
self.X: np.ones([self.num_path, 1]).dot(self.Xinit.reshape([1,-1])),
self.W: np.ones([self.num_path, 1]) * self.Winit,
self.pi: np.zeros([self.num_path, self.d]),
self.dZ: np.random.normal(size=[self.num_path, self.d, self.N]),
self.W_pos: np.ones([self.num_path, 1], dtype=bool),
self.Play_Opt_Policy: 0.0,
self.is_back_test: False,
self.is_training : False})
print('step = %d, loss = %f' %(step, loss))
if loss < 800:
print('good result!')
self.num_path = 1
step, loss, X_hist, W_hist, pi_hist, pi_plus_hist, dX_hist, dW_hist, \
lower_bound_hist, upper_bound_hist = sess.run([self.global_step, self.loss, \
self.X_hist, self.W_hist, self.pi_hist, self.pi_plus_hist, \
self.dX_hist, self.dW_hist, self.lower_bound_hist,
self.upper_bound_hist], feed_dict={
self.X: np.ones([self.num_path, 1]).dot(self.Xinit.reshape([1, -1])),
self.W: np.ones([self.num_path, 1]) * self.Winit,
self.pi: np.zeros([self.num_path, self.d]),
self.dZ: np.expand_dims(self.back_test_dX,axis=0),
self.W_pos: np.ones([self.num_path, 1], dtype=bool),
self.Play_Opt_Policy: 0.0,
self.is_back_test: True,
self.is_training: False})
self.process_save_data(X_hist, W_hist, pi_hist, pi_plus_hist, dX_hist, lower_bound_hist,
upper_bound_hist)
print("finished")
def train(self):
sess = self.sess
trainable_variables = tf.trainable_variables()
self.global_step = tf.get_variable('global_step', [],
initializer=const_init(1),
trainable=False,
dtype=tf.int32)
learning_rate = tf.train.exponential_decay(1.0,
self.global_step,
decay_steps=500,
decay_rate=0.5,
staircase=False)
grads = tf.gradients(self.loss, trainable_variables)
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
apply_op = optimizer.apply_gradients(zip(grads, trainable_variables),
global_step=self.global_step,
name='train_step')
train_ops = [apply_op] + self._extra_train_ops
train_op = tf.group(*train_ops)
multiplier = 10
sess.run(tf.global_variables_initializer())
limit = self.limit
sess.run(self.global_step.initializer)
self.num_path = 100 * multiplier
for i in range(self.n_steps):
sess.run(train_op,
feed_dict={
self.X: np.ones([self.num_path, 1]) * self.Xinit,
self.W: np.ones([self.num_path, 1]) * self.Winit,
self.pi: np.zeros([self.num_path, 1]),
self.dZ:
np.random.normal(size=[self.num_path, self.N]),
self.W_pos: np.ones([self.num_path, 1], dtype=bool),
self.upper_limit:
np.ones([1, 1], dtype=float) * limit,
self.lower_limit:
np.ones([1, 1], dtype=float) * (-limit),
self.is_training: True
})
if (i % 50 == 0) or (i == self.n_steps - 1):
self.num_path = 10000
step, loss, X_hist, W_hist, pi_hist, pi_plus_hist, dX_hist, \
lower_bound_hist, upper_bound_hist = sess.run([self.global_step,self.loss, \
self.X_hist, self.W_hist, self.pi_hist, self.pi_plus_hist, \
self.dX_hist, self.lower_bound_hist, self.upper_bound_hist],feed_dict={
self.X : np.ones([self.num_path,1]) * self.Xinit,
self.W : np.ones([self.num_path,1]) * self.Winit,
self.pi : np.zeros([self.num_path,1]),
self.dZ: np.random.normal(size=[self.num_path, self.N]),
self.W_pos: np.ones([self.num_path, 1], dtype=bool),
self.upper_limit: np.ones([1, 1], dtype=float) * limit,
self.lower_limit: np.ones([1, 1], dtype=float) * (-limit),
self.is_training : False})
self.process_save_data(limit,
X_hist,
W_hist,
pi_hist,
pi_plus_hist,
dX_hist,
lower_bound_hist,
upper_bound_hist,
make_plot=(i == self.n_steps - 1))
print('limit = %f, step = %d, loss = %f' % (limit, step, loss))
self.num_path = 100 * multiplier
