def loss(self, x, t, u_hat):
"""
Returns mse sums catching properties of 1d heat eq.
"""
#Parital derivatives
u = u_hat(x, t)
u_x = tf.gradients(u, x)[0]
u_xx = tf.gradients(u_x, x)[0]
u_t = tf.gradients(u, t)[0]
#Initial and boundary conditions
n = tf.size(x)
zeros, ones = tf.zeros([n, 1],
dtype=tf.float64), tf.ones([n, 1],
dtype=tf.float64)
input_values = tf.cast(tf.reshape(tf.linspace(0.0, 1.0, n), [-1, 1]),
dtype=tf.float64)
u_t0 = u_hat(input_values, zeros) #t=0
u_x0 = u_hat(zeros, input_values) #x=0
u_x1 = u_hat(ones, input_values) #x=1
return (mean_squared_error(u_t, u_xx) +
mean_squared_error(u_t0, tf.sin(np.pi * input_values)) +
mean_squared_error(u_x0, zeros) +
mean_squared_error(u_x1, zeros))
def eigen_ODE_loss(x0, t, x_hat, l=n * m, seed=seed):
"""
Custom loss function to be used in HeatLearner class.
( HeatLearner(custom_loss=eigen_ODE_loss) )
Uses a loop to compute f(x(t)) because of shape requirments in f transform.
x(t) output is shape (n*m, 1).
x(t=t_m) is shape (n, 1).
"""
np.random.seed(seed)
A = np.random.normal(0, 1, (n, n))
A = (A.T + A) / 2
A = tf.convert_to_tensor(A, dtype=tf.float64) #(n, n) symmetrical matrix
x = x_hat(x0, t) #First forward pass/prediction
#Loops over each (n, 1) at each time m in (n*m, 1)
Fx = []
for i in range(int(l / n)): #l/n = m
x_vec = x[n * i:n * (i + 1), 0] #x vector at time dictated by rows
x_vec = tf.reshape(x_vec, [-1, 1]) #Shape (n, 1)
#f(x(t))
xT = tf.transpose(x_vec)
m1 = tf.matmul(xT, x_vec) * A
m2 = (1 - tf.matmul(xT, tf.matmul(A, x_vec)))
fx = tf.matmul(m1 + m2, x_vec)
Fx.append(fx)
fx = tf.reshape(Fx, [-1, 1]) #Reshape to (n*m, 1)
x_t = tf.gradients(x_hat(x0, t), t)[0] #Gradient (n*m, 1)
return mean_squared_error(x_t, fx - x)
def model_fn(features, labels, mode):
tf.logging.set_verbosity(tf.logging.WARN)
model = hub.Module(IMG_ENCODER, trainable=True)
tf.logging.set_verbosity(tf.logging.INFO)
model = model(features['x'])
regularizer = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
output = tf.layers.dense(model,
VEC_SPACE_DIMENSIONS,
activation=tf.nn.relu)
output = tf.layers.dense(model,
VEC_SPACE_DIMENSIONS,
activation=tf.nn.relu)
output = tf.layers.dense(model,
VEC_SPACE_DIMENSIONS,
activation=tf.nn.tanh)
if mode == ModeKeys.TRAIN or mode == ModeKeys.EVAL:
loss = mean_squared_error(labels, output)
regularizer = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
loss = loss + 0.25 * sum(regularizer)
if mode == ModeKeys.TRAIN:
train_op = AdamOptimizer(learning_rate=0.00001).minimize(
loss=loss, global_step=get_global_step())
return EstimatorSpec(mode=mode, loss=loss, train_op=train_op)
elif mode == ModeKeys.EVAL:
eval_metric_ops = {
'accuracy': tf.metrics.mean_cosine_distance(labels, output, 0)
}
return EstimatorSpec(mode=mode,
loss=loss,
eval_metric_ops=eval_metric_ops)
elif mode == ModeKeys.PREDICT:
return tf.estimator.EstimatorSpec(mode=mode, predictions=output)
train_ds = Dataset.from_tensor_slices(
(x_train, y_train)).shuffle(n_train_samples).batch(batch_size).repeat()
test_ds = Dataset.from_tensor_slices((x_test, y_test))
for b, item in enumerate(train_ds):
print(b, item)
if b + 1 == math.ceil(n_train_samples / batch_size):
break
def create_model():
model = Sequential()
model.add(Dense(4, activation='tanh', input_dim=1))
model.add(Dense(1))
return model
eager_model = create_model()
optimizer = GradientDescentOptimizer(0.1)
for e in range(n_epochs):
for b, (x, y) in enumerate(train_ds):
with tf.GradientTape() as tape:
pred = eager_model.predict(x.numpy())
loss_value = mean_squared_error(pred, y.numpy().reshape(10, 1))
grads = tape.gradient(loss_value, eager_model.variables)
optimizer.apply_gradients([grads.eager_model.variables])
print(loss_value)
print(e)
def __init__(self,
policy,
env,
nsteps,
ent_coef=0.01,
vf_coef=0.5,
max_grad_norm=0.5,
lr=7e-4,
alpha=0.99,
epsilon=1e-5,
total_timesteps=int(80e6),
lrschedule='linear',
network='cnn',
prio_args=None):
self.prio_args = prio_args
sess = tf_util.get_session()
nenvs = self.get_active_envs(env)
nbatch = nenvs * nsteps
with tf.variable_scope('a2c_model', reuse=tf.AUTO_REUSE):
# step_model is used for sampling
step_model = policy(nenvs, 1, sess)
# train_model is used to train our network
train_model = policy(nbatch, nsteps, sess)
# our TD evaluating network
A = tf.placeholder(train_model.action.dtype, train_model.action.shape)
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
LR = tf.placeholder(tf.float32, [])
# Calculate the loss
# Total loss = Policy gradient loss - entropy * entropy coefficient + Value coefficient * value loss
# Policy loss
neglogpac = train_model.pd.neglogp(A)
# L = A(s,a) * -logpi(a|s)
pg_loss = tf.reduce_mean(ADV * neglogpac)
# Entropy is used to improve exploration by limiting the premature convergence to suboptimal policy.
entropy = tf.reduce_mean(train_model.pd.entropy())
# Value loss
vf_loss = losses.mean_squared_error(tf.squeeze(train_model.vf), R)
# TD loss
# td_loss = losses.mean_squared_error(tf.squeeze(train_model.dt), TD)
loss = pg_loss - entropy * ent_coef + vf_loss * vf_coef
"""prio model"""
with tf.variable_scope('a2c_model_prio', reuse=tf.AUTO_REUSE):
# prio_model = policy(nbatch, nsteps, sess)
prio_model = MyNN(env, nbatch, network)
P_R = tf.placeholder(tf.float32, [nbatch])
PRIO = tf.placeholder(tf.float32, [nbatch])
P_LR = tf.placeholder(tf.float32, [])
# prio_model_loss = losses.mean_squared_error(tf.squeeze(prio_model.out), P_R) # Reward
prio_model_loss = losses.mean_squared_error(tf.squeeze(prio_model.out),
PRIO) # TD Error
# Update parameters using loss
# 1. Get the model parameters
params = find_trainable_variables("a2c_model")
params_prio = find_trainable_variables("a2c_model_prio")
# 2. Calculate the gradients
grads = tf.gradients(loss, params)
prio_grads = tf.gradients(prio_model_loss, params_prio)
if max_grad_norm is not None:
# Clip the gradients (normalize)
grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
prio_grads, prio_grad_norm = tf.clip_by_global_norm(
prio_grads, max_grad_norm)
grads = list(zip(grads, params))
prio_grads = list(zip(prio_grads, params_prio))
# zip aggregate each gradient with parameters associated
# For instance zip(ABCD, xyza) => Ax, By, Cz, Da
# 3. Make op for one policy and value update step of A2C
trainer = tf.train.RMSPropOptimizer(learning_rate=LR,
decay=alpha,
epsilon=epsilon)
prio_trainer = tf.train.RMSPropOptimizer(learning_rate=P_LR,
decay=alpha,
epsilon=epsilon)
_train = trainer.apply_gradients(grads)
_prio_train = prio_trainer.apply_gradients(prio_grads)
lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
def train(obs, states, rewards, masks, actions, values):
# Here we calculate advantage A(s,a) = R + yV(s') - V(s)
# rewards = R + yV(s')
advs = rewards - values
for step in range(len(obs)):
cur_lr = lr.value()
td_map = {
train_model.X: obs,
A: actions,
ADV: advs,
R: rewards,
LR: cur_lr
}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, _train], td_map)
prio_loss = 0
if self.prio_args is not None:
prio_values = GetValuesForPrio(self.prio_args['prio_type'],
self.prio_args['prio_param'],
advs, rewards)
prio_td_map = {
prio_model.X: obs,
P_R: rewards,
P_LR: cur_lr,
PRIO: prio_values
}
prio_loss, _, p_td = sess.run(
[prio_model_loss, _prio_train, PRIO], prio_td_map)
# mb aranged as 1D-vector = [[env_1: n1, ..., n_nstep],...,[env_n_active]]
# need to take last value of each env's buffer
self.prio_score = prio_values[list(
filter(lambda x: x % nsteps == (nsteps - 1),
range(len(prio_values))))]
return policy_loss, value_loss, policy_entropy, prio_loss
self.train = train
self.train_model = train_model
self.step_model = step_model
self.prio_model = prio_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
self.save = functools.partial(tf_util.save_variables, sess=sess)
self.load = functools.partial(tf_util.load_variables, sess=sess)
tf.global_variables_initializer().run(session=sess)
def __init__(self,
policy,
env,
nsteps,
ent_coef=0.01,
vf_coef=0.5,
r0_coef=0.05,
max_grad_norm=0.5,
lr=7e-4,
alpha=0.99,
epsilon=1e-5,
total_timesteps=int(80e6),
lrschedule='linear',
head=-1,
step_placeholder=None,
train_placeholder=None,
encoded_x_1=None,
encoded_x_2=None):
sess = tf_util.get_session()
nenvs = env.num_envs
nbatch = nenvs * nsteps
self.step_placeholder = step_placeholder
self.train_placeholder = train_placeholder
self.encoded_x_1 = encoded_x_1
self.encoded_x_2 = encoded_x_2
with tf.variable_scope('aux_model' + str(head), reuse=tf.AUTO_REUSE):
step_model, self.step_placeholder, self.encoded_x_1 = policy(
nenvs,
1,
sess,
observ_placeholder=self.step_placeholder,
encoded_x=self.encoded_x_1)
train_model, self.train_placeholder, self.encoded_x_2 = policy(
nbatch,
nsteps,
sess,
observ_placeholder=self.train_placeholder,
encoded_x=self.encoded_x_2)
A = tf.placeholder(train_model.action.dtype,
train_model.action.shape)
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
LR = tf.placeholder(tf.float32, [])
neglogpac = train_model.pd.neglogp(A)
entropy = tf.reduce_mean(train_model.pd.entropy())
pg_loss = tf.reduce_mean(ADV * neglogpac)
print(train_model.vf)
print(R)
vf_loss = losses.mean_squared_error(tf.squeeze(train_model.vf), R)
#r0_loss = losses.mean_squared_error(tf.squeeze(train_model.r), R)
loss = pg_loss - entropy * ent_coef + vf_loss * vf_coef
params = find_trainable_variables('aux_model' + str(head))
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
#print("gradiants to update: ", grads)
trainer = tf.train.RMSPropOptimizer(learning_rate=LR,
decay=alpha,
epsilon=epsilon)
_train = trainer.apply_gradients(grads)
with tf.name_scope('summaries'):
a_r = tf.summary.scalar('avg_reward', tf.reduce_mean(R))
#a_p_l = tf.summary.scalar('avg_pg_loss', tf.reduce_mean(pg_loss))
#a_v_l = tf.summary.scalar('avg_vf_loss', tf.reduce_mean(vf_loss))
#a_l = tf.summary.scalar('avg_loss', tf.reduce_mean(loss))
#merged = tf.summary.merge([a_r, a_p_l, a_v_l, a_l])
merged = tf.summary.merge([a_r])
lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
def train(
obs, states, rewards, masks, actions, values
): #Policy already sampled!! We need to update the critic now.
advs = rewards - values #For a set of (s, a), we get (r0 - v0, r1 - v1, ...)
#print("advs: ", advs)
for step in range(len(obs)):
cur_lr = lr.value()
td_map = {
train_model.X: obs,
A: actions,
ADV: advs,
R: rewards,
LR: cur_lr
}
#print(td_map)
#we train out model with observed actions, advs, rewards, cur_lr
#how is values and rewards calculated though? These cannot be sampled from a single state.
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, summary, _ = sess.run(
[pg_loss, vf_loss, entropy, merged, _train], td_map)
return policy_loss, value_loss, policy_entropy, summary
self.train = train
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
self.save = functools.partial(tf_util.save_variables, sess=sess)
self.load = functools.partial(tf_util.load_variables, sess=sess)
self.train_writer = tf.summary.FileWriter('logs/aux/' + str(head),
sess.graph)
tf.global_variables_initializer().run(session=sess)
def __init__(self, policy, env, nsteps, icm,idf,
ent_coef=0.01, vf_coef=0.5, max_grad_norm=0.5, lr=7e-4,
alpha=0.99, epsilon=1e-5, total_timesteps=int(80e6), lrschedule='linear'):
sess = tf_util.get_session()
nenvs = env.num_envs
nbatch = nenvs*nsteps
self.idf=idf
print("This is Icm in Model Init function " , type(icm))
with tf.variable_scope('a2c_model', reuse=tf.AUTO_REUSE):
# step_model is used for sampling
step_model = policy(nenvs, 1, sess)
# train_model is used to train our network
train_model = policy(nbatch, nsteps, sess)
A = tf.placeholder(train_model.action.dtype, train_model.action.shape)
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
LR = tf.placeholder(tf.float32, [])
# Calculate the loss
# Total loss = Policy gradient loss - entropy * entropy coefficient + Value coefficient * value loss
# Policy loss
neglogpac = train_model.pd.neglogp(A)
# L = A(s,a) * -logpi(a|s)
pg_loss = tf.reduce_mean(ADV * neglogpac)
# Entropy is used to improve exploration by limiting the premature convergence to suboptimal policy.
entropy = tf.reduce_mean(train_model.pd.entropy())
# Value loss
vf_loss = losses.mean_squared_error(tf.squeeze(train_model.vf), R)
loss = pg_loss - entropy*ent_coef + vf_loss * vf_coef
# Update parameters using loss
# 1. Get the model parameters
params = find_trainable_variables("a2c_model")
# 2. Calculate the gradients
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
# Clip the gradients (normalize)
grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
if icm is not None :
grads = grads + icm.pred_grads_and_vars
# print("Gradients added ")
# print("independetly there shape were a2c : {} icm :{} and together {} ".format(np.shape(grads),np.shape(icm.pred_grads_and_vars),
# np.shape(grads_and_vars)))
# zip aggregate each gradient with parameters associated
# For instance zip(ABCD, xyza) => Ax, By, Cz, Da
# 3. Make op for one policy and value update step of A2C
trainer = tf.train.RMSPropOptimizer(learning_rate=LR, decay=alpha, epsilon=epsilon)
_train = trainer.apply_gradients(grads)
lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
def train(obs, states, rewards, masks, actions, values , next_obs ) :
#, icm_rewards,cumulative_dicounted_icm): #, new_rew):
# Here we calculate advantage A(s,a) = R + yV(s') - V(s)
# rewards = R + yV(s')
# print(" icm called in train function ", type(icm))
advs = rewards - values
# print("Now the advantage ", advs )
# icm_adv = icm_rewards - values
# m , s = get_mean_and_std(icm_adv)
# > adv Normaliztion
# m , s = get_mean_and_std(advs)
# advs = (advs - m) / (s + 1e-7)
# advs = (icm_adv - m) / (s + 1e-7)
# icm_adv = (icm_adv - icm_adv.mean()) / ( + 1e-7)
# print("icm advantage ", icm_adv)
# advs = new_rew - values
# print("Advantage :", advs)
# print("On train shapes are ")
# print(" obs {} states {} rewards {} masks {} actions {} values {} ".
# format(np.shape(obs) , np.shape(states) , np.shape(rewards) , np.shape(masks) ,np.shape(actions) ,
# np.shape(values) ))
# print("Received Advantage {} rewards {} values {}".format(
# advs , rewards , values) )
# print("advantage reward and values shape ")
# print("advs {} , rewards shape {} , values {}".format(np.shape(advs) , np.shape(rewards) , np.shape(values)))
for step in range(len(obs)):
cur_lr = lr.value()
if icm is None :
td_map = {train_model.X:obs, A:actions, ADV:advs, R:rewards, LR:cur_lr}
else :
# print("curiosity Td Map ")
# print(" obs {} , next obs {} , actions {} ".format(np.shape(obs) , np.shape(next_obs),
# np.shape(actions)))
td_map = {train_model.X:obs, A:actions, ADV:advs, R:rewards, LR:cur_lr ,
icm.state_:obs, icm.next_state_ : next_obs , icm.action_ : actions }# , icm.R :rewards }
if icm is None :
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, _train],
td_map
)
return policy_loss, value_loss, policy_entropy
else :
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
if self.idf :
policy_loss, value_loss, policy_entropy,forward_loss , inverse_loss , icm_loss, _ = sess.run(
[pg_loss, vf_loss, entropy, icm.forw_loss , icm.inv_loss, icm.icm_loss ,_train],
td_map)
return policy_loss, value_loss, policy_entropy,forward_loss , inverse_loss , icm_loss, advs
else :
policy_loss, value_loss, policy_entropy,forward_loss , icm_loss, _ = sess.run(
[pg_loss, vf_loss, entropy, icm.forw_loss , icm.icm_loss ,_train],
td_map)
return policy_loss, value_loss, policy_entropy,forward_loss , 0.0 , icm_loss, advs
self.train = train
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
self.save = functools.partial(tf_util.save_variables, sess=sess)
self.load = functools.partial(tf_util.load_variables, sess=sess)
tf.global_variables_initializer().run(session=sess)
# +
from tensorflow.contrib.layers import xavier_initializer
from tensorflow.losses import mean_squared_error
from tensorflow.train import AdamOptimizer
tf.reset_default_graph()
X_data = tf.placeholder(tf.float32, shape=[None, x_vals.shape[1]])
y_target = tf.placeholder(tf.float32, shape=[None, 1])
W = tf.get_variable(shape=[x_vals.shape[1], 1], name="W", initializer=xavier_initializer())
b = tf.get_variable(shape=[1, 1], name="b", initializer=xavier_initializer())
output = tf.matmul(X_data, W) - b
l2_norm = mean_squared_error(output, y_target)
# -
# $$ Loss = \max(0, 1 - \hat{y(i)} \cdot y(i)) + \alpha ||X \cdot W - b||^2 $$
loss = tf.reduce_mean(tf.maximum(0., 1. - output * y_target)) + 0.01 * l2_norm
optimizer = AdamOptimizer(0.01).minimize(loss)
# +
batch_size = 1024
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for i in range(20000):
rand_index = np.random.choice(len(X_train), size=batch_size)
rand_x = X_train[rand_index]
def __init__(self,
policy,
env,
nsteps,
ent_coef=0.01,
vf_coef=0.5,
max_grad_norm=0.5,
lr=7e-4,
alpha=0.99,
epsilon=1e-5,
total_timesteps=int(80e6),
lrschedule='linear'):
sess = tf_util.get_session()
nenvs = env.num_envs
nbatch = nenvs * nsteps
with tf.variable_scope('a2c_model', reuse=tf.AUTO_REUSE):
step_model = policy(nenvs, 1, sess)
train_model = policy(nbatch, nsteps, sess)
A = tf.placeholder(train_model.action.dtype, train_model.action.shape)
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
LR = tf.placeholder(tf.float32, [])
neglogpac = train_model.pd.neglogp(A)
entropy = tf.reduce_mean(train_model.pd.entropy())
pg_loss = tf.reduce_mean(ADV * neglogpac)
vf_loss = losses.mean_squared_error(tf.squeeze(train_model.vf), R)
loss = pg_loss - entropy * ent_coef + vf_loss * vf_coef
params = find_trainable_variables("a2c_model")
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
trainer = tf.train.RMSPropOptimizer(learning_rate=LR,
decay=alpha,
epsilon=epsilon)
_train = trainer.apply_gradients(grads)
lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
def train(obs, states, rewards, masks, actions, values):
advs = rewards - values
for step in range(len(obs)):
cur_lr = lr.value()
td_map = {
train_model.X: obs,
A: actions,
ADV: advs,
R: rewards,
LR: cur_lr
}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, _train], td_map)
return policy_loss, value_loss, policy_entropy
self.train = train
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
self.save = functools.partial(tf_util.save_variables, sess=sess)
self.load = functools.partial(tf_util.load_variables, sess=sess)
tf.global_variables_initializer().run(session=sess)
def _mse(y_true: Tensor, y_pred: Tensor) -> Tensor:
"""均方误差损失(losses.mse())
:param y_true: shape = (N, In), float32
:param y_pred: shape = (N, In), float32
:return: shape = (N,), float32"""
return tf.reduce_mean((y_true - y_pred)**2, axis=-1)
tf.random.set_seed(0)
y_true = tf.random.normal((16, 10))
y_pred = tf.random.normal((16, 10))
print(losses.mse(y_true, y_pred))
print(losses.mean_squared_error(y_true, y_pred)) # 答案同上
print(_mse(y_true, y_pred))
# tf.Tensor(
# [1.1952267 3.4243941 2.1024227 2.3010921 2.3643446 1.8302895 1.6360563
# 3.5714912 2.3740485 4.2296114 1.4224513 4.019039 0.7188259 1.5340036
# 1.5875269 2.435854 ], shape=(16,), dtype=float32)
# tf.Tensor(
# [1.1952267 3.4243941 2.1024227 2.3010921 2.3643446 1.8302895 1.6360563
# 3.5714912 2.3740485 4.2296114 1.4224513 4.019039 0.7188259 1.5340036
# 1.5875269 2.435854 ], shape=(16,), dtype=float32)
# tf.Tensor(
# [1.1952267 3.4243941 2.1024227 2.301092 2.3643446 1.8302895 1.6360562
# 3.5714912 2.3740482 4.2296114 1.4224513 4.019039 0.7188258 1.5340036
# 1.587527 2.435854 ], shape=(16,), dtype=float32)
def __init__(self,
policy,
env,
nsteps,
ent_coef=0.01,
vf_coef=0.5,
max_grad_norm=0.5,
lr=7e-4,
alpha=0.99,
epsilon=1e-5,
diverse_r_coef=0.1,
gamma=0.99,
total_timesteps=int(80e6),
lrschedule='linear'):
sess = tf_util.get_session()
nenvs = env.num_envs
nbatch = nenvs * nsteps
with tf.variable_scope('vfo_model', reuse=tf.AUTO_REUSE):
step_model = policy(nbatch=nenvs, nsteps=1, sess=sess)
train_model = policy(nbatch=nbatch, nsteps=nsteps, sess=sess)
A = tf.placeholder(train_model.action.dtype, train_model.action.shape)
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
LR = tf.placeholder(tf.float32, [])
params = find_trainable_variables('vfo_model')
print(params)
# ==============================
# model-free actor-critic loss
# ==============================
with tf.variable_scope('mf_loss'):
neglogpac = train_model.pd.neglogp(A)
entropy = tf.reduce_mean(train_model.pd.entropy())
pg_loss = tf.reduce_mean(ADV * neglogpac)
vf_loss = losses.mean_squared_error(tf.squeeze(train_model.vf), R)
loss = pg_loss - entropy * ent_coef + vf_loss * vf_coef
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
trainer = tf.train.RMSPropOptimizer(learning_rate=LR,
decay=alpha,
epsilon=epsilon)
_train = trainer.apply_gradients(grads)
# ==============================
# diverse options policy loss
# ==============================
option_train_ops = []
option_losses = []
option_losses_names = []
option_distil_train_op = None
with tf.variable_scope('options_loss'):
diversity_reward = -1 * tf.nn.softmax_cross_entropy_with_logits_v2(
labels=train_model.op_z,
logits=train_model.option_discriminator)
diversity_reward = tf.check_numerics(
diversity_reward, 'Check numerics (1): diversity_reward')
diversity_reward -= tf.log(
tf.reduce_sum(train_model.prior_op_z * train_model.op_z) +
1e-6)
print('d_reward:', diversity_reward.get_shape().as_list())
intrinsic_reward = tf.multiply(
train_model.next_pvfs - train_model.pvfs, train_model.op_z)
intrinsic_reward = tf.reduce_sum(intrinsic_reward, 1)
print('i_reward:', intrinsic_reward.get_shape().as_list())
reward = diverse_r_coef * diversity_reward + intrinsic_reward
with tf.variable_scope('critic'):
next_vf = tf.reduce_sum(
tf.multiply(train_model.next_pvfs, train_model.op_z), 1)
print('next_vf:', next_vf.get_shape().as_list())
option_q_y = tf.stop_gradient(reward +
(1 - train_model.dones) * gamma *
next_vf)
option_q = tf.squeeze(train_model.option_q, 1)
print('option_q_y:', option_q_y.get_shape().as_list())
print('option_q:', option_q.get_shape().as_list())
option_q_loss = 0.5 * tf.reduce_mean(
(option_q_y - option_q)**2)
with tf.variable_scope('actor'):
log_op_pi_t = train_model.option_pd.logp(A)
log_target_t = tf.squeeze(train_model.option_q, 1)
pvf = tf.reduce_sum(
tf.multiply(train_model.pvfs, train_model.op_z), 1)
print('op_pi:', log_op_pi_t.get_shape().as_list())
print('op_t:', log_target_t.get_shape().as_list())
print('pvf:', pvf.get_shape().as_list())
kl_surrogate_loss = tf.reduce_mean(
log_op_pi_t *
tf.stop_gradient(log_op_pi_t - log_target_t - pvf))
with tf.variable_scope('discriminator'):
print('op_z:', train_model.op_z.get_shape().as_list())
print('op_dis:',
train_model.option_discriminator.get_shape().as_list())
discriminator_loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits_v2(
labels=train_model.op_z,
logits=train_model.option_discriminator_logits))
with tf.variable_scope('distillation'):
# NOTE: to train distillation, op_z should be feed with q(z|s)
print('mf_pi:', train_model.pi.get_shape().as_list())
print('op_pi:', train_model.option_pi.get_shape().as_list())
distillation_loss = losses.mean_squared_error(
tf.stop_gradient(train_model.pi), train_model.option_pi)
_train_option_q = tf.train.AdamOptimizer(lr).minimize(
loss=option_q_loss, var_list=params)
option_train_ops.append(_train_option_q)
option_losses.append(option_q_loss)
option_losses_names.append('option_critic')
_train_option_policy = tf.train.AdamOptimizer(lr).minimize(
loss=kl_surrogate_loss, var_list=params)
option_train_ops.append(_train_option_policy)
option_losses.append(kl_surrogate_loss)
option_losses_names.append('option_actor')
_train_option_disc = tf.train.AdamOptimizer(lr).minimize(
loss=discriminator_loss, var_list=params)
option_train_ops.append(_train_option_disc)
option_losses.append(discriminator_loss)
option_losses_names.append('option_discriminator')
option_distil_train_op = tf.train.AdamOptimizer(lr).minimize(
loss=distillation_loss, var_list=params)
tf.summary.FileWriter(logger.get_dir(), sess.graph)
lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
def train(obs, states, rewards, masks, actions, values):
advs = rewards - values
for step in range(len(obs)):
cur_lr = lr.value()
td_map = {
train_model.X: obs,
A: actions,
ADV: advs,
R: rewards,
LR: cur_lr
}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, _train], td_map)
return policy_loss, value_loss, policy_entropy
def train_options(obs, next_obs, states, next_states, masks,
next_masks, actions, actions_full, dones, options_z):
feed = {
train_model.X: obs,
train_model.X_next: next_obs,
A: actions,
train_model.ac: actions_full,
train_model.dones: dones,
train_model.op_z: options_z
}
if states is not None:
feed[train_model.S] = states
feed[train_model.next_S] = next_states
feed[train_model.M] = masks
feed[train_model.next_M] = next_masks
record_loss_values = []
for name, loss, train_op in zip(option_losses_names, option_losses,
option_train_ops):
loss_value, _ = sess.run([loss, train_op], feed)
record_loss_values.append((name + '_loss', loss_value))
return record_loss_values
def distill_mf_to_options(obs, states, masks):
feed = {train_model.X: obs}
if states is not None:
feed[train_model.S] = states
feed[train_model.M] = masks
option_ensembles = sess.run(train_model.option_discriminator, feed)
feed[train_model.op_z] = option_ensembles
distillation_loss_value, _ = sess.run(
[distillation_loss, option_distil_train_op], feed)
return distillation_loss_value
self.train = train
self.train_options = train_options
self.distill_mf_to_options = distill_mf_to_options
self.train_model = train_model
self.prior_op_z = train_model.prior_op_z
self.step_model = step_model
self.step = step_model.step
self.option_step = step_model.option_step
self.option_select = step_model.option_select
self.selective_option_step = step_model.selective_option_step
self.value = step_model.value
self.proto_value = step_model.proto_value
self.initial_state = step_model.initial_state
self.save = functools.partial(tf_util.save_variables, sess=sess)
self.load = functools.partial(tf_util.load_variables, sess=sess)
tf.global_variables_initializer().run(session=sess)
def __init__(
self,
ob_size, # dimension of observation vector
act_size, # dimension of action vector
latents, # network hidden layer sizes
learning_rate=1e-5, # learning rate
activation='relu', # activation function
optimizer='adam', # optimization function
vf_coef=0.1, # vf_loss weight
ent_coef=0.01, # ent_loss weight
max_grad_norm=0.5): # how frequently the logs are printed out
sess = tf_util.get_session()
activation = tf_util.get_activation(activation)
optimizer = tf_util.get_optimizer(optimizer)
# learning_rate = tf.train.polynomial_decay(
# learning_rate=learning_rate,
# global_step=tf.train.get_or_create_global_step(),
# decay_steps=total_epoches,
# end_learning_rate=learning_rate / 10,
# )
# placeholders for use
X = tf.placeholder(tf.float32, [None, None, ob_size], 'observation')
A = tf.placeholder(tf.int32, [None], 'action')
ADV = tf.placeholder(tf.float32, [None], 'advantage')
R = tf.placeholder(tf.float32, [None], 'reward')
with tf.variable_scope('a2c'):
policy = build_policy(
observations=X,
act_size=act_size,
latents=latents,
vf_latents=latents,
activation=activation
)
# Calculate the loss
# Total loss = Policy gradient loss - entropy * entropy coefficient + Value coefficient * value loss
# Policy loss
neglogpac = policy.neglogp(A)
# L = A(s,a) * -logpi(a|s)
pg_loss = tf.reduce_mean(ADV * neglogpac)
# Entropy is used to improve exploration by limiting the premature convergence to suboptimal policy.
entropy = tf.reduce_mean(policy.entropy())
# Value loss
vf_loss = losses.mean_squared_error(R, policy.vf)
loss = pg_loss - entropy * ent_coef + vf_loss * vf_coef
# gradients and optimizer
params = tf.trainable_variables('a2c')
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
# Clip the gradients (normalize)
grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
# 3. Make op for one policy and value update step of A2C
trainer = optimizer(learning_rate=learning_rate)
_train = trainer.apply_gradients(grads)
# Add ops to save and restore all the variables.
saver = tf.train.Saver(tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='a2c'))
def step(obs):
action, value = sess.run([policy.action, policy.vf], feed_dict={
X: obs
})
return action, value
def value(obs):
return sess.run(policy.vf, feed_dict={
X: obs
})
def debug_output(obs):
"""
This function is only for debug
"""
return sess.run([policy.logits, policy.latent, policy.vf_latent], feed_dict={
X: obs
})
def train(obs, actions, rewards, values):
# Here we calculate advantage A(s,a) = R + yV(s') - V(s)
# rewards = R + yV(s')
advs = rewards - values
td_map = {X:obs, A:actions, ADV:advs, R:rewards}
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, _train],
td_map
)
return policy_loss, value_loss, policy_entropy
def save(save_path):
saver.save(sess, save_path)
print(f'Model saved to {save_path}')
def load(load_path):
saver.restore(sess, load_path)
print(f'Model restored from {load_path}')
self.train = train
self.step = step
self.value = value
self.save = save
self.load = load
# for debug
self.debug_output = debug_output
tf.global_variables_initializer().run(session=sess)
def __init__(
self,
policy,
env,
nsteps,
ent_coef=0.01 # 엔트로피계수
,
vf_coef=0.5 # 가치계수
,
max_grad_norm=0.5,
lr=7e-4,
alpha=0.99 # 벼림비 에누리
,
epsilon=1e-5 #
,
total_timesteps=int(80e6),
lrschedule='linear'):
sess = tf_util.get_session()
nenvs = env.num_envs
nbatch = nenvs * nsteps
with tf.variable_scope('a2c_model', reuse=tf.AUTO_REUSE):
# step_model 은 표집을 위해 사용한다
step_model = policy(nenvs, 1, sess)
# train_model 은 망을 벼림하기위해 사용한다
train_model = policy(nbatch, nsteps, sess)
A = tf.placeholder(train_model.action.dtype, train_model.action.shape)
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
LR = tf.placeholder(tf.float32, [])
# 손실(loss)을 계산한다
# Total loss = Policy gradient loss - entropy coefficient * entropy + Value coefficient * value loss
# 총손실 = 정책 기울기손실 - 엔트로피계수 * 엔트로피 + 가치계수 * 가치손실
# 정책 손실(Policy loss)
neglogpac = train_model.pd.neglogp(A)
# L = A(s,a) * -logpi(a|s)
pg_loss = tf.reduce_mean(ADV * neglogpac)
# 엔트로피는 부적절한(suboptimal) 정책으로 조기수렴하는 것을 제한하여 탐사를 개선하는데 사용한다.
entropy = tf.reduce_mean(train_model.pd.entropy())
# 가치 손실(Value loss)
vf_loss = losses.mean_squared_error(tf.squeeze(train_model.vf), R)
loss = pg_loss - entropy * ent_coef + vf_loss * vf_coef
# 손실(loss)을 사용하여 참여값(parameter: 가중값, 편향값)을 갱신한다
# 1. 모형의 참여값(parameter)을 가져온다
params = find_trainable_variables("a2c_model")
# 2. 기울기(gradient)를 계산한다
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
# 기울기를 제한한다: Clip the gradients ( normalize )
grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
# zip 은 참여값(parameter)에 관련된 각각의 기울기를 합산(aggregate)한다.
# 예를들어 zip(ABCD, xyza) => Ax, By, Cz, Da
grads = list(zip(grads, params))
# 3. A2C 정책과 가치 갱신 한단계에 대한 동작을 만든다.
trainer = tf.train.RMSPropOptimizer(learning_rate=LR,
decay=alpha,
epsilon=epsilon)
_train = trainer.apply_gradients(grads)
lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
def train(obs, states, rewards, masks, actions, values):
# 여기에서 강점을 계산한다: advantage A(s,a) = R + yV(s') - V(s)
# rewards = R + yV(s')
advs = rewards - values
for step in range(len(obs)):
cur_lr = lr.value()
td_map = {
train_model.X: obs,
A: actions,
ADV: advs,
R: rewards,
LR: cur_lr
}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, _train], td_map)
return policy_loss, value_loss, policy_entropy
self.train = train
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
self.save = functools.partial(tf_util.save_variables, sess=sess)
self.load = functools.partial(tf_util.load_variables, sess=sess)
tf.global_variables_initializer().run(session=sess)
def _build_net(self, s, h, scope, trainable, ent_coef=0.01, vf_coef=0.5):
# s is the state of the current market
# h is the number of hand 0-11
with tf.variable_scope(scope):
init_w = tf.random_normal_initializer(0., 0.3)
init_b = tf.constant_initializer(0.1)
net = tf.layers.dense(s,
30,
activation=tf.nn.relu,
kernel_initializer=init_w,
bias_initializer=init_b,
name='l1',
trainable=trainable)
with tf.variable_scope('tcn'):
tcndropout = tf.placeholder_with_default(0., shape=())
value_map = build_tcn(s,
tcndropout,
kernel_size=3,
num_channels=[256, 64, 32, 10])
if (modelDebug):
print("value_map shape",
value_map.shape) #value_map shape (?, 20, 10)
with tf.variable_scope('vin'):
v = value_map[:, -1, tf.
newaxis, :] # get the values of the last time step
vi_w = tf.get_variable('vi_w', [3, 1, 3],
initializer=init_w,
trainable=trainable)
for i in range(-2, -5, -1):
q = tf.pad(v, tf.constant([[0, 0], [0, 0], [1, 1]]))
q = tfnn.conv1d(q, vi_w, 1, "VALID", data_format="NCW")
#v: [?,1,1,12] vi_w:[1,3,1,3]
if (modelDebug):
print("q shape", q.shape) # q shape (?, 3, 10)
v = tf.reduce_max(q, axis=1, keepdims=True, name="v%d" % i)
v = v + value_map[:, i, tf.newaxis, :]
# print(v.shape)
with tf.variable_scope('a'):
v = v[:, 0, :] # reshape v into rank2
paddings = tf.constant([[0, 0], [3, 3]])
v = tf.pad(v, paddings, "SYMMETRIC")
h_pos = tf.one_hot(h, depth=10)
# att_v = v[:,0,h:h+7]# the attentioned value function
att_v = tf.concat([v, h_pos], 1) # concat the onehot position
if (modelDebug):
print("att_v", att_v.shape) #att_v (?, 26)
action = tf.layers.dense(att_v,
self.a_dim,
kernel_initializer=init_w,
bias_initializer=init_b,
name='a',
trainable=trainable)
action = tf.nn.softmax(action) #action (?, 3)
if (modelDebug):
print("action", action.shape)
value = tf.layers.dense(att_v,
1,
kernel_initializer=init_w,
bias_initializer=init_b,
name="v",
trainable=trainable)
if (modelDebug):
print("value :", value.shape)
a = tf.argmax(
action, axis=1
) # the optimal action selected by algorithm for inference
if (modelDebug):
print("a:", a.shape)
a_hot = tf.one_hot(
A, depth=3
) # the one_hot vector from A(place holder of explored action) for training
prob = tf.reduce_sum(tf.multiply(action, a_hot),
reduction_indices=[1])
eligibility = tf.log(prob) * (R - value)
loss = -tf.reduce_sum(eligibility)
entropy = tf.reduce_mean(tf.multiply(
tf.log(action),
action)) # the entropy term promotes exploration
if (modelDebug):
print(" tf.multiply( tf.log(action), action )",
tf.multiply(tf.log(action), action).shape)
print("entropy", entropy.shape)
loss += entropy * ent_coef
vf_loss = losses.mean_squared_error(value, R)
loss -= vf_loss * vf_coef
optimizer = tf.train.AdamOptimizer(0.01).minimize(loss)
return a, optimizer, value
def __init__(self,
policy,
ob_space,
ac_space,
nenvs,
total_timesteps,
nprocs=32,
nscripts=16,
nsteps=20,
nstack=4,
ent_coef=0.1,
vf_coef=0.5,
vf_fisher_coef=1.0,
lr=0.25,
max_grad_norm=0.001,
kfac_clip=0.001,
lrschedule='linear',
alpha=0.99,
epsilon=1e-5):
config = tf.ConfigProto(allow_soft_placement=True,
intra_op_parallelism_threads=nprocs,
inter_op_parallelism_threads=nprocs)
config.gpu_options.allow_growth = True
self.sess = sess = tf.Session(config=config)
nsml.bind(sess=sess)
#nact = ac_space.n
nbatch = nenvs * nsteps
A = tf.placeholder(tf.int32, [nbatch])
XY0 = tf.placeholder(tf.int32, [nbatch])
XY1 = tf.placeholder(tf.int32, [nbatch])
# ADV == TD_TARGET - values
ADV = tf.placeholder(tf.float32, [nbatch])
TD_TARGET = tf.placeholder(tf.float32, [nbatch])
PG_LR = tf.placeholder(tf.float32, [])
VF_LR = tf.placeholder(tf.float32, [])
self.model = step_model = policy(sess,
ob_space,
ac_space,
nenvs,
1,
nstack,
reuse=False)
self.model2 = train_model = policy(sess,
ob_space,
ac_space,
nenvs,
nsteps,
nstack,
reuse=True)
# Policy 1 : Base Action : train_model.pi label = A
script_mask = tf.concat([
tf.zeros([nscripts * nsteps, 1]),
tf.ones([(nprocs - nscripts) * nsteps, 1])
],
axis=0)
pi = train_model.pi
pac_weight = script_mask * (tf.nn.softmax(pi) - 1.0) + 1.0
pac_weight = tf.reduce_sum(pac_weight * tf.one_hot(A, depth=3), axis=1)
neglogpac = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=pi,
labels=A)
neglogpac *= tf.stop_gradient(pac_weight)
inv_A = 1.0 - tf.cast(A, tf.float32)
xy0_mask = tf.cast(A, tf.float32)
xy1_mask = tf.cast(A, tf.float32)
condition0 = tf.equal(xy0_mask, 2)
xy0_mask = tf.where(condition0, tf.ones(tf.shape(xy0_mask)), xy0_mask)
xy0_mask = 1.0 - xy0_mask
condition1 = tf.equal(xy1_mask, 2)
xy1_mask = tf.where(condition1, tf.zeros(tf.shape(xy1_mask)), xy1_mask)
# One hot representation of chosen marine.
# [batch_size, 2]
pi_xy0 = train_model.pi_xy0
pac_weight = script_mask * (tf.nn.softmax(pi_xy0) - 1.0) + 1.0
pac_weight = tf.reduce_sum(pac_weight * tf.one_hot(XY0, depth=1024),
axis=1)
logpac_xy0 = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=pi_xy0, labels=XY0)
logpac_xy0 *= tf.stop_gradient(pac_weight)
logpac_xy0 *= tf.cast(xy0_mask, tf.float32)
pi_xy1 = train_model.pi_xy1
pac_weight = script_mask * (tf.nn.softmax(pi_xy1) - 1.0) + 1.0
pac_weight = tf.reduce_sum(pac_weight * tf.one_hot(XY0, depth=1024),
axis=1)
# 1D? 2D?
logpac_xy1 = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=pi_xy1, labels=XY1)
logpac_xy1 *= tf.stop_gradient(pac_weight)
logpac_xy1 *= tf.cast(xy1_mask, tf.float32)
pg_loss = tf.reduce_mean(ADV * neglogpac)
pg_loss_xy0 = tf.reduce_mean(ADV * logpac_xy0)
pg_loss_xy1 = tf.reduce_mean(ADV * logpac_xy1)
vf_ = tf.squeeze(train_model.vf)
vf_r = tf.concat([
tf.ones([nscripts * nsteps]),
tf.zeros([(nprocs - nscripts) * nsteps])
],
axis=0) * TD_TARGET
vf_masked = vf_ * tf.squeeze(script_mask) + vf_r
#vf_mask[0:nscripts * nsteps] = R[0:nscripts * nsteps]
vf_loss = losses.mean_squared_error(vf_masked, TD_TARGET)
entropy_a = tf.reduce_mean(cat_entropy(train_model.pi))
entropy_xy0 = tf.reduce_mean(cat_entropy(train_model.pi_xy0))
entropy_xy1 = tf.reduce_mean(cat_entropy(train_model.pi_xy1))
entropy = entropy_a + entropy_xy0 + entropy_xy1
loss = pg_loss - entropy * ent_coef + vf_loss * vf_coef
params = find_trainable_variables("model")
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
grads, _ = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
trainer = tf.train.RMSPropOptimizer(learning_rate=lr,
decay=alpha,
epsilon=epsilon)
_train = trainer.apply_gradients(grads)
self.logits = logits = train_model.pi
# xy0
self.params_common = params_common = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, scope='model/common')
self.params_xy0 = params_xy0 = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES,
scope='model/xy0') + params_common
train_loss_xy0 = pg_loss_xy0 - entropy * ent_coef + vf_coef * vf_loss
self.grads_check_xy0 = grads_xy0 = tf.gradients(
train_loss_xy0, params_xy0)
if max_grad_norm is not None:
grads_xy0, _ = tf.clip_by_global_norm(grads_xy0, max_grad_norm)
grads_xy0 = list(zip(grads_xy0, params_xy0))
trainer_xy0 = tf.train.RMSPropOptimizer(learning_rate=lr,
decay=alpha,
epsilon=epsilon)
_train_xy0 = trainer_xy0.apply_gradients(grads_xy0)
# xy1
self.params_xy1 = params_xy1 = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES,
scope='model/xy1') + params_common
train_loss_xy1 = pg_loss_xy1 - entropy * ent_coef + vf_coef * vf_loss
self.grads_check_xy1 = grads_xy1 = tf.gradients(
train_loss_xy1, params_xy1)
if max_grad_norm is not None:
grads_xy1, _ = tf.clip_by_global_norm(grads_xy1, max_grad_norm)
grads_xy1 = list(zip(grads_xy1, params_xy1))
trainer_xy1 = tf.train.RMSPropOptimizer(learning_rate=lr,
decay=alpha,
epsilon=epsilon)
_train_xy1 = trainer_xy1.apply_gradients(grads_xy1)
self.lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
def train(obs, states, td_targets, masks, actions, xy0, xy1, values):
advs = td_targets - values
for step in range(len(obs)):
cur_lr = self.lr.value()
td_map = {
train_model.X: obs,
A: actions,
XY0: xy0,
XY1: xy1,
ADV: advs,
TD_TARGET: td_targets,
PG_LR: cur_lr
}
if states != []:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, _, \
policy_loss_xy0, policy_entropy_xy0, _, \
policy_loss_xy1, policy_entropy_xy1, _ = sess.run(
[pg_loss, vf_loss, entropy, _train,
pg_loss_xy0, entropy_xy0, _train_xy0,
pg_loss_xy1, entropy_xy1, _train_xy1],
td_map)
return policy_loss, value_loss, policy_entropy, \
policy_loss_xy0, policy_entropy_xy0, \
policy_loss_xy1, policy_entropy_xy1
def save(save_path):
ps = sess.run(params)
joblib.dump(ps, save_path)
def load(load_path):
loaded_params = joblib.load(load_path)
restores = []
for p, loaded_p in zip(params, loaded_params):
restores.append(p.assign(loaded_p))
sess.run(restores)
self.train = train
self.save = save
self.load = load
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
print("global_variables_initializer start")
tf.global_variables_initializer().run(session=sess)
print("global_variables_initializer complete")
def initialize(self) -> None:
keras.backend.set_session(self._session)
##############################################################################
# Q-network #
##############################################################################
self._sensor_input_tensor = tf.placeholder(
dtype=tf.float32,
shape=(None, *self._sensor_state_shape),
name="sensor_input",
)
self._heat_input_tensor = tf.placeholder(
dtype=tf.float32,
shape=(None, *self._heat_state_shape),
name="heat_input")
# self._position_input_tensor = tf.placeholder(
# dtype=tf.float32,
# shape=(None, *self._position_state_shape),
# name="position_input"
# )
# conv1 = Conv2D(
# filters=self._n_sensors // 2,
# kernel_size=(1, self._window_size),
# data_format="channels_last",
# activation="relu",
# name="conv1",
# )
# conv_output = conv1(self._sensor_input_tensor)
#
# conv2 = Conv2D(
# filters=self._n_sensors // 4,
# kernel_size=(1, self._window_size),
# data_format="channels_last",
# activation="relu",
# name="conv2",
# )
# conv_output = conv2(conv_output)
#
# flatten_conv = Flatten(name="flatten_conv")
# flattened_conv = flatten_conv(conv_output)
concat = Concatenate(name="concat")
concatenated_input = concat([
self._sensor_input_tensor,
self._heat_input_tensor,
# self._position_input_tensor,
])
hidden_dense0 = Dense(units=self._n_sensor_inputs * 8,
activation="relu",
name="hidden_dense1")
x = hidden_dense0(concatenated_input)
hidden_dense1 = Dense(units=self._n_sensor_inputs * 4,
activation="relu",
name="hidden_dense1")
x = hidden_dense1(x)
hidden_dense2 = Dense(units=self._n_sensor_inputs * 2,
activation="relu",
name="hidden_dense2")
x = hidden_dense2(x)
hidden_dense3 = Dense(units=self._n_sensor_inputs,
activation="relu",
name="hidden_dense3")
x = hidden_dense3(x)
output_layer = Dense(units=self._n_output_angles + 1,
name="action_quality")
self._actions_qualities_tensor = output_layer(x)
self._action_index_tensor = tf.argmax(self._actions_qualities_tensor,
axis=1,
name="output")
################################################################################
# Updating Q-network #
################################################################################
self._chosen_actions_tensor = tf.placeholder(dtype=tf.int32,
shape=(None, ),
name="chosen_actions")
self._rewards_tensor = tf.placeholder(dtype=tf.float32,
shape=(None, ),
name="discounted_rewards")
self._terminates_tensor = tf.placeholder(dtype=tf.float32,
shape=(None, ),
name="episode_terminated")
self._replay_next_states_qualities_tensor = tf.placeholder(
dtype=tf.float32,
shape=self._actions_qualities_tensor.shape,
name="replay_next_states_qualities")
next_state_indices = tf.stack((tf.range(
0,
tf.shape(self._rewards_tensor)[0]), self._chosen_actions_tensor),
axis=1)
responsible_qualities = tf.gather_nd(
self._replay_next_states_qualities_tensor, next_state_indices)
# noinspection PyTypeChecker
target_quality = (self._rewards_tensor +
self._terminates_tensor * responsible_qualities *
self._process_config.reward_discount_coef)
tf_range = tf.range(0,
tf.shape(self._rewards_tensor)[0],
dtype=tf.int32)
state_indices = tf.stack((tf_range, self._chosen_actions_tensor),
axis=1)
current_quality = tf.gather_nd(self._actions_qualities_tensor,
state_indices)
loss = losses.mean_squared_error(target_quality,
current_quality,
reduction=losses.Reduction.MEAN)
optimizer = tf.train.AdamOptimizer(learning_rate=0.0001)
self._update_model = optimizer.minimize(loss)
self._session.run(tf.global_variables_initializer())
self._saver = tf.train.Saver()
def __init__(self,
optimiser,
policy,
env,
nsteps,
vf_coef=0.5,
max_grad_norm=0.5,
alpha=0.99,
epsilon=1e-5,
total_timesteps=int(80e6)):
sess = tf_util.get_session()
nenvs = env.num_envs
nbatch = nenvs * nsteps
with tf.variable_scope('a2c_model', reuse=tf.AUTO_REUSE):
# step_model is used for sampling
step_model = policy(nenvs, 1, sess)
# train_model is used to train our network
train_model = policy(nbatch, nsteps, sess)
A = tf.placeholder(train_model.action.dtype, train_model.action.shape)
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
LR = tf.placeholder(tf.float32, [])
Ent_Coeff = tf.placeholder(tf.float32, []) # for Entropy
# Calculate the loss
# Total loss = Policy gradient loss - entropy * entropy coefficient + Value coefficient * value loss
# Policy loss
neglogpac = train_model.pd.neglogp(A)
# L = A(s,a) * -logpi(a|s)
pg_loss = tf.reduce_mean(ADV * neglogpac)
# Entropy is used to improve exploration by limiting the premature convergence to suboptimal policy.
entropy = tf.reduce_mean(train_model.pd.entropy())
# Value loss
vf_loss = losses.mean_squared_error(tf.squeeze(train_model.vf), R)
loss = pg_loss - entropy * Ent_Coeff + vf_loss * vf_coef
# Update parameters using loss
# 1. Get the model parameters
params = find_trainable_variables("a2c_model")
# 2. Calculate the gradients
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
# Clip the gradients (normalize)
grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
# zip aggregate each gradient with parameters associated
# For instance zip(ABCD, xyza) => Ax, By, Cz, Da
# 3. Make op for one policy and value update step of A2C
if optimiser == 'RMSProp':
trainer = tf.train.RMSPropOptimizer(learning_rate=LR,
decay=alpha,
epsilon=epsilon)
elif optimiser == 'SGD':
trainer = tf.train.GradientDescentOptimizer(learning_rate=LR)
_train = trainer.apply_gradients(grads)
#https://stackoverflow.com/a/45624533
_slot_vars = [
trainer.get_slot(var, name) for name in trainer.get_slot_names()
for var in params
]
SLOTS = [tf.placeholder(tf.float32, slot.shape) for slot in _slot_vars]
_set_slots = [var.assign(SLOTS[i]) for i, var in enumerate(_slot_vars)]
def get_opt_state():
return sess.run(_slot_vars)
def set_opt_state(state):
feed = {k: v for k, v in zip(SLOTS, state)}
return sess.run(_set_slots, feed)
def train(obs, states, rewards, masks, actions, values, ent_coeff):
# Here we calculate advantage A(s,a) = R + yV(s') - V(s)
# rewards = R + yV(s')
advs = rewards - values
td_map = {
train_model.X: obs,
A: actions,
ADV: advs,
R: rewards,
Ent_Coeff: ent_coeff,
LR: 1.0
}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, _train], td_map)
return policy_loss, value_loss, policy_entropy
# Only this bit added
def get_mean_std_neg_ll(obs, actions):
td_map = {train_model.X: obs, A: actions}
vals = sess.run(
[train_model.pd.mean, train_model.pd.std, neglogpac], td_map)
return vals
self.get_mean_std_neg_ll = get_mean_std_neg_ll
self.train = train
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
self.save = functools.partial(tf_util.save_variables, sess=sess)
self.load = functools.partial(tf_util.load_variables, sess=sess)
self.get_opt_state = get_opt_state
self.set_opt_state = set_opt_state
tf.global_variables_initializer().run(session=sess)
def __init__(self,
policy,
env,
nsteps,
ent_coef=0.01,
vf_coef=0.5,
lr=1e-3,
alpha=0.99,
epsilon=1e-5):
sess = tf.Session()
with tf.variable_scope('a2c_model', reuse=tf.AUTO_REUSE):
# step_model for sampling
step_model = policy(nenvs, 1, sess)
# train_model to train our network
train_model = policy(nbatch, nsteps, sess)
A = tf.placeholder(train_model.action.dtype, train_model.action.shape)
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
LR = tf.placeholder(tf.float32, [])
# total_loss = Policy gradient loss - entropy * entropy coeff + value coeff * value loss
# policy loss
neglogpac = train_model.pd.neglogp(A)
# L = -log(pi(a|s)) * Adv(s, a)
pg_loss = tf.reduce_mean(ADV * neglogpac)
#entropy is used to improve exploration by limiting the premature
# convergence to suboptimal policy
entropy = tf.reduce_mean(train_model.pd.entropy())
# value loss
vf_loss = losses.mean_squared_error(tf.squeeze(train_model.vf), R)
loss = pg_loss - entropy * ent_coef + vf_loss * vf_coef
# update params using loss
# 1. get model params
params = find_trainable_variables("a2c_model")
# 2. calculate the gradients
grads = tf.gradients(loss, params)
grads = list(zip(grads, params))
# 3. make op for one policy and value update step of A2C
trainer = tf.train.RMSPropOptimizer(learning_rate=LR,
decay=alpha,
epsilon=epsilon)
_train = trainer.apply_gradients(grads)
def train(obs, states, rewards, mask, actions, values):
# we calculate advantage A(s, a) = R + yV(s') - V(s)
# rewards = R + yV(s')
advs = rewards - values
#for step in range(len(obs)):
td_map = {
train_model.X: obs,
A: actions,
ADV: advs,
R: rewards,
LR: lr
}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, _train], td_map)
return policy_loss, value_loss, policy_entropy
self.train = train
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
def __init__(self,
action_dim,
state_dim,
lr=0.001,
ent_coef=0.01,
value_coef=0.5,
reward_decay=0.95,
output_graph=False):
self.action_dim = action_dim
self.state_dim = state_dim
self.lr = lr
self.ent_coef = ent_coef
self.value_coef = value_coef
# self.actor_lr = actor_lr
# self.critic_lr = critic_lr
self.gamma = reward_decay
self.output_graph = output_graph
tf.reset_default_graph()
self.sess = tf.Session()
self.sess.run(tf.global_variables_initializer())
if output_graph:
tf.summary.FileWriter("logs/", self.sess.graph)
with tf.name_scope("inputs"):
self.tf_obs = tf.placeholder(tf.float32, [None, self.state_dim],
name="observations")
self.tf_ac = tf.placeholder(tf.float32, [None, self.action_dim],
name="actions")
self.advantage = tf.placeholder(tf.float32, [
None,
],
name="advantage")
self.R = tf.placeholder(tf.float32, [
None,
], name="return")
# fc1
layer = tf.layers.dense(
inputs=self.tf_obs,
units=10,
activation=tf.nn.tanh,
kernel_initializer=tf.random_normal_initializer(mean=0,
stddev=0.3),
bias_initializer=tf.constant_initializer(0.1),
name='actor_fc1')
# fc2
mean_all_act = tf.layers.dense(
inputs=layer,
units=self.action_dim,
activation=None,
kernel_initializer=tf.random_normal_initializer(mean=0,
stddev=0.3),
bias_initializer=tf.constant_initializer(0.1),
name='actor_fc2')
logstd_act = tf.get_variable(name="logstd",
shape=[1, self.action_dim],
initializer=tf.zeros_initializer())
pdparam = tf.concat([mean_all_act, mean_all_act * 0.0 + logstd_act],
axis=1)
self.pd = DiagGaussianPd(pdparam)
self.action = self.pd.sample()
self.neglogp = self.pd.neglogp(self.action)
# for critic network, we share the first layer with actor
self.value = tf.layers.dense(
input=layer,
units=1,
activation=None,
kernel_initializer=tf.random_normal_initializer(mean=0,
stddev=0.3),
bias_initializer=tf.constant_initializer(0.1),
name="critic_fc2")
with tf.name_scope("Loss"):
# Total loss = Policy loss - entropy * ent_coef + value loss * value_coef
pg_loss = tf.reduce_mean(self.advantage * self.neglogp)
value_loss = losses.mean_squared_error(
tf.squeeze(self.value, self.R))
entropy = tf.reduce_mean(self.pd.entropy())
# total loss
loss = pg_loss - entropy * self.ent_coef + value_loss * self.value_coef
with tf.name_scope("Train"):
train_op = tf.train.AdamOptimizer(self.lr).minimize(loss)
def step(self, observation):
actions, values, neglogp = self.sess.run(
[self.action, self.value, self.neglogp],
feed_dict={self.tf_obs: observation[np.newaxis, :]})
return actions, values, neglogp
def learn(self, obs, actions, rewards, values):
# calculate adv = reward - V(s)
# reward = r + yV(s')
advs = rewards - values
# value = self.sess.run(self.value, feed_dict={self.obs: state})
td_map = {
self.tf_obs: obs,
self.tf_ac: actions,
self.advantage: advs,
self.R: rewards
}
policy_loss, value_loss, policy_entropy, _ = self.sess.run(
[pg_loss, value_loss, entropy, train_op], feed_dict=td_map)
return policy_loss, value_loss, policy_entropy
self.step = step
self.learn = learn
def __init__(self,
network,
env,
*,
seed=None,
nsteps=5,
total_timesteps=int(80e6),
vf_coef=0.5,
ent_coef=0.5,
max_grad_norm=0.5,
lr=1e-5,
lrschedule='constant',
gamma=0.99,
alpha=0.99,
epsilon=1e-5,
model_save_path=None,
tb_log_path=None):
"""
Main entrypoint for A2C algorithm. Train a policy with given network architecture on a given environment using a2c algorithm.
Parameters:
-----------
:param network: policy network architecture. Either string (mlp, lstm, lnlstm, cnn_lstm, cnn, cnn_small, conv_only - see policies.py.py for full list)
specifying the standard network architecture, or a function that takes tensorflow tensor as input and returns
tuple (output_tensor, extra_feed) where output tensor is the last network layer output, extra_feed is None for feed-forward
neural nets, and extra_feed is a dictionary describing how to feed state into the network for recurrent neural nets.
See policies.py.py for more details on using recurrent nets in policies.py
:param env: RL environment. Should implement interface similar to VecEnv (baselines.common/vec_env) or be wrapped with DummyVecEnv (baselines.common/vec_env/dummy_vec_env.py)
:param seed: seed to make random number sequence in the alorightm reproducible. By default is None which means seed from system noise generator (not reproducible)
:param nsteps: int, number of steps of the vectorized environment per update (i.e. batch size is nsteps * nenv where
nenv is number of environment copies simulated in parallel)
:param total_timesteps: int, total number of timesteps to train on (default: 80M)
:param vf_coef: float, coefficient in front of value function loss in the total loss function (default: 0.5)
:param ent_coef: float, coeffictiant in front of the policy entropy in the total loss function (default: 0.01)
:param max_grad_norm: float, gradient is clipped to have global L2 norm no more than this value (default: 0.5)
:param lr: float, learning rate for RMSProp (current implementation has RMSProp hardcoded in) (default: 7e-4)
:param lrschedule: schedule of learning rate. Can be 'linear', 'constant', or a function [0..1] -> [0..1] that takes fraction of
the training progress as input and returns fraction of the learning rate (specified as lr) as output
:param epsilon: float, RMSProp epsilon (stabilizes square root computation in denominator of RMSProp update) (default: 1e-5)
:param alpha: float, RMSProp decay parameter (default: 0.99)
:param gamma: float, reward discounting parameter (default: 0.99)
:param model_save_path: str, the location to save model parameters (if None, auto saving)
:param tb_log_path: str, the log location for tensorboard (if None, no logging)
"""
self.policy = build_policy(network)
self.env = env
self.nenvs = env.num_envs
self.nsteps = nsteps
nbatch = self.nenvs * nsteps
self.seed = seed
self.ent_coef = ent_coef
self.vf_coef = vf_coef
self.max_grad_norm = max_grad_norm
self.lr = lr
self.gamma = gamma
self.alpha = alpha
self.epsilon = epsilon
self.total_timesteps = total_timesteps
self.lrschedule = lrschedule
self.model_save_path = model_save_path
self.tb_log_path = None # tb_log_path
self.sess = get_session()
self.graph = self.sess.graph
self.episode_reward = np.zeros((self.nenvs, ))
with tf.variable_scope('a2c_model', reuse=tf.AUTO_REUSE):
# step_model is used for sampling
self.step_model = self.policy(self.sess,
env.observation_space,
env.action_space,
self.nenvs,
1,
self.nenvs,
reuse=False)
# train_model is used to train our network
self.train_model = self.policy(self.sess,
env.observation_space,
env.action_space,
self.nenvs,
self.nsteps,
nbatch,
reuse=True)
with tf.variable_scope('loss', reuse=False):
self.action_ph = tf.placeholder(self.train_model.action.dtype,
self.train_model.action.shape)
self.adv_ph = tf.placeholder(tf.float32, [nbatch])
self.reward_ph = tf.placeholder(tf.float32, [nbatch])
self.lr_ph = tf.placeholder(tf.float32, [])
# Calculate the loss
# Total loss = Policy gradient loss - entropy * entropy coefficient + Value coefficient * value loss
# Policy loss
neglogpac = self.train_model.proba_distribution.neglogp(
self.action_ph)
# L = A(s,a) * -logpi(a|s)
self.pg_loss = tf.reduce_mean(self.adv_ph * neglogpac)
# Entropy is used to improve exploration by limiting the premature convergence to suboptimal policy.
self.entropy = tf.reduce_mean(
self.train_model.proba_distribution.entropy())
# Value loss
self.vf_loss = losses.mean_squared_error(
tf.squeeze(self.train_model.value_fn), self.reward_ph)
self.reg_loss = tf.contrib.layers.apply_regularization(
tf.contrib.layers.l2_regularizer(0.8),
tf.trainable_variables())
self.loss = self.pg_loss - self.entropy * ent_coef + self.vf_loss * vf_coef + self.reg_loss
tf.summary.scalar('lr', self.lr_ph)
tf.summary.scalar('pg_loss', self.pg_loss)
tf.summary.scalar('entropy', self.entropy)
tf.summary.scalar('vf_loss', self.vf_loss)
tf.summary.scalar('loss', self.loss)
tf.summary.histogram('obs', self.train_model.obs_ph)
# Update parameters using loss
# 1. Get the model parameters
params = tf.trainable_variables("a2c_model")
# 2. Calculate the gradients
self.grads = grads = tf.gradients(self.loss, params)
if max_grad_norm is not None:
# Clip the gradients (normalize)
grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
# 3. Make op for one policy and value update step of A2C
trainer = tf.train.RMSPropOptimizer(learning_rate=self.lr_ph,
decay=alpha,
epsilon=epsilon)
self.apply_backprop = trainer.apply_gradients(grads)
self.lr_schedule = Scheduler(initial_value=lr,
n_values=total_timesteps,
schedule=lrschedule)
self.step = self.step_model.step
self.value = self.step_model.value
self.initial_state = self.step_model.initial_state
self.def_path_pre = os.path.dirname(
os.path.abspath(__file__)) + '/tmp/' # default path prefix
self.summary = tf.summary.merge_all()
tf.global_variables_initializer().run(session=self.sess)
def __init__(self, policy, env, nsteps,
ent_coef=0.01, vf_coef=0.5, max_grad_norm=0.5, lr=7e-4,
alpha=0.99, epsilon=1e-5, total_timesteps=int(80e6), lrschedule='linear'):
sess = tf_util.get_session()
nenvs = env.num_envs
nbatch = nenvs*nsteps
with tf.variable_scope('a2c_model', reuse=tf.AUTO_REUSE):
# step_model is used for sampling
step_model = policy(nenvs, 1, sess)
# train_model is used to train our network
train_model = policy(nbatch, nsteps, sess)
A = tf.placeholder(train_model.action.dtype, train_model.action.shape)
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
LR = tf.placeholder(tf.float32, [])
# Calculate the loss
# Total loss = Policy gradient loss - entropy * entropy coefficient + Value coefficient * value loss
# Policy loss
neglogpac = train_model.pd.neglogp(A)
# L = A(s,a) * -logpi(a|s)
pg_loss = tf.reduce_mean(ADV * neglogpac)
# Entropy is used to improve exploration by limiting the premature convergence to suboptimal policy.
entropy = tf.reduce_mean(train_model.pd.entropy())
# Value loss
vf_loss = losses.mean_squared_error(tf.squeeze(train_model.vf), R)
loss = pg_loss - entropy*ent_coef + vf_loss * vf_coef
# Update parameters using loss
# 1. Get the model parameters
params = find_trainable_variables("a2c_model")
# 2. Calculate the gradients
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
# Clip the gradients (normalize)
grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
# zip aggregate each gradient with parameters associated
# For instance zip(ABCD, xyza) => Ax, By, Cz, Da
# 3. Make op for one policy and value update step of A2C
trainer = tf.train.RMSPropOptimizer(learning_rate=LR, decay=alpha, epsilon=epsilon)
_train = trainer.apply_gradients(grads)
lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
def train(obs, states, rewards, masks, actions, values):
# Here we calculate advantage A(s,a) = R + yV(s') - V(s)
# rewards = R + yV(s')
advs = rewards - values
for step in range(len(obs)):
cur_lr = lr.value()
td_map = {train_model.X:obs, A:actions, ADV:advs, R:rewards, LR:cur_lr}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
policy_loss, value_loss, policy_entropy, _ = sess.run(
[pg_loss, vf_loss, entropy, _train],
td_map
)
return policy_loss, value_loss, policy_entropy
self.train = train
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
self.save = functools.partial(tf_util.save_variables, sess=sess)
self.load = functools.partial(tf_util.load_variables, sess=sess)
tf.global_variables_initializer().run(session=sess)
def __init__(self,
policy,
env,
nsteps,
dropoutpi_keep_prob,
dropoutpi_keep_prob_value,
dropoutvf_keep_prob,
dropoutvf_keep_prob_value,
isbnpitrainmode,
isbnvftrainmode,
l1regpi,
l2regpi,
l1regvf,
l2regvf,
wclippi,
wclipvf,
ent_coef=0.01,
vf_coef=0.5,
max_grad_norm=0.5,
lr=7e-4,
alpha=0.99,
epsilon=1e-5,
total_timesteps=int(80e6),
lrschedule='linear',
regnologstd=False,
regonlylogstd=False):
sess = tf_util.get_session()
nenvs = env.num_envs
nbatch = nenvs * nsteps
with tf.variable_scope('a2c_model', reuse=tf.AUTO_REUSE):
# step_model is used for sampling
step_model = policy(nenvs, 1, sess)
# train_model is used to train our network
train_model = policy(nbatch, nsteps, sess)
A = tf.placeholder(train_model.action.dtype, train_model.action.shape)
ADV = tf.placeholder(tf.float32, [nbatch])
R = tf.placeholder(tf.float32, [nbatch])
LR = tf.placeholder(tf.float32, [])
self.dropoutpi_keep_prob = dropoutpi_keep_prob
self.dropoutpi_keep_prob_value = dropoutpi_keep_prob_value
self.dropoutvf_keep_prob = dropoutvf_keep_prob
self.dropoutvf_keep_prob_value = dropoutvf_keep_prob_value
self.isbnpitrainmode = isbnpitrainmode
self.isbnvftrainmode = isbnvftrainmode
#REGULARIZATION
self.toregularizepi = l1regpi > 0 or l2regpi > 0
self.toregularizevf = l1regvf > 0 or l2regvf > 0
self.toweightclippi = wclippi > 0
self.toweightclipvf = wclipvf > 0
# Calculate the loss
# Total loss = Policy gradient loss - entropy * entropy coefficient + Value coefficient * value loss
# Policy loss
neglogpac = train_model.pd.neglogp(A)
# L = A(s,a) * -logpi(a|s)
pg_loss = tf.reduce_mean(ADV * neglogpac)
# Entropy is used to improve exploration by limiting the premature convergence to suboptimal policy.
entropy = tf.reduce_mean(train_model.pd.entropy())
# Value loss
vf_loss = losses.mean_squared_error(tf.squeeze(train_model.vf), R)
loss = pg_loss - entropy * ent_coef + vf_loss * vf_coef
if self.toregularizepi:
print("Regularizing policy network: L1 = {}, L2 = {}".format(
l1regpi, l2regpi))
regularizerpi = tf.contrib.layers.l1_l2_regularizer(
scale_l1=l1regpi, scale_l2=l2regpi, scope='a2c_model/pi')
all_trainable_weights_pi = find_trainable_variables('a2c_model/pi')
regularization_penalty_pi = tf.contrib.layers.apply_regularization(
regularizerpi, all_trainable_weights_pi)
loss = loss + regularization_penalty_pi
if self.toregularizevf:
print("Regularizing value network: L1 = {}, L2 = {}".format(
l1regvf, l2regvf))
regularizervf = tf.contrib.layers.l1_l2_regularizer(
scale_l1=l1regvf, scale_l2=l2regvf, scope='a2c_model/vf')
all_trainable_weights_vf = find_trainable_variables('a2c_model/vf')
regularization_penalty_vf = tf.contrib.layers.apply_regularization(
regularizervf, all_trainable_weights_vf)
loss = loss + regularization_penalty_vf
# Update parameters using loss
# 1. Get the model parameters
params = find_trainable_variables("a2c_model")
# 2. Calculate the gradients
grads = tf.gradients(loss, params)
if max_grad_norm is not None:
# Clip the gradients (normalize)
grads, grad_norm = tf.clip_by_global_norm(grads, max_grad_norm)
grads = list(zip(grads, params))
# zip aggregate each gradient with parameters associated
# For instance zip(ABCD, xyza) => Ax, By, Cz, Da
# 3. Make op for one policy and value update step of A2C
trainer = tf.train.RMSPropOptimizer(learning_rate=LR,
decay=alpha,
epsilon=epsilon)
_train = trainer.apply_gradients(grads)
if self.toweightclippi:
print("Weight clipping policy network = {}".format(wclippi))
policyparams = find_trainable_variables('a2c_model/pi')
self._wclip_ops_pi = []
self.wclip_bounds_pi = [-wclippi, wclippi]
for toclipvar in policyparams:
if 'logstd' in toclipvar.name:
continue
self._wclip_ops_pi.append(
tf.assign(
toclipvar,
tf.clip_by_value(toclipvar, self.wclip_bounds_pi[0],
self.wclip_bounds_pi[1])))
self._wclip_op_pi = tf.group(*self._wclip_ops_pi)
if self.toweightclipvf:
print("Weight clipping value network = {}".format(wclipvf))
valueparams = find_trainable_variables('a2c_model/vf')
self._wclip_ops_vf = []
self.wclip_bounds_vf = [-wclipvf, wclipvf]
for toclipvar in valueparams:
self._wclip_ops_vf.append(
tf.assign(
toclipvar,
tf.clip_by_value(toclipvar, self.wclip_bounds_vf[0],
self.wclip_bounds_vf[1])))
self._wclip_op_vf = tf.group(*self._wclip_ops_vf)
lr = Scheduler(v=lr, nvalues=total_timesteps, schedule=lrschedule)
def train(obs, states, rewards, masks, actions, values):
# Here we calculate advantage A(s,a) = R + yV(s') - V(s)
# rewards = R + yV(s')
advs = rewards - values
for step in range(len(obs)):
cur_lr = lr.value()
td_map = {
train_model.X: obs,
A: actions,
ADV: advs,
R: rewards,
LR: cur_lr
}
if states is not None:
td_map[train_model.S] = states
td_map[train_model.M] = masks
if self.dropoutpi_keep_prob is not None:
td_map[
self.dropoutpi_keep_prob] = self.dropoutpi_keep_prob_value
if self.dropoutvf_keep_prob is not None:
td_map[
self.dropoutvf_keep_prob] = self.dropoutvf_keep_prob_value
if self.isbnpitrainmode is not None:
td_map[self.isbnpitrainmode] = True
if self.isbnvftrainmode is not None:
td_map[self.isbnvftrainmode] = True
train_tensors = [pg_loss, vf_loss, entropy, _train]
if self.toweightclippi:
train_tensors.append(self._wclip_op_pi)
if self.toweightclipvf:
train_tensors.append(self._wclip_op_vf)
policy_loss, value_loss, policy_entropy, _ = sess.run(
train_tensors, td_map)[:4]
return policy_loss, value_loss, policy_entropy
self.train = train
self.train_model = train_model
self.step_model = step_model
self.step = step_model.step
self.value = step_model.value
self.initial_state = step_model.initial_state
self.save = functools.partial(tf_util.save_variables, sess=sess)
self.load = functools.partial(tf_util.load_variables, sess=sess)
tf.global_variables_initializer().run(session=sess)
