def create_normalizer():
if len(self.args.obs_dims)==1:
with tf.variable_scope('normalizer'):
self.obs_normalizer = Normalizer(self.args.obs_dims, self.sess)
self.obs_ph = self.obs_normalizer.normalize(self.raw_obs_ph)
self.obs_next_ph = self.obs_normalizer.normalize(self.raw_obs_next_ph)
else:
self.obs_normalizer = None
self.obs_ph = self.raw_obs_ph
self.obs_next_ph = self.raw_obs_next_ph
class DDPG:
def __init__(self, args):
self.args = args
self.create_model()
self.train_info_pi = {
'Pi_q_loss': self.pi_q_loss,
'Pi_l2_loss': self.pi_l2_loss
}
self.train_info_q = {
'Q_loss': self.q_loss
}
self.train_info = {**self.train_info_pi, **self.train_info_q}
self.step_info = {
'Q_average': self.q_pi
}
def create_model(self):
def create_session():
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
self.sess = tf.Session(config=config)
def create_inputs():
self.raw_obs_ph = tf.placeholder(tf.float32, [None]+self.args.obs_dims)
self.raw_obs_next_ph = tf.placeholder(tf.float32, [None]+self.args.obs_dims)
self.acts_ph = tf.placeholder(tf.float32, [None]+self.args.acts_dims)
self.rews_ph = tf.placeholder(tf.float32, [None, 1])
def create_normalizer():
with tf.variable_scope('normalizer'):
self.obs_normalizer = Normalizer(self.args.obs_dims, self.sess)
self.obs_ph = self.obs_normalizer.normalize(self.raw_obs_ph)
self.obs_next_ph = self.obs_normalizer.normalize(self.raw_obs_next_ph)
def create_network():
def mlp_policy(obs_ph):
with tf.variable_scope('net', initializer=tf.contrib.layers.xavier_initializer()):
pi_dense1 = tf.layers.dense(obs_ph, 256, activation=tf.nn.relu, name='pi_dense1')
pi_dense2 = tf.layers.dense(pi_dense1, 256, activation=tf.nn.relu, name='pi_dense2')
pi_dense3 = tf.layers.dense(pi_dense2, 256, activation=tf.nn.relu, name='pi_dense3')
pi = tf.layers.dense(pi_dense3, self.args.acts_dims[0], activation=tf.nn.tanh, name='pi')
return pi
def mlp_value(obs_ph, acts_ph):
state_ph = tf.concat([obs_ph, acts_ph], axis=1)
with tf.variable_scope('net', initializer=tf.contrib.layers.xavier_initializer()):
q_dense1 = tf.layers.dense(state_ph, 256, activation=tf.nn.relu, name='q_dense1')
q_dense2 = tf.layers.dense(q_dense1, 256, activation=tf.nn.relu, name='q_dense2')
q_dense3 = tf.layers.dense(q_dense2, 256, activation=tf.nn.relu, name='q_dense3')
q = tf.layers.dense(q_dense3, 1, name='q')
return q
with tf.variable_scope('main'):
with tf.variable_scope('policy'):
self.pi = mlp_policy(self.obs_ph)
with tf.variable_scope('value'):
self.q = mlp_value(self.obs_ph, self.acts_ph)
with tf.variable_scope('value', reuse=True):
self.q_pi = mlp_value(self.obs_ph, self.pi)
with tf.variable_scope('target'):
with tf.variable_scope('policy'):
self.pi_t = mlp_policy(self.obs_next_ph)
with tf.variable_scope('value'):
self.q_t = mlp_value(self.obs_next_ph, self.pi_t)
def create_operators():
self.pi_q_loss = -tf.reduce_mean(self.q_pi)
self.pi_l2_loss = self.args.act_l2*tf.reduce_mean(tf.square(self.pi))
self.pi_optimizer = tf.train.AdamOptimizer(self.args.pi_lr)
self.pi_train_op = self.pi_optimizer.minimize(self.pi_q_loss+self.pi_l2_loss, var_list=get_vars('main/policy'))
if self.args.clip_return:
return_value = tf.clip_by_value(self.q_t, self.args.clip_return_l, self.args.clip_return_r)
else:
return_value = self.q_t
target = tf.stop_gradient(self.rews_ph+self.args.gamma*return_value)
self.q_loss = tf.reduce_mean(tf.square(self.q-target))
self.q_optimizer = tf.train.AdamOptimizer(self.args.q_lr)
self.q_train_op = self.q_optimizer.minimize(self.q_loss, var_list=get_vars('main/value'))
self.target_update_op = tf.group([
v_t.assign(self.args.polyak*v_t + (1.0-self.args.polyak)*v)
for v, v_t in zip(get_vars('main'), get_vars('target'))
])
self.saver=tf.train.Saver()
self.init_op = tf.global_variables_initializer()
self.target_init_op = tf.group([
v_t.assign(v)
for v, v_t in zip(get_vars('main'), get_vars('target'))
])
self.graph = tf.Graph()
with self.graph.as_default():
create_session()
create_inputs()
create_normalizer()
create_network()
create_operators()
self.init_network()
def init_network(self):
self.sess.run(self.init_op)
self.sess.run(self.target_init_op)
def step(self, obs, explore=False, test_info=False):
if (not test_info) and (self.args.buffer.steps_counter<self.args.warmup):
return np.random.uniform(-1, 1, size=self.args.acts_dims)
if self.args.goal_based: obs = goal_based_process(obs)
# eps-greedy exploration
if explore and np.random.uniform()<=self.args.eps_act:
return np.random.uniform(-1, 1, size=self.args.acts_dims)
feed_dict = {
self.raw_obs_ph: [obs]
}
action, info = self.sess.run([self.pi, self.step_info], feed_dict)
action = action[0]
# uncorrelated gaussian explorarion
if explore: action += np.random.normal(0, self.args.std_act, size=self.args.acts_dims)
action = np.clip(action, -1, 1)
if test_info: return action, info
return action
def step_batch(self, obs):
actions = self.sess.run(self.pi, {self.raw_obs_ph:obs})
return actions
def feed_dict(self, batch):
return {
self.raw_obs_ph: batch['obs'],
self.raw_obs_next_ph: batch['obs_next'],
self.acts_ph: batch['acts'],
self.rews_ph: batch['rews']
}
def train(self, batch):
feed_dict = self.feed_dict(batch)
info, _, _ = self.sess.run([self.train_info, self.pi_train_op, self.q_train_op], feed_dict)
return info
def train_pi(self, batch):
feed_dict = self.feed_dict(batch)
info, _ = self.sess.run([self.train_info_pi, self.pi_train_op], feed_dict)
return info
def train_q(self, batch):
feed_dict = self.feed_dict(batch)
info, _ = self.sess.run([self.train_info_q, self.q_train_op], feed_dict)
return info
def normalizer_update(self, batch):
self.obs_normalizer.update(np.concatenate([batch['obs'], batch['obs_next']], axis=0))
def target_update(self):
self.sess.run(self.target_update_op)
class DDQ6(BaseLearner):
def __init__(self, args, flags={}):
super(DDQ6, self).__init__()
self.args = args
self.gpu = args.gpu
self.flags = flags
self.acts_num = args.acts_dims[0]
self.inner_q_type = args.inner_q_type
self.num_q = self.args.num_q
self.tau = self.args.tau
self.alpha = self.args.alpha
self.beta = self.args.beta
self.q_funcs = []
self.q_pi_funcs = []
self.target_q_funcs = []
self.target_q_pi_funcs = []
self.meta_q_funcs = []
self.meta_q_pi_funcs = []
self.target_meta_q_funcs = []
self.target_meta_q_pi_funcs = []
self.target_qs = None
self.qs = None
self.meta_q_funcs_stack = None
self.meta_q_pi = None
self.meta_target_check_range = None
self.create_model()
self.train_info = {
'Q_loss': self.q_loss,
'Meta_Q_loss': self.meta_q_loss,
# 'Q_target_0': self.q_step_target[:, 0],
# 'Q_target_1': self.q_step_target[:, 1],
'difference': self.buffer_target_diffence,
'target_range': self.target_check_range,
'meta_target_range': self.meta_target_check_range,
'regression_target': self.qvalues_ph,
'true_return': self.true_rews_ph,
}
self.step_info = {
'Q_average': self.meta_q_pi,
'sub_Q_average': self.q_pi,
}
self.args.buffer.update_func(self)
def create_model(self):
# def create_session():
# config = tf.ConfigProto()
# config.gpu_options.allow_growth = True
# self.sess = tf.Session(config=config)
def create_inputs():
self.raw_obs_ph = tf.placeholder(tf.float32,
[None] + self.args.obs_dims)
self.em_raw_obs_ph = tf.placeholder(tf.float32,
[None] + self.args.obs_dims)
self.acts_ph = tf.placeholder(tf.float32,
[None] + self.args.acts_dims + [1])
self.rews_ph = tf.placeholder(tf.float32, [None, 1])
self.true_rews_ph = tf.placeholder(tf.float32, [None, 1])
self.done_ph = tf.placeholder(tf.float32, [None, 1])
self.qvalues_ph = tf.placeholder(tf.float32, [None, 2])
def create_normalizer():
if len(self.args.obs_dims) == 1:
with tf.variable_scope('normalizer'):
self.obs_normalizer = Normalizer(self.args.obs_dims,
self.sess)
self.obs_ph = self.obs_normalizer.normalize(self.raw_obs_ph)
self.em_obs_ph = self.obs_normalizer.normalize(
self.em_raw_obs_ph)
else:
self.obs_normalizer = None
self.obs_ph = self.raw_obs_ph
self.em_obs_ph = self.em_raw_obs_ph
def create_network():
value_net = self.mlp_value if len(
self.args.obs_dims) == 1 else self.conv_value
with tf.variable_scope('main'):
with tf.variable_scope('sub'):
for i in range(self.num_q):
with tf.variable_scope('value_{}'.format(i)):
q = value_net(self.obs_ph)
q_pi = tf.reduce_max(q, axis=1, keepdims=True)
self.q_funcs.append(q)
self.q_pi_funcs.append(q_pi)
with tf.variable_scope('meta'):
for i in range(2):
with tf.variable_scope('meta_value_{}'.format(i)):
q = value_net(self.obs_ph)
q_pi = tf.reduce_max(q, axis=1, keepdims=True)
self.meta_q_funcs.append(q)
self.meta_q_pi_funcs.append(q_pi)
self.qs = tf.stack(self.q_funcs, axis=-1)
self.q_funcs_stack = tf.reshape(
self.qs, [-1] + self.args.acts_dims + [2, self.num_q // 2])
self.q_pi = tf.reduce_mean(tf.reduce_max(tf.reduce_mean(
self.q_funcs_stack, axis=-1),
axis=1),
axis=-1)
self.meta_q_funcs_stack = tf.stack(self.meta_q_funcs, axis=-1)
self.q_step = tf.reduce_mean(self.meta_q_funcs_stack, axis=-1)
self.meta_q_pi = tf.reduce_mean(tf.reduce_max(self.q_step,
axis=1),
axis=-1)
with tf.variable_scope('target'):
with tf.variable_scope('sub'):
for i in range(self.num_q):
with tf.variable_scope('value_{}'.format(i)):
tar_q = value_net(self.em_obs_ph)
tar_q_pi = tf.reduce_max(tar_q,
axis=1,
keepdims=True)
self.target_q_funcs.append(tar_q)
self.target_q_pi_funcs.append(tar_q_pi)
with tf.variable_scope('meta'):
for i in range(2):
with tf.variable_scope('meta_value_{}'.format(i)):
tar_q = value_net(self.em_obs_ph)
tar_q_pi = tf.reduce_max(tar_q,
axis=1,
keepdims=True)
self.target_meta_q_funcs.append(tar_q)
self.target_meta_q_pi_funcs.append(tar_q_pi)
self.target_qs = tf.stack(self.target_q_funcs, axis=-1)
self.target_q_funcs_stack = tf.reshape(
self.target_qs,
[-1] + self.args.acts_dims + [2, self.num_q // 2])
if self.inner_q_type == "min":
self.em_q = tf.reduce_max(tf.reduce_min(
self.target_q_funcs_stack, axis=-1),
axis=1)
else:
self.em_q = tf.reduce_max(tf.reduce_mean(
self.target_q_funcs_stack, axis=-1),
axis=1)
# self.q_target_mean = tf.reduce_mean(self.em_q, axis=1, keepdims=True)
# self.q_step_target = tf.reduce_mean(self.target_q_funcs_stack, axis=-1)
def create_operators():
self.target_q_max = tf.reduce_max(self.target_qs, axis=1)
self.target = tf.stop_gradient(self.rews_ph +
(1.0 - self.done_ph) *
self.args.gamma * self.target_q_max)
self.q_acts = tf.reduce_sum(self.qs * self.acts_ph,
axis=1,
keepdims=True)
q_acts = tf.reshape(self.q_acts, shape=(-1, self.num_q))
self.meta_q_acts = tf.reduce_sum(self.meta_q_funcs_stack *
self.acts_ph,
axis=1,
keepdims=True)
meta_q_acts = tf.reshape(self.meta_q_acts, shape=(-1, 2))
self.meta_target_check_range = tf.reduce_max(
tf.abs(self.qvalues_ph))
self.target_check_range = tf.reduce_max(tf.abs(self.target))
duplicate_qvalues = tf.stack(
[self.qvalues_ph for _ in range(self.num_q // 2)], axis=-1)
duplicate_qvalues = tf.reshape(duplicate_qvalues, (-1, self.num_q))
self.buffer_target_diffence = tf.reduce_mean(
(duplicate_qvalues - self.target)**2)
# self.q_loss = tf.reduce_mean(tf.abs(q_acts - self.qvalues_ph))
# self.q_loss = tf.reduce_mean(tf.nn.leaky_relu(q_acts - self.qvalues_ph, alpha=self.alpha)**2)
self.q_loss = tf.losses.huber_loss(q_acts,
self.target,
reduction=Reduction.SUM)
# self.q_loss = tf.losses.huber_loss(q_acts, self.target)
sym_q_loss = huber_loss(meta_q_acts, self.qvalues_ph)
overestimate = (meta_q_acts - self.qvalues_ph) > 0
self.meta_q_loss = tf.reduce_sum(
tf.where(overestimate, sym_q_loss, self.alpha * sym_q_loss))
if self.args.optimizer == 'adam':
self.q_optimizer = tf.train.AdamOptimizer(
self.args.q_lr, epsilon=self.args.Adam_eps)
self.meta_q_optimizer = tf.train.AdamOptimizer(
self.args.q_lr, epsilon=self.args.Adam_eps)
elif self.args.optimizer == 'rmsprop':
self.q_optimizer = tf.train.RMSPropOptimizer(
self.args.q_lr,
decay=self.args.RMSProp_decay,
epsilon=self.args.RMSProp_eps)
self.meta_q_optimizer = tf.train.RMSPropOptimizer(
self.args.q_lr,
decay=self.args.RMSProp_decay,
epsilon=self.args.RMSProp_eps)
self.q_train_op = self.q_optimizer.minimize(
self.q_loss, var_list=get_vars('main/sub'))
self.meta_q_train_op = self.meta_q_optimizer.minimize(
10 * self.meta_q_loss, var_list=get_vars('main/meta'))
self.target_update_op = tf.group([
v_t.assign(self.tau * v + (1 - self.tau) * v_t)
for v, v_t in zip(get_vars('main'), get_vars('target'))
])
self.saver = tf.train.Saver()
self.init_op = tf.global_variables_initializer()
self.target_init_op = tf.group([
v_t.assign(v)
for v, v_t in zip(get_vars('main'), get_vars('target'))
])
self.graph = tf.Graph()
with self.graph.as_default():
# with tf.device("/gpu:{}".format(self.gpu)):
self.create_session()
create_inputs()
create_normalizer()
create_network()
create_operators()
self.init_network()
def init_network(self):
self.sess.run(self.init_op)
self.sess.run(self.target_init_op)
def step(self, obs, explore=False, test_info=False):
if (not test_info) and (self.args.buffer.steps_counter <
self.args.warmup):
return np.random.randint(self.acts_num)
# eps-greedy exploration
if explore and np.random.uniform() <= self.args.eps_act:
return np.random.randint(self.acts_num)
feed_dict = {
self.raw_obs_ph: [obs / 255.0],
}
q_values, info = self.sess.run([self.q_step, self.step_info],
feed_dict)
# q_value = np.mean(q_values, axis=-1)
action = np.argmax(q_values[0])
if test_info: return action, info
return action
def step_with_q(self, obs, explore=False, target=True):
feed_dict = {
self.raw_obs_ph: [obs / 255.0],
self.em_raw_obs_ph: [obs / 255.0],
}
q_values, q_values_target = self.sess.run([self.q_step, self.em_q],
feed_dict)
# q_value = np.mean(q_values, axis=-1)
action = np.argmax(q_values[0])
q = q_values_target if target else q_values
if self.args.buffer.steps_counter < self.args.warmup:
return np.random.randint(self.acts_num), q
# eps-greedy exploration
if explore and np.random.uniform() <= self.args.eps_act:
return np.random.randint(self.acts_num), q
return action, q
def feed_dict(self, batch):
def one_hot(idx):
idx = np.array(idx).reshape(-1)
batch_size = idx.shape[0]
res = np.zeros((batch_size, self.acts_num), dtype=np.float32)
res[np.arange(batch_size), idx] = 1.0
return res
batch_obs, batch_obs_next, batch_actions, batch_rewards, batch_next_obs, batch_dones, batch_returns, batch_true_returns = \
batch['obs0'], batch['obs1'], batch[
'actions'], batch['rewards'], batch['obs1'], batch['terminals1'], batch['return'], batch['true_return']
# if self.num_q == 4:
# batch_returns = np.repeat(batch_returns, 2, axis=1)
feed_dict = {
self.raw_obs_ph: batch_obs,
self.em_raw_obs_ph: batch_obs_next,
self.acts_ph: one_hot(batch_actions)[..., np.newaxis],
self.rews_ph: batch_rewards,
self.done_ph: batch_dones,
self.qvalues_ph: batch_returns,
self.true_rews_ph: batch_true_returns,
}
return feed_dict
def train(self, batch):
feed_dict = self.feed_dict(batch)
info, _, _ = self.sess.run(
[self.train_info, self.q_train_op, self.meta_q_train_op],
feed_dict)
return info
def test_q(self, batch):
feed_dict = self.feed_dict(batch)
q_loss, meta_q_loss = self.sess.run([self.q_loss, self.meta_q_loss],
feed_dict)
return q_loss, meta_q_loss
def normalizer_update(self, batch):
if not (self.obs_normalizer is None):
self.obs_normalizer.update(
np.concatenate([batch['obs'], batch['obs_next']], axis=0))
def target_update(self):
self.sess.run(self.target_update_op)
def save_model(self, save_path):
with self.graph.as_default():
saver = tf.train.Saver()
saver.save(self.sess, save_path)
def load_model(self, load_path):
with self.graph.as_default():
saver = tf.train.Saver()
saver.restore(self.sess, load_path)
class LowRankDQN:
def __init__(self, args, flags={}):
self.args = args
self.flags = flags
self.acts_num = args.acts_dims[0]
self.create_model()
self.train_info = {
'Q_loss': self.q_loss,
'Q_L1_loss': self.q_l1_loss,
'Q_sparsity': self.q_sparse_abs_avg,
'target_range': self.target_check_range
}
self.step_info = {
'Q_average': self.q_pi
}
if self.args.learn[-2:]=='lb':
self.train_info = {
**self.train_info,
**{
'Q_target': self.target,
'Q_LB': self.q_lb_ph,
'LB_ratio': self.lb_ratio
}
}
def create_model(self):
def create_session():
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
self.sess = tf.Session(config=config)
def create_inputs():
self.raw_obs_ph = tf.placeholder(tf.float32, [None]+self.args.obs_dims)
self.raw_obs_next_ph = tf.placeholder(tf.float32, [None]+self.args.obs_dims)
self.acts_ph = tf.placeholder(tf.float32, [None]+self.args.acts_dims)
self.rews_ph = tf.placeholder(tf.float32, [None, 1])
self.done_ph = tf.placeholder(tf.float32, [None, 1])
if self.args.learn[-2:]=='lb':
self.q_lb_ph = tf.placeholder(tf.float32, [None, 1])
def create_normalizer():
if len(self.args.obs_dims)==1:
with tf.variable_scope('normalizer'):
self.obs_normalizer = Normalizer(self.args.obs_dims, self.sess)
self.obs_ph = self.obs_normalizer.normalize(self.raw_obs_ph)
self.obs_next_ph = self.obs_normalizer.normalize(self.raw_obs_next_ph)
else:
self.obs_normalizer = None
self.obs_ph = self.raw_obs_ph
self.obs_next_ph = self.raw_obs_next_ph
def create_network():
def mlp_value(obs_ph):
with tf.variable_scope('net', initializer=tf.contrib.layers.xavier_initializer()):
q_dense1 = tf.layers.dense(obs_ph, 256, activation=tf.nn.relu, name='q_dense1')
q_dense2 = tf.layers.dense(q_dense1, 256, activation=tf.nn.relu, name='q_dense2')
q = tf.layers.dense(q_dense2, self.acts_num, name='q')
return q
def conv_value(obs_ph):
with tf.variable_scope('net', initializer=tf.contrib.layers.xavier_initializer()):
q_conv1 = tf.layers.conv2d(obs_ph, 32, 8, 4, 'same', activation=tf.nn.relu, name='q_conv1')
q_conv2 = tf.layers.conv2d(q_conv1, 64, 4, 2, 'same', activation=tf.nn.relu, name='q_conv2')
q_conv3 = tf.layers.conv2d(q_conv2, 64, 3, 1, 'same', activation=tf.nn.relu, name='q_conv3')
q_conv3_flat = tf.layers.flatten(q_conv3)
q_dense_act = tf.layers.dense(q_conv3_flat, 512, activation=tf.nn.relu, name='q_dense_act')
q_lowrank_hidden = tf.layers.dense(q_dense_act, self.args.rank, activation=tf.nn.relu, name='q_lowrank_hidden')
q_lowrank_act = tf.layers.dense(q_lowrank_hidden, self.acts_num, name='q_lowrank_act')
q_sparse_act = tf.layers.dense(q_dense_act, self.acts_num, name='q_sparse_act')
return q_lowrank_act+q_sparse_act, q_sparse_act
value_net = mlp_value if len(self.args.obs_dims)==1 else conv_value
with tf.variable_scope('main'):
with tf.variable_scope('value'):
self.q, self.q_sparse = value_net(self.obs_ph)
self.q_sparse_avg = tf.reduce_mean(self.q_sparse)
self.q_sparse_abs_avg = tf.reduce_mean(tf.abs(self.q_sparse))
self.q_pi = tf.reduce_max(self.q, axis=1, keepdims=True)
if self.args.double:
with tf.variable_scope('value', reuse=True):
self.q_next, _ = value_net(self.obs_next_ph)
self.pi_next = tf.one_hot(tf.argmax(self.q_next, axis=1), self.acts_num, dtype=tf.float32)
with tf.variable_scope('target'):
with tf.variable_scope('value'):
if self.args.double:
self.q_t = tf.reduce_sum(value_net(self.obs_next_ph)[0]*self.pi_next, axis=1, keepdims=True)
else:
self.q_t = tf.reduce_max(value_net(self.obs_next_ph)[0], axis=1, keepdims=True)
def create_operators():
self.target = tf.stop_gradient(self.rews_ph+(1.0-self.done_ph)*(self.args.gamma**self.args.nstep)*self.q_t)
target = self.target
if self.args.learn[-2:]=='lb':
self.lb_ratio = tf.less(target, self.q_lb_ph)
target = tf.maximum(target, self.q_lb_ph)
self.target_check_range = tf.reduce_max(tf.abs(target))
self.q_acts = tf.reduce_sum(self.q*self.acts_ph, axis=1, keepdims=True)
self.q_loss = tf.losses.huber_loss(target, self.q_acts) + self.args.beta*tf.losses.huber_loss(0, self.q_sparse_avg)
self.q_l1_loss = tf.reduce_mean(tf.abs(target-self.q_acts))
if self.args.optimizer=='adam':
self.q_optimizer = tf.train.AdamOptimizer(self.args.q_lr, epsilon=self.args.Adam_eps)
elif self.args.optimizer=='rmsprop':
self.q_optimizer = tf.train.RMSPropOptimizer(self.args.q_lr, decay=self.args.RMSProp_decay, epsilon=self.args.RMSProp_eps)
self.q_train_op = self.q_optimizer.minimize(self.q_loss, var_list=get_vars('main/value'))
self.target_update_op = tf.group([
v_t.assign(v)
for v, v_t in zip(get_vars('main'), get_vars('target'))
])
self.saver=tf.train.Saver()
self.init_op = tf.global_variables_initializer()
self.target_init_op = tf.group([
v_t.assign(v)
for v, v_t in zip(get_vars('main'), get_vars('target'))
])
self.graph = tf.Graph()
with self.graph.as_default():
create_session()
create_inputs()
create_normalizer()
create_network()
create_operators()
self.init_network()
def init_network(self):
self.sess.run(self.init_op)
self.sess.run(self.target_init_op)
def step(self, obs, explore=False, test_info=False):
if (not test_info) and (self.args.buffer.steps_counter<self.args.warmup):
return np.random.randint(self.acts_num)
# eps-greedy exploration
if explore and np.random.uniform()<=self.args.eps_act:
return np.random.randint(self.acts_num)
feed_dict = {
self.raw_obs_ph: [obs/255.0]
}
q_value, info = self.sess.run([self.q, self.step_info], feed_dict)
action = np.argmax(q_value[0])
if test_info: return action, info
return action
def feed_dict(self, batch):
def one_hot(idx):
idx = np.array(idx)
batch_size = idx.shape[0]
res = np.zeros((batch_size, self.acts_num), dtype=np.float32)
res[np.arange(batch_size),idx] = 1.0
return res
feed_dict = {
self.raw_obs_ph: np.array(batch['obs']),
self.raw_obs_next_ph: np.array(batch['obs_next']),
self.acts_ph: one_hot(batch['acts']),
self.rews_ph: np.clip(np.array(batch['rews']), -self.args.rews_scale, self.args.rews_scale),
self.done_ph: batch['done']
}
if self.args.learn[-2:]=='lb':
feed_dict[self.q_lb_ph]= batch['rets']
return feed_dict
def train(self, batch):
feed_dict = self.feed_dict(batch)
info, _ = self.sess.run([self.train_info, self.q_train_op], feed_dict)
return info
def test_q(self, batch):
feed_dict = self.feed_dict(batch)
q_loss = self.sess.run(self.q_loss, feed_dict)
return q_loss
def normalizer_update(self, batch):
if not(self.obs_normalizer is None):
self.obs_normalizer.update(np.concatenate([batch['obs'], batch['obs_next']], axis=0))
def target_update(self):
self.sess.run(self.target_update_op)
def save_model(self, save_path):
with self.graph.as_default():
saver = tf.train.Saver()
saver.save(self.sess, save_path)
def load_model(self, load_path):
with self.graph.as_default():
saver = tf.train.Saver()
saver.restore(self.sess, load_path)
