def build_graph(self, graph):
self.env.seed(self.random_seed)
np.random.seed(self.random_seed)
with graph.as_default():
tf.set_random_seed(self.random_seed)
# Graph of the LSTM model of the world
input_scope = tf.VariableScope(reuse=False, name="inputs")
with tf.variable_scope(input_scope):
self.state_input_plh = tf.placeholder(
tf.float32,
shape=[None, None, self.m_params['env_state_size']],
name='state_input_plh')
self.action_input_plh = tf.placeholder(tf.int32,
shape=[None, None, 1],
name='action_input_plh')
self.mask_plh = tf.placeholder(tf.float32,
shape=[None, None, 1],
name="mask_plh")
input_shape = tf.shape(self.state_input_plh)
dynamic_batch_size, dynamic_num_steps = input_shape[
0], input_shape[1]
action_input = tf.one_hot(indices=tf.squeeze(
self.action_input_plh, 2),
depth=self.m_params['nb_actions'])
m_inputs = tf.concat([self.state_input_plh, action_input],
2,
name="m_inputs")
m_scope = tf.VariableScope(reuse=False, name="m")
with tf.variable_scope(m_scope):
self.state_reward_preds, self.m_final_state, self.m_initial_state = capacities.predictive_model(
self.m_params,
m_inputs,
dynamic_batch_size,
None,
summary_collections=[self.M_SUMMARIES])
fixed_m_scope = tf.VariableScope(reuse=False, name='FixedM')
with tf.variable_scope(fixed_m_scope):
self.update_m_fixed_vars_op = capacities.fix_scope(m_scope)
m_training_scope = tf.VariableScope(reuse=False, name='m_training')
with tf.variable_scope(m_training_scope):
self.m_next_states = tf.placeholder(
tf.float32,
shape=[None, None, self.m_params['env_state_size']],
name="m_next_states")
self.m_rewards = tf.placeholder(tf.float32,
shape=[None, None, 1],
name="m_rewards")
y_true = tf.concat([self.m_rewards, self.m_next_states], 2)
with tf.control_dependencies([self.state_reward_preds]):
self.m_loss = 1 / 2 * tf.reduce_mean(
tf.square(self.state_reward_preds - y_true) *
self.mask_plh)
tf.summary.scalar('m_loss',
self.m_loss,
collections=[self.M_SUMMARIES])
m_adam = tf.train.AdamOptimizer(self.m_params['lr'])
self.m_global_step = tf.Variable(0,
trainable=False,
name="m_global_step")
tf.summary.scalar('m_global_step',
self.m_global_step,
collections=[self.M_SUMMARIES])
self.m_train_op = m_adam.minimize(
self.m_loss, global_step=self.m_global_step)
self.all_m_summary_t = tf.summary.merge_all(key=self.M_SUMMARIES)
# Graph of the controller
c_scope = tf.VariableScope(reuse=False, name="c")
c_summary_collection = [self.C_SUMMARIES]
with tf.variable_scope(c_scope):
# c_cell = LSTMCell(
# num_units=self.c_params['nb_units']
# , initializer=tf.truncated_normal_initializer(
# mean=self.c_params['initial_mean']
# , stddev=self.c_params['initial_stddev']
# )
# )
# self.c_initial_state = c_cell.zero_state(dynamic_batch_size, dtype=tf.float32)
# c_c_h_states, self.c_final_state = tf.nn.dynamic_rnn(c_cell, self.state_input_plh, initial_state=self.c_initial_state)
# c_c_states, c_h_states = tf.split(value=c_c_h_states, num_or_size_splits=[self.c_params['nb_units'], self.c_params['nb_units']], axis=2)
# # Compute the Controller projection
# self.probs_t, self.actions_t = projection_func(c_h_states)
m_params = self.m_params
model_func = lambda m_inputs, m_state: capacities.predictive_model(
m_params, m_inputs, dynamic_batch_size, m_state)
c_params = self.c_params
projection_func = lambda inputs: capacities.projection(
c_params, inputs)
cm_cell = CMCell(num_units=self.c_params['nb_units'],
m_units=self.m_params['nb_units'],
fixed_model_scope=fixed_m_scope,
model_func=model_func,
projection_func=projection_func,
num_proj=self.c_params['nb_actions'],
initializer=tf.truncated_normal_initializer(
mean=self.c_params['initial_mean'],
stddev=self.c_params['initial_stddev']))
self.cm_initial_state = cm_cell.zero_state(dynamic_batch_size,
dtype=tf.float32)
probs_and_actions_t, self.cm_final_state = tf.nn.dynamic_rnn(
cm_cell,
self.state_input_plh,
initial_state=self.cm_initial_state)
self.probs_t, actions_t = tf.split(
value=probs_and_actions_t,
num_or_size_splits=[self.c_params['nb_actions'], 1],
axis=2)
self.actions_t = tf.cast(actions_t, tf.int32)
# helper tensor used for inference
self.action_t = self.actions_t[0, 0, 0]
c_training_scope = tf.VariableScope(reuse=False, name='c_training')
with tf.variable_scope(c_training_scope):
self.c_rewards_plh = tf.placeholder(tf.float32,
shape=[None, None, 1],
name="c_rewards_plh")
baseline = tf.reduce_mean(self.c_rewards_plh)
batch_size, num_steps = tf.shape(self.actions_t)[0], tf.shape(
self.actions_t)[1]
line_indices = tf.matmul( # Line indice
tf.reshape(tf.range(0, batch_size), [-1, 1]),
tf.ones([1, num_steps], dtype=tf.int32))
column_indices = tf.matmul( # Column indice
tf.ones([batch_size, 1], dtype=tf.int32),
tf.reshape(tf.range(0, num_steps), [1, -1]))
depth_indices = tf.squeeze(self.actions_t, 2)
stacked_actions = tf.stack(
[line_indices, column_indices, depth_indices], 2)
with tf.control_dependencies([self.probs_t]):
log_probs = tf.expand_dims(
tf.log(tf.gather_nd(self.probs_t, stacked_actions)), 2)
masked_log_probs = log_probs * self.mask_plh
self.c_loss = tf.reduce_mean(-tf.reduce_sum(
masked_log_probs * (self.c_rewards_plh - baseline), 1))
tf.summary.scalar('c_loss',
self.c_loss,
collections=c_summary_collection)
c_adam = tf.train.AdamOptimizer(self.c_params['lr'])
self.c_global_step = tf.Variable(
0,
trainable=False,
name="global_step",
collections=[
tf.GraphKeys.GLOBAL_STEP, tf.GraphKeys.GLOBAL_VARIABLES
],
dtype=tf.int32)
tf.summary.scalar('c_global_step',
self.c_global_step,
collections=c_summary_collection)
self.c_train_op = c_adam.minimize(
self.c_loss, global_step=self.c_global_step)
self.all_c_summary_t = tf.summary.merge_all(key=self.C_SUMMARIES)
self.score_plh = tf.placeholder(tf.float32, shape=[])
self.score_sum_t = tf.summary.scalar('score', self.score_plh)
self.episode_id, self.inc_ep_id_op = capacities.counter(
"episode_id")
self.episode_id_sum = tf.summary.scalar('episode_id',
self.episode_id)
self.time, self.inc_time_op = capacities.counter("time")
# Playing part
self.pscore_plh = tf.placeholder(tf.float32, shape=[])
self.pscore_sum_t = tf.summary.scalar('play_score',
self.pscore_plh)
return graph
def build_graph(self, graph):
with graph.as_default():
tf.set_random_seed(self.random_seed)
self.inputs_plh = tf.placeholder(tf.int32,
shape=[None],
name="inputs_plh")
q_scope = tf.VariableScope(reuse=False, name='QValues')
with tf.variable_scope(q_scope):
self.Qs = tf.get_variable(
'Qs',
shape=[self.nb_state, self.action_space.n],
initializer=tf.constant_initializer(self.initial_q_value),
dtype=tf.float32)
tf.summary.histogram('Qarray', self.Qs)
self.q_preds_t = tf.gather(self.Qs, self.inputs_plh)
fixed_q_scope = tf.VariableScope(reuse=False, name='FixedQValues')
with tf.variable_scope(fixed_q_scope):
self.update_fixed_vars_op = capacities.fix_scope(q_scope)
policy_scope = tf.VariableScope(reuse=False, name='Policy')
with tf.variable_scope(policy_scope):
if 'UCB' in self.config and self.config['UCB']:
self.actions_t, self.probs_t = capacities.tabular_UCB(
self.Qs, self.inputs_plh)
else:
self.actions_t, self.probs_t = capacities.tabular_eps_greedy(
self.inputs_plh, self.q_preds_t, self.nb_state,
self.env.action_space.n, self.N0, self.min_eps)
self.action_t = self.actions_t[0]
self.q_value_t = self.q_preds_t[0][self.action_t]
# Experienced replay part
with tf.variable_scope('Learning'):
with tf.variable_scope(fixed_q_scope, reuse=True):
fixed_Qs = tf.get_variable('Qs')
self.rewards_plh = tf.placeholder(tf.float32,
shape=[None],
name="rewards_plh")
self.next_states_plh = tf.placeholder(tf.int32,
shape=[None],
name="next_states_plh")
# Note that we use the fixed Qs to create the targets
self.targets_t = capacities.get_q_learning_target(
fixed_Qs, self.rewards_plh, self.next_states_plh,
self.discount)
self.loss, self.train_op = capacities.tabular_learning_with_lr(
self.lr, self.lr_decay_steps, self.Qs, self.inputs_plh,
self.actions_t, self.targets_t)
self.score_plh = tf.placeholder(tf.float32, shape=[])
self.score_sum_t = tf.summary.scalar('score', self.score_plh)
self.loss_plh = tf.placeholder(tf.float32, shape=[])
self.loss_sum_t = tf.summary.scalar('loss', self.loss_plh)
self.all_summary_t = tf.summary.merge_all()
self.episode_id, self.inc_ep_id_op = capacities.counter(
"episode_id")
self.event_count, self.inc_event_count_op = capacities.counter(
"event_count")
# Playing part
self.pscore_plh = tf.placeholder(tf.float32, shape=[])
self.pscore_sum_t = tf.summary.scalar('play_score',
self.pscore_plh)
return graph
def build_graph(self, graph):
with graph.as_default():
tf.set_random_seed(self.random_seed)
self.inputs = tf.placeholder(
tf.float32,
shape=[None, self.observation_space.shape[0] + 1],
name='inputs')
q_scope = tf.VariableScope(reuse=False, name='QValues')
with tf.variable_scope(q_scope):
self.q_values = tf.squeeze(
capacities.value_f(self.q_params, self.inputs))
self.action_t = capacities.eps_greedy(self.inputs, self.q_values,
self.env.action_space.n,
self.N0, self.min_eps)
self.q_t = self.q_values[self.action_t]
fixed_q_scope = tf.VariableScope(reuse=False, name='FixedQValues')
with tf.variable_scope(fixed_q_scope):
self.update_fixed_vars_op = capacities.fix_scope(q_scope)
with tf.variable_scope('ExperienceReplay'):
self.er_inputs = tf.placeholder(
tf.float32,
shape=[None, self.observation_space.shape[0] + 1],
name="ERInputs")
self.er_actions = tf.placeholder(tf.int32,
shape=[None],
name="ERInputs")
self.er_rewards = tf.placeholder(tf.float32,
shape=[None],
name="ERReward")
self.er_next_states = tf.placeholder(
tf.float32,
shape=[None, self.observation_space.shape[0] + 1],
name="ERNextState")
with tf.variable_scope(q_scope, reuse=True):
er_q_values = capacities.value_f(self.q_params,
self.er_inputs)
er_stacked_actions = tf.stack([
tf.range(0,
tf.shape(self.er_actions)[0]), self.er_actions
], 1)
er_qs = tf.gather_nd(er_q_values, er_stacked_actions)
with tf.variable_scope(fixed_q_scope, reuse=True):
er_fixed_next_q_values = capacities.value_f(
self.q_params, self.er_next_states)
with tf.variable_scope(q_scope, reuse=True):
er_next_q_values = capacities.value_f(
self.q_params, self.er_next_states)
er_next_max_action_t = tf.cast(tf.argmax(er_next_q_values, 1),
tf.int32)
er_next_stacked_actions = tf.stack([
tf.range(0,
tf.shape(self.er_next_states)[0]),
er_next_max_action_t
], 1)
er_next_qs = tf.gather_nd(er_fixed_next_q_values,
er_next_stacked_actions)
er_target_qs1 = tf.stop_gradient(self.er_rewards +
self.discount * er_next_qs)
er_target_qs2 = self.er_rewards
er_stacked_targets = tf.stack([er_target_qs1, er_target_qs2],
1)
select_targets = tf.stack([
tf.range(0,
tf.shape(self.er_next_states)[0]),
tf.cast(self.er_next_states[:, -1], tf.int32)
], 1)
er_target_qs = tf.gather_nd(er_stacked_targets, select_targets)
self.er_loss = 1 / 2 * tf.reduce_sum(
tf.square(er_target_qs - er_qs))
er_adam = tf.train.AdamOptimizer(self.lr)
self.global_step = tf.Variable(
0,
trainable=False,
name="global_step",
collections=[
tf.GraphKeys.GLOBAL_STEP, tf.GraphKeys.GLOBAL_VARIABLES
])
self.er_train_op = er_adam.minimize(
self.er_loss, global_step=self.global_step)
self.score_plh = tf.placeholder(tf.float32, shape=[])
self.score_sum_t = tf.summary.scalar('score', self.score_plh)
self.loss_plh = tf.placeholder(tf.float32, shape=[])
self.loss_sum_t = tf.summary.scalar('loss', self.loss_plh)
self.all_summary_t = tf.summary.merge_all()
self.episode_id, self.inc_ep_id_op = capacities.counter(
"episode_id")
self.timestep, self.inc_timestep_op = capacities.counter(
"timestep")
# Playing part
self.pscore_plh = tf.placeholder(tf.float32, shape=[])
self.pscore_sum_t = tf.summary.scalar('play_score',
self.pscore_plh)
return graph
def build_graph(self, graph):
np.random.seed(self.random_seed)
with graph.as_default():
tf.set_random_seed(self.random_seed)
self.inputs = tf.placeholder(tf.float32, shape=[None, None, self.policy_params['nb_inputs']], name='inputs')
input_shape = tf.shape(self.inputs)
dynamic_batch_size, dynamic_num_steps = input_shape[0], input_shape[1]
inputs_mat = tf.reshape(self.inputs, [-1, self.policy_params['nb_inputs']])
policy_scope = tf.VariableScope(reuse=False, name='Policy')
with tf.variable_scope(policy_scope):
probs, actions = capacities.policy(self.policy_params, inputs_mat)
self.probs = tf.reshape(probs, [dynamic_batch_size, dynamic_num_steps, self.policy_params['nb_outputs']])
self.actions = tf.reshape(actions, [dynamic_batch_size, dynamic_num_steps, 1])
self.action_t = self.actions[0, 0, 0]
critic_scope = tf.VariableScope(reuse=False, name='QValues')
with tf.variable_scope(critic_scope):
critic_values_mat = capacities.value_f(self.critic_params, inputs_mat)
self.critic_values = tf.reshape(critic_values_mat, [dynamic_batch_size, dynamic_num_steps, self.critic_params['nb_outputs']])
fixed_critic_scope = tf.VariableScope(reuse=False, name='FixedQValues')
with tf.variable_scope(fixed_critic_scope):
self.update_fixed_vars_op = capacities.fix_scope(critic_scope)
with tf.variable_scope('Training'):
self.expected_rewards = tf.placeholder(tf.float32, shape=[None, None, 1], name="reward")
self.mask_plh = tf.placeholder(tf.float32, shape=[None, None, 1], name="mask_plh")
batch_size, num_steps = tf.shape(self.actions)[0], tf.shape(self.actions)[1]
line_indices = tf.matmul( # Line indice
tf.reshape(tf.range(0, batch_size), [-1, 1])
, tf.ones([1, num_steps], dtype=tf.int32)
)
column_indices = tf.matmul( # Column indice
tf.ones([batch_size, 1], dtype=tf.int32)
, tf.reshape(tf.range(0, num_steps), [1, -1])
)
depth_indices = tf.cast(tf.squeeze(self.actions, 2), tf.int32)
stacked_actions = tf.stack(
[line_indices, column_indices, depth_indices], 2
)
log_probs = tf.expand_dims(tf.log(tf.gather_nd(self.probs, stacked_actions)), 2)
self.policy_loss = tf.reduce_mean( - tf.reduce_sum((log_probs * (self.expected_rewards - tf.stop_gradient(self.critic_values))) * self.mask_plh, 1))
adam = tf.train.AdamOptimizer(self.lr)
self.train_policy_op = adam.minimize(self.policy_loss)
self.rewards = tf.placeholder(tf.float32, shape=[None, None, 1], name="reward")
self.next_states = tf.placeholder(tf.float32, shape=[None, None, self.critic_params['nb_inputs']], name="next_states")
with tf.variable_scope(fixed_critic_scope, reuse=True):
next_states_mat = tf.reshape(self.next_states, [-1, self.critic_params['nb_inputs']])
next_critic_values_mat = capacities.value_f(self.critic_params, next_states_mat)
next_critic_values = tf.reshape(next_critic_values_mat, [dynamic_batch_size, dynamic_num_steps, self.critic_params['nb_outputs']])
target_critics1 = tf.stop_gradient(self.rewards + self.discount * next_critic_values)
target_critics2 = self.rewards
stacked_targets = tf.stack([tf.squeeze(target_critics1, 2), tf.squeeze(target_critics2, 2)], 2)
batch_size, num_steps = tf.shape(self.next_states)[0], tf.shape(self.next_states)[1]
line_indices = tf.matmul( # Line indice
tf.reshape(tf.range(0, batch_size), [-1, 1])
, tf.ones([1, num_steps], dtype=tf.int32)
)
column_indices = tf.matmul( # Column indice
tf.ones([batch_size, 1], dtype=tf.int32)
, tf.reshape(tf.range(0, num_steps), [1, -1])
)
depth_indices = tf.cast(self.next_states[:, :, -1], tf.int32)
select_targets = tf.stack(
[line_indices, column_indices, depth_indices], 2
)
target_critics = tf.expand_dims(tf.gather_nd(stacked_targets, select_targets), 2)
self.critic_loss = 1/2 * tf.reduce_sum(tf.square(target_critics - self.critic_values) * self.mask_plh)
adam = tf.train.AdamOptimizer(self.critic_lr)
self.global_step = tf.Variable(0, trainable=False, name="global_step", collections=[tf.GraphKeys.GLOBAL_STEP, tf.GraphKeys.GLOBAL_VARIABLES])
self.train_critic_op = adam.minimize(self.critic_loss, global_step=self.global_step)
self.policy_loss_plh = tf.placeholder(tf.float32, shape=[])
self.policy_loss_sum_t = tf.summary.scalar('policy_loss', self.policy_loss_plh)
self.critic_loss_plh = tf.placeholder(tf.float32, shape=[])
self.critic_loss_sum_t = tf.summary.scalar('critic_loss', self.critic_loss_plh)
# self.loss_plh = tf.placeholder(tf.float32, shape=[])
# self.loss_sum_t = tf.summary.scalar('loss', self.loss_plh)
self.all_summary_t = tf.summary.merge_all()
self.score_plh = tf.placeholder(tf.float32, shape=[])
self.score_sum_t = tf.summary.scalar('av_score', self.score_plh)
self.episode_id, self.inc_ep_id_op = capacities.counter("episode_id")
# Playing part
self.pscore_plh = tf.placeholder(tf.float32, shape=[])
self.pscore_sum_t = tf.summary.scalar('play_score', self.pscore_plh)
return graph