def __init__(self, sess, config):
BasePlayer.__init__(self, sess, config)
self.network = config['network']
self.use_action_masks = config.get('use_action_masks', False)
self.obs_ph = tf.placeholder(self.obs_space.dtype,
(None, ) + self.obs_space.shape,
name='obs')
self.actions_num = self.action_space.n
if self.use_action_masks:
print('using masks for action')
self.action_mask_ph = tf.placeholder('int32',
(None, self.actions_num),
name='actions_mask')
else:
self.action_mask_ph = None
self.mask = [False] * self.num_agents
self.epoch_num = tf.Variable(tf.constant(0, shape=(),
dtype=tf.float32),
trainable=False)
self.normalize_input = self.config['normalize_input']
self.input_obs = self.obs_ph
if self.obs_space.dtype == np.uint8:
self.input_obs = tf.to_float(self.input_obs) / 255.0
if self.normalize_input:
self.moving_mean_std = MovingMeanStd(shape=self.obs_space.shape,
epsilon=1e-5,
decay=0.99)
self.input_obs = self.moving_mean_std.normalize(self.input_obs,
train=False)
self.run_dict = {
'name': 'agent',
'inputs': self.input_obs,
'batch_num': self.num_agents,
'games_num': self.num_agents,
'actions_num': self.actions_num,
'prev_actions_ph': None,
'action_mask_ph': self.action_mask_ph
}
self.last_state = None
if self.network.is_rnn():
self.neglop, self.value, self.action, _, self.states_ph, self.masks_ph, self.lstm_state, self.initial_state, self.logits = self.network(
self.run_dict, reuse=False)
self.last_state = self.initial_state * self.num_agents
else:
self.neglop, self.value, self.action, _, self.logits = self.network(
self.run_dict, reuse=False)
self.variables = TensorFlowVariables(
[self.neglop, self.value, self.action], self.sess)
self.saver = tf.train.Saver()
self.sess.run(tf.global_variables_initializer())
def __init__(self,
sess,
base_name,
observation_space,
action_space,
config,
logger,
central_state_space=None):
observation_shape = observation_space.shape
actions_num = action_space.n
self.config = config
self.is_adaptive_lr = config['lr_schedule'] == 'adaptive'
self.is_polynom_decay_lr = config['lr_schedule'] == 'polynom_decay'
self.is_exp_decay_lr = config['lr_schedule'] == 'exp_decay'
self.lr_multiplier = tf.constant(1, shape=(), dtype=tf.float32)
self.learning_rate_ph = tf.placeholder('float32', (), name='lr_ph')
self.games_to_track = tr_helpers.get_or_default(
config, 'games_to_track', 100)
self.max_epochs = tr_helpers.get_or_default(self.config, 'max_epochs',
1e6)
self.games_to_log = self.config.get('games_to_track', 100)
self.game_rewards = deque([], maxlen=self.games_to_track)
self.game_lengths = deque([], maxlen=self.games_to_track)
self.game_scores = deque([], maxlen=self.games_to_log)
self.epoch_num = tf.Variable(tf.constant(0, shape=(),
dtype=tf.float32),
trainable=False)
self.update_epoch_op = self.epoch_num.assign(self.epoch_num + 1)
self.current_lr = self.learning_rate_ph
if self.is_adaptive_lr:
self.lr_threshold = config['lr_threshold']
if self.is_polynom_decay_lr:
self.lr_multiplier = tf.train.polynomial_decay(
1.0,
global_step=self.epoch_num,
decay_steps=self.max_epochs,
end_learning_rate=0.001,
power=tr_helpers.get_or_default(config, 'decay_power', 1.0))
if self.is_exp_decay_lr:
self.lr_multiplier = tf.train.exponential_decay(
1.0,
global_step=self.epoch_num,
decay_steps=self.max_epochs,
decay_rate=config['decay_rate'])
self.env_name = config['env_name']
self.network = config['network']
self.batch_size = self.config['batch_size']
self.obs_shape = observation_shape
self.actions_num = actions_num
self.writer = SummaryWriter('runs/' + config['name'] +
datetime.now().strftime("%d, %H:%M:%S"))
self.epsilon = self.config['epsilon']
self.rewards_shaper = self.config['reward_shaper']
self.epsilon_processor = tr_helpers.LinearValueProcessor(
self.config['epsilon'], self.config['min_epsilon'],
self.config['epsilon_decay_frames'])
self.beta_processor = tr_helpers.LinearValueProcessor(
self.config['priority_beta'], self.config['max_beta'],
self.config['beta_decay_frames'])
if self.env_name:
self.env_config = config.get('env_config', {})
self.env = env_configurations.configurations[
self.env_name]['env_creator'](**self.env_config)
self.sess = sess
self.steps_num = self.config['steps_num']
self.obs_act_rew = deque([], maxlen=self.steps_num)
self.is_prioritized = config['replay_buffer_type'] != 'normal'
self.atoms_num = self.config['atoms_num']
assert self.atoms_num == 1
if central_state_space is not None:
self.state_shape = central_state_space.shape
else:
raise NotImplementedError(
"central_state_space input to VDN is NONE!")
self.n_agents = self.env.env_info['n_agents']
if not self.is_prioritized:
self.exp_buffer = experience.ReplayBufferCentralState(
config['replay_buffer_size'], observation_space,
central_state_space, self.n_agents)
else:
raise NotImplementedError(
"Not implemented! PrioritizedReplayBuffer with CentralState")
#self.exp_buffer = experience.PrioritizedReplayBufferCentralState(config['replay_buffer_size'], config['priority_alpha'])
#self.sample_weights_ph = tf.placeholder(tf.float32, shape=[None, 1], name='sample_weights')
self.batch_size_ph = tf.placeholder(tf.int32, name='batch_size_ph')
self.obs_ph = tf.placeholder(observation_space.dtype,
shape=(None, ) + self.obs_shape,
name='obs_ph')
self.state_ph = tf.placeholder(observation_space.dtype,
shape=(None, ) + self.state_shape,
name='state_ph')
self.actions_ph = tf.placeholder(tf.int32,
shape=[None, 1],
name='actions_ph')
self.rewards_ph = tf.placeholder(tf.float32,
shape=[None, 1],
name='rewards_ph')
self.next_obs_ph = tf.placeholder(observation_space.dtype,
shape=(None, ) + self.obs_shape,
name='next_obs_ph')
self.is_done_ph = tf.placeholder(tf.float32,
shape=[None, 1],
name='is_done_ph')
self.is_not_done = 1 - self.is_done_ph
self.name = base_name
self.gamma = self.config['gamma']
self.gamma_step = self.gamma**self.steps_num
self.grad_norm = config['grad_norm']
self.input_obs = self.obs_ph
self.input_next_obs = self.next_obs_ph
if observation_space.dtype == np.uint8:
print('scaling obs')
self.input_obs = tf.to_float(self.input_obs) / 255.0
self.input_next_obs = tf.to_float(self.input_next_obs) / 255.0
self.setup_qvalues(actions_num)
self.reg_loss = tf.losses.get_regularization_loss()
self.td_loss_mean += self.reg_loss
self.learning_rate = self.config['learning_rate']
self.train_step = tf.train.AdamOptimizer(
self.learning_rate * self.lr_multiplier
) # .minimize(self.td_loss_mean, var_list=self.weights)
grads = tf.gradients(self.td_loss_mean, self.weights)
if self.config['truncate_grads']:
grads, _ = tf.clip_by_global_norm(grads, self.grad_norm)
grads = list(zip(grads, self.weights))
self.train_op = self.train_step.apply_gradients(grads)
self.saver = tf.train.Saver()
self.assigns_op = [
tf.assign(w_target, w_self, validate_shape=True)
for w_self, w_target in zip(self.weights, self.target_weights)
]
self.variables = TensorFlowVariables(self.qvalues, self.sess)
if self.env_name:
sess.run(tf.global_variables_initializer())
self._reset()
self.logger = logger
self.num_env_steps_train = 0
class VDNAgent:
def __init__(self,
sess,
base_name,
observation_space,
action_space,
config,
logger,
central_state_space=None):
observation_shape = observation_space.shape
actions_num = action_space.n
self.config = config
self.is_adaptive_lr = config['lr_schedule'] == 'adaptive'
self.is_polynom_decay_lr = config['lr_schedule'] == 'polynom_decay'
self.is_exp_decay_lr = config['lr_schedule'] == 'exp_decay'
self.lr_multiplier = tf.constant(1, shape=(), dtype=tf.float32)
self.learning_rate_ph = tf.placeholder('float32', (), name='lr_ph')
self.games_to_track = tr_helpers.get_or_default(
config, 'games_to_track', 100)
self.max_epochs = tr_helpers.get_or_default(self.config, 'max_epochs',
1e6)
self.games_to_log = self.config.get('games_to_track', 100)
self.game_rewards = deque([], maxlen=self.games_to_track)
self.game_lengths = deque([], maxlen=self.games_to_track)
self.game_scores = deque([], maxlen=self.games_to_log)
self.epoch_num = tf.Variable(tf.constant(0, shape=(),
dtype=tf.float32),
trainable=False)
self.update_epoch_op = self.epoch_num.assign(self.epoch_num + 1)
self.current_lr = self.learning_rate_ph
if self.is_adaptive_lr:
self.lr_threshold = config['lr_threshold']
if self.is_polynom_decay_lr:
self.lr_multiplier = tf.train.polynomial_decay(
1.0,
global_step=self.epoch_num,
decay_steps=self.max_epochs,
end_learning_rate=0.001,
power=tr_helpers.get_or_default(config, 'decay_power', 1.0))
if self.is_exp_decay_lr:
self.lr_multiplier = tf.train.exponential_decay(
1.0,
global_step=self.epoch_num,
decay_steps=self.max_epochs,
decay_rate=config['decay_rate'])
self.env_name = config['env_name']
self.network = config['network']
self.batch_size = self.config['batch_size']
self.obs_shape = observation_shape
self.actions_num = actions_num
self.writer = SummaryWriter('runs/' + config['name'] +
datetime.now().strftime("%d, %H:%M:%S"))
self.epsilon = self.config['epsilon']
self.rewards_shaper = self.config['reward_shaper']
self.epsilon_processor = tr_helpers.LinearValueProcessor(
self.config['epsilon'], self.config['min_epsilon'],
self.config['epsilon_decay_frames'])
self.beta_processor = tr_helpers.LinearValueProcessor(
self.config['priority_beta'], self.config['max_beta'],
self.config['beta_decay_frames'])
if self.env_name:
self.env_config = config.get('env_config', {})
self.env = env_configurations.configurations[
self.env_name]['env_creator'](**self.env_config)
self.sess = sess
self.steps_num = self.config['steps_num']
self.obs_act_rew = deque([], maxlen=self.steps_num)
self.is_prioritized = config['replay_buffer_type'] != 'normal'
self.atoms_num = self.config['atoms_num']
assert self.atoms_num == 1
if central_state_space is not None:
self.state_shape = central_state_space.shape
else:
raise NotImplementedError(
"central_state_space input to VDN is NONE!")
self.n_agents = self.env.env_info['n_agents']
if not self.is_prioritized:
self.exp_buffer = experience.ReplayBufferCentralState(
config['replay_buffer_size'], observation_space,
central_state_space, self.n_agents)
else:
raise NotImplementedError(
"Not implemented! PrioritizedReplayBuffer with CentralState")
#self.exp_buffer = experience.PrioritizedReplayBufferCentralState(config['replay_buffer_size'], config['priority_alpha'])
#self.sample_weights_ph = tf.placeholder(tf.float32, shape=[None, 1], name='sample_weights')
self.batch_size_ph = tf.placeholder(tf.int32, name='batch_size_ph')
self.obs_ph = tf.placeholder(observation_space.dtype,
shape=(None, ) + self.obs_shape,
name='obs_ph')
self.state_ph = tf.placeholder(observation_space.dtype,
shape=(None, ) + self.state_shape,
name='state_ph')
self.actions_ph = tf.placeholder(tf.int32,
shape=[None, 1],
name='actions_ph')
self.rewards_ph = tf.placeholder(tf.float32,
shape=[None, 1],
name='rewards_ph')
self.next_obs_ph = tf.placeholder(observation_space.dtype,
shape=(None, ) + self.obs_shape,
name='next_obs_ph')
self.is_done_ph = tf.placeholder(tf.float32,
shape=[None, 1],
name='is_done_ph')
self.is_not_done = 1 - self.is_done_ph
self.name = base_name
self.gamma = self.config['gamma']
self.gamma_step = self.gamma**self.steps_num
self.grad_norm = config['grad_norm']
self.input_obs = self.obs_ph
self.input_next_obs = self.next_obs_ph
if observation_space.dtype == np.uint8:
print('scaling obs')
self.input_obs = tf.to_float(self.input_obs) / 255.0
self.input_next_obs = tf.to_float(self.input_next_obs) / 255.0
self.setup_qvalues(actions_num)
self.reg_loss = tf.losses.get_regularization_loss()
self.td_loss_mean += self.reg_loss
self.learning_rate = self.config['learning_rate']
self.train_step = tf.train.AdamOptimizer(
self.learning_rate * self.lr_multiplier
) # .minimize(self.td_loss_mean, var_list=self.weights)
grads = tf.gradients(self.td_loss_mean, self.weights)
if self.config['truncate_grads']:
grads, _ = tf.clip_by_global_norm(grads, self.grad_norm)
grads = list(zip(grads, self.weights))
self.train_op = self.train_step.apply_gradients(grads)
self.saver = tf.train.Saver()
self.assigns_op = [
tf.assign(w_target, w_self, validate_shape=True)
for w_self, w_target in zip(self.weights, self.target_weights)
]
self.variables = TensorFlowVariables(self.qvalues, self.sess)
if self.env_name:
sess.run(tf.global_variables_initializer())
self._reset()
self.logger = logger
self.num_env_steps_train = 0
def get_weights(self):
return self.variables.get_flat()
def set_weights(self, weights):
return self.variables.set_flat(weights)
def update_epoch(self):
return self.sess.run([self.update_epoch_op])[0]
def setup_qvalues(self, actions_num):
config = {
'input_obs': self.input_obs,
'input_next_obs': self.input_next_obs,
'actions_num': actions_num,
'is_double': self.config['is_double'],
'actions_ph': self.actions_ph,
'batch_size_ph': self.batch_size_ph,
'n_agents': self.n_agents
}
# (bs, n_agents, n_actions), (bs, 1), (bs, 1)
self.qvalues, self.current_action_qvalues_mix, self.target_action_qvalues_mix = self.network(
config)
self.weights = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope='agent')
self.target_weights = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, scope='target')
self.reference_qvalues = self.rewards_ph + self.gamma_step * self.is_not_done * self.target_action_qvalues_mix
if self.is_prioritized:
# we need to return l1 loss to update priority buffer
self.abs_errors = tf.abs(self.current_action_qvalues_mix -
self.reference_qvalues) + 1e-5
# the same as multiply gradients later (other way is used in different examples over internet)
self.td_loss = tf.losses.huber_loss(
self.current_action_qvalues_mix,
self.reference_qvalues,
reduction=tf.losses.Reduction.NONE) * self.sample_weights_ph
self.td_loss_mean = tf.reduce_mean(self.td_loss)
else:
self.td_loss_mean = tf.losses.huber_loss(
self.current_action_qvalues_mix,
self.reference_qvalues,
reduction=tf.losses.Reduction.MEAN)
self.reg_loss = tf.losses.get_regularization_loss()
self.td_loss_mean += self.reg_loss
self.learning_rate = self.config['learning_rate']
if self.env_name:
self.train_step = tf.train.AdamOptimizer(
self.learning_rate * self.lr_multiplier).minimize(
self.td_loss_mean, var_list=self.weights)
def save(self, fn):
self.saver.save(self.sess, fn)
def restore(self, fn):
self.saver.restore(self.sess, fn)
def _reset(self):
self.obs_act_rew.clear()
if self.env_name:
self.current_obs = self.env.reset()
self.total_reward = 0.0
self.total_shaped_reward = 0.0
self.step_count = 0
def get_action(self, obs, avail_acts, epsilon=0.0):
if np.random.random() < epsilon:
action = tfp.distributions.Categorical(
probs=avail_acts.astype(float)).sample().eval(
session=self.sess)
else:
obs = obs.reshape((self.n_agents, ) + self.obs_shape)
# (n_agents, num_actions)
qvals = self.get_qvalues(obs).squeeze(0)
qvals[avail_acts == False] = -9999999
action = np.argmax(qvals, axis=1)
return action
def get_qvalues(self, obs):
return self.sess.run(self.qvalues, {
self.obs_ph: obs,
self.batch_size_ph: 1
})
def play_steps(self, steps, epsilon=0.0):
done_reward = None
done_shaped_reward = None
done_steps = None
done_info = None
steps_rewards = 0
cur_gamma = 1
cur_obs_act_rew_len = len(self.obs_act_rew)
# always break after one
while True:
if cur_obs_act_rew_len > 0:
obs = self.obs_act_rew[-1][0]
else:
obs = self.current_obs
obs = np.reshape(obs, ((self.n_agents, ) + self.obs_shape))
state = self.env.get_state()
action = self.get_action(obs, self.env.get_action_mask(), epsilon)
new_obs, reward, is_done, info = self.env.step(action)
# reward = reward * (1 - is_done)
# Increase step count by 1 - we do not use vec env! (WHIRL)
self.num_env_steps_train += 1
# Same reward, done for all agents
reward = reward[0]
is_done = all(is_done)
state = state[0]
self.step_count += 1
self.total_reward += reward
shaped_reward = self.rewards_shaper(reward)
self.total_shaped_reward += shaped_reward
self.obs_act_rew.append([new_obs, action, shaped_reward, state])
if len(self.obs_act_rew) < steps:
break
for i in range(steps):
sreward = self.obs_act_rew[i][2]
steps_rewards += sreward * cur_gamma
cur_gamma = cur_gamma * self.gamma
next_obs, current_action, _, current_st = self.obs_act_rew[0]
self.exp_buffer.add(self.current_obs, current_action, current_st,
steps_rewards, new_obs, is_done)
self.current_obs = next_obs
break
if is_done:
done_reward = self.total_reward
done_steps = self.step_count
done_shaped_reward = self.total_shaped_reward
done_info = info
self._reset()
return done_reward, done_shaped_reward, done_steps, done_info
def load_weights_into_target_network(self):
self.sess.run(self.assigns_op)
def sample_batch(self, exp_replay, batch_size):
obs_batch, act_batch, st_batch, reward_batch, next_obs_batch, is_done_batch = exp_replay.sample(
batch_size)
obs_batch = obs_batch.reshape((batch_size * self.n_agents, ) +
self.obs_shape)
act_batch = act_batch.reshape((batch_size * self.n_agents, 1))
st_batch = st_batch.reshape((batch_size, ) + self.state_shape)
next_obs_batch = next_obs_batch.reshape((batch_size *
self.n_agents, ) +
self.obs_shape)
reward_batch = reward_batch.reshape((batch_size, 1))
is_done_batch = is_done_batch.reshape((batch_size, 1))
return {
self.obs_ph: obs_batch,
self.actions_ph: act_batch,
self.state_ph: st_batch,
self.rewards_ph: reward_batch,
self.is_done_ph: is_done_batch,
self.next_obs_ph: next_obs_batch,
self.batch_size_ph: batch_size
}
def sample_prioritized_batch(self, exp_replay, batch_size, beta):
obs_batch, act_batch, st_batch, reward_batch, next_obs_batch, is_done_batch, sample_weights, sample_idxes = exp_replay.sample(
batch_size, beta)
obs_batch = obs_batch.reshape((batch_size * self.n_agents, ) +
self.obs_shape)
act_batch = act_batch.reshape((batch_size * self.n_agents, 1))
st_batch = st_batch.reshape((batch_size, ) + self.state_shape)
next_obs_batch = next_obs_batch.reshape((batch_size *
self.n_agents, ) +
self.obs_shape)
reward_batch = reward_batch.reshape((batch_size, 1))
is_done_batch = is_done_batch.reshape((batch_size, 1))
sample_weights = sample_weights.reshape((batch_size, 1))
batch = {
self.obs_ph: obs_batch,
self.actions_ph: act_batch,
self.state_ph: st_batch,
self.rewards_ph: reward_batch,
self.is_done_ph: is_done_batch,
self.next_obs_ph: next_obs_batch,
self.sample_weights_ph: sample_weights,
self.batch_size_ph: batch_size
}
return [batch, sample_idxes]
def train(self):
mem_free_steps = 0
last_mean_rewards = -100500
epoch_num = 0
frame = 0
update_time = 0
play_time = 0
start_time = time.time()
total_time = 0
self.load_weights_into_target_network()
for _ in range(0, self.config['num_steps_fill_buffer']):
self.play_steps(self.steps_num, self.epsilon)
steps_per_epoch = self.config['steps_per_epoch']
num_epochs_to_copy = self.config['num_epochs_to_copy']
frame = 0
play_time = 0
update_time = 0
rewards = []
shaped_rewards = []
steps = []
losses = deque([], maxlen=100)
while True:
epoch_num = self.update_epoch()
t_play_start = time.time()
self.epsilon = self.epsilon_processor(frame)
self.beta = self.beta_processor(frame)
for _ in range(0, steps_per_epoch):
reward, shaped_reward, step, info = self.play_steps(
self.steps_num, self.epsilon)
if reward != None:
self.game_lengths.append(step)
self.game_rewards.append(reward)
game_res = info.get('battle_won', 0.5)
self.game_scores.append(game_res)
# shaped_rewards.append(shaped_reward)
t_play_end = time.time()
play_time += t_play_end - t_play_start
# train
frame = frame + steps_per_epoch
t_start = time.time()
if self.is_prioritized:
batch, idxes = self.sample_prioritized_batch(
self.exp_buffer,
batch_size=self.batch_size,
beta=self.beta)
_, loss_t, errors_update, lr_mul = self.sess.run([
self.train_op, self.td_loss_mean, self.abs_errors,
self.lr_multiplier
], batch)
self.exp_buffer.update_priorities(idxes, errors_update)
else:
batch = self.sample_batch(self.exp_buffer,
batch_size=self.batch_size)
_, loss_t, lr_mul = self.sess.run(
[self.train_op, self.td_loss_mean, self.lr_multiplier],
batch)
losses.append(loss_t)
t_end = time.time()
update_time += t_end - t_start
total_time += update_time
if frame % 1000 == 0:
mem_free_steps += 1
if mem_free_steps == 10:
mem_free_steps = 0
tr_helpers.free_mem()
sum_time = update_time + play_time
print('frames per seconds: ', 1000 / (sum_time))
self.writer.add_scalar('performance/fps', 1000 / sum_time,
frame)
self.writer.add_scalar('performance/upd_time', update_time,
frame)
self.writer.add_scalar('performance/play_time', play_time,
frame)
self.writer.add_scalar('losses/td_loss', np.mean(losses),
frame)
self.writer.add_scalar('info/lr_mul', lr_mul, frame)
self.writer.add_scalar('info/lr', self.learning_rate * lr_mul,
frame)
self.writer.add_scalar('info/epochs', epoch_num, frame)
self.writer.add_scalar('info/epsilon', self.epsilon, frame)
self.logger.log_stat("whirl/performance/fps", 1000 / sum_time,
self.num_env_steps_train)
self.logger.log_stat("whirl/performance/upd_time", update_time,
self.num_env_steps_train)
self.logger.log_stat("whirl/performance/play_time", play_time,
self.num_env_steps_train)
self.logger.log_stat("losses/td_loss", np.mean(losses),
self.num_env_steps_train)
self.logger.log_stat("whirl/info/last_lr",
self.learning_rate * lr_mul,
self.num_env_steps_train)
self.logger.log_stat("whirl/info/lr_mul", lr_mul,
self.num_env_steps_train)
self.logger.log_stat("whirl/epochs", epoch_num,
self.num_env_steps_train)
self.logger.log_stat("whirl/epsilon", self.epsilon,
self.num_env_steps_train)
if self.is_prioritized:
self.writer.add_scalar('beta', self.beta, frame)
update_time = 0
play_time = 0
num_games = len(self.game_rewards)
if num_games > 10:
mean_rewards = np.sum(self.game_rewards) / num_games
mean_lengths = np.sum(self.game_lengths) / num_games
mean_scores = np.mean(self.game_scores)
self.writer.add_scalar('rewards/mean', mean_rewards, frame)
self.writer.add_scalar('rewards/time', mean_rewards,
total_time)
self.writer.add_scalar('episode_lengths/mean',
mean_lengths, frame)
self.writer.add_scalar('episode_lengths/time',
mean_lengths, total_time)
self.logger.log_stat("whirl/rewards/mean",
np.asscalar(mean_rewards),
self.num_env_steps_train)
self.logger.log_stat("whirl/rewards/time", mean_rewards,
total_time)
self.logger.log_stat("whirl/episode_lengths/mean",
np.asscalar(mean_lengths),
self.num_env_steps_train)
self.logger.log_stat("whirl/episode_lengths/time",
mean_lengths, total_time)
self.logger.log_stat("whirl/win_rate/mean",
np.asscalar(mean_scores),
self.num_env_steps_train)
self.logger.log_stat("whirl/win_rate/time",
np.asscalar(mean_scores), total_time)
if mean_rewards > last_mean_rewards:
print('saving next best rewards: ', mean_rewards)
last_mean_rewards = mean_rewards
self.save("./nn/" + self.config['name'] + 'ep=' +
str(epoch_num) + 'rew=' + str(mean_rewards))
if last_mean_rewards > self.config['score_to_win']:
print('network won!')
return last_mean_rewards, epoch_num
if frame % num_epochs_to_copy == 0:
self.load_weights_into_target_network()
if epoch_num >= self.max_epochs:
print('Max epochs reached')
self.save(
"./nn/" + 'last_' + self.config['name'] + 'ep=' +
str(epoch_num) + 'rew=' +
str(np.sum(self.game_rewards) / len(self.game_rewards)))
return last_mean_rewards, epoch_num
def __init__(self, sess, base_name, observation_space, action_space,
config):
observation_shape = observation_space.shape
self.use_action_masks = config.get('use_action_masks', False)
self.is_train = config.get('is_train', True)
self.self_play = config.get('self_play', False)
self.name = base_name
self.config = config
self.env_name = config['env_name']
self.ppo = config['ppo']
self.is_adaptive_lr = config['lr_schedule'] == 'adaptive'
self.is_polynom_decay_lr = config['lr_schedule'] == 'polynom_decay'
self.is_exp_decay_lr = config['lr_schedule'] == 'exp_decay'
self.lr_multiplier = tf.constant(1, shape=(), dtype=tf.float32)
self.epoch_num = tf.Variable(tf.constant(0, shape=(),
dtype=tf.float32),
trainable=False)
self.e_clip = config['e_clip']
self.clip_value = config['clip_value']
self.network = config['network']
self.rewards_shaper = config['reward_shaper']
self.num_actors = config['num_actors']
self.env_config = self.config.get('env_config', {})
self.vec_env = vecenv.create_vec_env(self.env_name, self.num_actors,
**self.env_config)
self.num_agents = self.vec_env.get_number_of_agents()
self.steps_num = config['steps_num']
self.seq_len = self.config['seq_len']
self.normalize_advantage = config['normalize_advantage']
self.normalize_input = self.config['normalize_input']
self.state_shape = observation_shape
self.critic_coef = config['critic_coef']
self.writer = SummaryWriter('runs/' + config['name'] +
datetime.now().strftime("%d, %H:%M:%S"))
self.sess = sess
self.grad_norm = config['grad_norm']
self.gamma = self.config['gamma']
self.tau = self.config['tau']
self.ignore_dead_batches = self.config.get('ignore_dead_batches',
False)
self.dones = np.asarray([False] * self.num_actors * self.num_agents,
dtype=np.bool)
self.current_rewards = np.asarray([0] * self.num_actors *
self.num_agents,
dtype=np.float32)
self.current_lengths = np.asarray([0] * self.num_actors *
self.num_agents,
dtype=np.float32)
self.games_to_log = self.config.get('games_to_track', 100)
self.game_rewards = deque([], maxlen=self.games_to_log)
self.game_lengths = deque([], maxlen=self.games_to_log)
self.game_scores = deque([], maxlen=self.games_to_log)
self.obs_ph = tf.placeholder(observation_space.dtype,
(None, ) + observation_shape,
name='obs')
self.target_obs_ph = tf.placeholder(observation_space.dtype,
(None, ) + observation_shape,
name='target_obs')
self.actions_num = action_space.n
self.actions_ph = tf.placeholder('int32', (None, ), name='actions')
if self.use_action_masks:
self.action_mask_ph = tf.placeholder('int32',
(None, self.actions_num),
name='actions_mask')
else:
self.action_mask_ph = None
self.old_logp_actions_ph = tf.placeholder('float32', (None, ),
name='old_logpactions')
self.rewards_ph = tf.placeholder('float32', (None, ), name='rewards')
self.old_values_ph = tf.placeholder('float32', (None, ),
name='old_values')
self.advantages_ph = tf.placeholder('float32', (None, ),
name='advantages')
self.learning_rate_ph = tf.placeholder('float32', (), name='lr_ph')
self.update_epoch_op = self.epoch_num.assign(self.epoch_num + 1)
self.current_lr = self.learning_rate_ph
self.input_obs = self.obs_ph
self.input_target_obs = self.target_obs_ph
if observation_space.dtype == np.uint8:
self.input_obs = tf.to_float(self.input_obs) / 255.0
self.input_target_obs = tf.to_float(self.input_target_obs) / 255.0
if self.is_adaptive_lr:
self.lr_threshold = config['lr_threshold']
if self.is_polynom_decay_lr:
self.lr_multiplier = tf.train.polynomial_decay(
1.0,
self.epoch_num,
config['max_epochs'],
end_learning_rate=0.001,
power=tr_helpers.get_or_default(config, 'decay_power', 1.0))
if self.is_exp_decay_lr:
self.lr_multiplier = tf.train.exponential_decay(
1.0,
self.epoch_num,
config['max_epochs'],
decay_rate=config['decay_rate'])
if self.normalize_input:
self.moving_mean_std = MovingMeanStd(shape=observation_space.shape,
epsilon=1e-5,
decay=0.99)
self.input_obs = self.moving_mean_std.normalize(self.input_obs,
train=True)
self.input_target_obs = self.moving_mean_std.normalize(
self.input_target_obs, train=False)
games_num = self.config[
'minibatch_size'] // self.seq_len # it is used only for current rnn implementation
self.train_dict = {
'name': 'agent',
'inputs': self.input_obs,
'batch_num': self.config['minibatch_size'],
'games_num': games_num,
'actions_num': self.actions_num,
'prev_actions_ph': self.actions_ph,
'action_mask_ph': None
}
self.run_dict = {
'name': 'agent',
'inputs': self.input_target_obs,
'batch_num': self.num_actors * self.num_agents,
'games_num': self.num_actors * self.num_agents,
'actions_num': self.actions_num,
'prev_actions_ph': None,
'action_mask_ph': self.action_mask_ph
}
self.states = None
if self.network.is_rnn():
self.logp_actions, self.state_values, self.action, self.entropy, self.states_ph, self.masks_ph, self.lstm_state, self.initial_state = self.network(
self.train_dict, reuse=False)
self.target_neglogp, self.target_state_values, self.target_action, _, self.target_states_ph, self.target_masks_ph, self.target_lstm_state, self.target_initial_state, self.logits = self.network(
self.run_dict, reuse=True)
self.states = self.target_initial_state
else:
self.logp_actions, self.state_values, self.action, self.entropy = self.network(
self.train_dict, reuse=False)
self.target_neglogp, self.target_state_values, self.target_action, _, self.logits = self.network(
self.run_dict, reuse=True)
self.saver = tf.train.Saver()
self.variables = TensorFlowVariables([
self.target_action, self.target_state_values, self.target_neglogp
], self.sess)
if self.is_train:
self.setup_losses()
self.sess.run(tf.global_variables_initializer())
class A2CAgent:
def __init__(self, sess, base_name, observation_space, action_space,
config):
observation_shape = observation_space.shape
self.use_action_masks = config.get('use_action_masks', False)
self.is_train = config.get('is_train', True)
self.self_play = config.get('self_play', False)
self.name = base_name
self.config = config
self.env_name = config['env_name']
self.ppo = config['ppo']
self.is_adaptive_lr = config['lr_schedule'] == 'adaptive'
self.is_polynom_decay_lr = config['lr_schedule'] == 'polynom_decay'
self.is_exp_decay_lr = config['lr_schedule'] == 'exp_decay'
self.lr_multiplier = tf.constant(1, shape=(), dtype=tf.float32)
self.epoch_num = tf.Variable(tf.constant(0, shape=(),
dtype=tf.float32),
trainable=False)
self.e_clip = config['e_clip']
self.clip_value = config['clip_value']
self.network = config['network']
self.rewards_shaper = config['reward_shaper']
self.num_actors = config['num_actors']
self.env_config = self.config.get('env_config', {})
self.vec_env = vecenv.create_vec_env(self.env_name, self.num_actors,
**self.env_config)
self.num_agents = self.vec_env.get_number_of_agents()
self.steps_num = config['steps_num']
self.seq_len = self.config['seq_len']
self.normalize_advantage = config['normalize_advantage']
self.normalize_input = self.config['normalize_input']
self.state_shape = observation_shape
self.critic_coef = config['critic_coef']
self.writer = SummaryWriter('runs/' + config['name'] +
datetime.now().strftime("%d, %H:%M:%S"))
self.sess = sess
self.grad_norm = config['grad_norm']
self.gamma = self.config['gamma']
self.tau = self.config['tau']
self.ignore_dead_batches = self.config.get('ignore_dead_batches',
False)
self.dones = np.asarray([False] * self.num_actors * self.num_agents,
dtype=np.bool)
self.current_rewards = np.asarray([0] * self.num_actors *
self.num_agents,
dtype=np.float32)
self.current_lengths = np.asarray([0] * self.num_actors *
self.num_agents,
dtype=np.float32)
self.games_to_log = self.config.get('games_to_track', 100)
self.game_rewards = deque([], maxlen=self.games_to_log)
self.game_lengths = deque([], maxlen=self.games_to_log)
self.game_scores = deque([], maxlen=self.games_to_log)
self.obs_ph = tf.placeholder(observation_space.dtype,
(None, ) + observation_shape,
name='obs')
self.target_obs_ph = tf.placeholder(observation_space.dtype,
(None, ) + observation_shape,
name='target_obs')
self.actions_num = action_space.n
self.actions_ph = tf.placeholder('int32', (None, ), name='actions')
if self.use_action_masks:
self.action_mask_ph = tf.placeholder('int32',
(None, self.actions_num),
name='actions_mask')
else:
self.action_mask_ph = None
self.old_logp_actions_ph = tf.placeholder('float32', (None, ),
name='old_logpactions')
self.rewards_ph = tf.placeholder('float32', (None, ), name='rewards')
self.old_values_ph = tf.placeholder('float32', (None, ),
name='old_values')
self.advantages_ph = tf.placeholder('float32', (None, ),
name='advantages')
self.learning_rate_ph = tf.placeholder('float32', (), name='lr_ph')
self.update_epoch_op = self.epoch_num.assign(self.epoch_num + 1)
self.current_lr = self.learning_rate_ph
self.input_obs = self.obs_ph
self.input_target_obs = self.target_obs_ph
if observation_space.dtype == np.uint8:
self.input_obs = tf.to_float(self.input_obs) / 255.0
self.input_target_obs = tf.to_float(self.input_target_obs) / 255.0
if self.is_adaptive_lr:
self.lr_threshold = config['lr_threshold']
if self.is_polynom_decay_lr:
self.lr_multiplier = tf.train.polynomial_decay(
1.0,
self.epoch_num,
config['max_epochs'],
end_learning_rate=0.001,
power=tr_helpers.get_or_default(config, 'decay_power', 1.0))
if self.is_exp_decay_lr:
self.lr_multiplier = tf.train.exponential_decay(
1.0,
self.epoch_num,
config['max_epochs'],
decay_rate=config['decay_rate'])
if self.normalize_input:
self.moving_mean_std = MovingMeanStd(shape=observation_space.shape,
epsilon=1e-5,
decay=0.99)
self.input_obs = self.moving_mean_std.normalize(self.input_obs,
train=True)
self.input_target_obs = self.moving_mean_std.normalize(
self.input_target_obs, train=False)
games_num = self.config[
'minibatch_size'] // self.seq_len # it is used only for current rnn implementation
self.train_dict = {
'name': 'agent',
'inputs': self.input_obs,
'batch_num': self.config['minibatch_size'],
'games_num': games_num,
'actions_num': self.actions_num,
'prev_actions_ph': self.actions_ph,
'action_mask_ph': None
}
self.run_dict = {
'name': 'agent',
'inputs': self.input_target_obs,
'batch_num': self.num_actors * self.num_agents,
'games_num': self.num_actors * self.num_agents,
'actions_num': self.actions_num,
'prev_actions_ph': None,
'action_mask_ph': self.action_mask_ph
}
self.states = None
if self.network.is_rnn():
self.logp_actions, self.state_values, self.action, self.entropy, self.states_ph, self.masks_ph, self.lstm_state, self.initial_state = self.network(
self.train_dict, reuse=False)
self.target_neglogp, self.target_state_values, self.target_action, _, self.target_states_ph, self.target_masks_ph, self.target_lstm_state, self.target_initial_state, self.logits = self.network(
self.run_dict, reuse=True)
self.states = self.target_initial_state
else:
self.logp_actions, self.state_values, self.action, self.entropy = self.network(
self.train_dict, reuse=False)
self.target_neglogp, self.target_state_values, self.target_action, _, self.logits = self.network(
self.run_dict, reuse=True)
self.saver = tf.train.Saver()
self.variables = TensorFlowVariables([
self.target_action, self.target_state_values, self.target_neglogp
], self.sess)
if self.is_train:
self.setup_losses()
self.sess.run(tf.global_variables_initializer())
def setup_losses(self):
curr_e_clip = self.e_clip * self.lr_multiplier
if (self.ppo):
self.prob_ratio = tf.exp(self.old_logp_actions_ph -
self.logp_actions)
self.pg_loss_unclipped = -tf.multiply(self.advantages_ph,
self.prob_ratio)
self.pg_loss_clipped = -tf.multiply(
self.advantages_ph,
tf.clip_by_value(self.prob_ratio, 1. - curr_e_clip,
1. + curr_e_clip))
self.actor_loss = tf.maximum(self.pg_loss_unclipped,
self.pg_loss_clipped)
else:
self.actor_loss = self.logp_actions * self.advantages_ph
self.actor_loss = tf.reduce_mean(self.actor_loss)
self.c_loss = (tf.squeeze(self.state_values) - self.rewards_ph)**2
if self.clip_value:
self.cliped_values = self.old_values_ph + tf.clip_by_value(
tf.squeeze(self.state_values) - self.old_values_ph,
-curr_e_clip, curr_e_clip)
self.c_loss_clipped = tf.square(self.cliped_values -
self.rewards_ph)
self.critic_loss = tf.maximum(self.c_loss, self.c_loss_clipped)
else:
self.critic_loss = self.c_loss
self.critic_loss = tf.reduce_mean(self.critic_loss)
self.kl_approx = 0.5 * tf.stop_gradient(
tf.reduce_mean((self.old_logp_actions_ph - self.logp_actions)**2))
if self.is_adaptive_lr:
self.current_lr = tf.where(
self.kl_approx > (2.0 * self.lr_threshold),
tf.maximum(self.current_lr / 1.5, 1e-6), self.current_lr)
self.current_lr = tf.where(
self.kl_approx < (0.5 * self.lr_threshold),
tf.minimum(self.current_lr * 1.5, 1e-2), self.current_lr)
self.loss = self.actor_loss + 0.5 * self.critic_coef * self.critic_loss - self.config[
'entropy_coef'] * self.entropy
self.reg_loss = tf.losses.get_regularization_loss()
self.loss += self.reg_loss
self.train_step = tf.train.AdamOptimizer(self.current_lr *
self.lr_multiplier)
self.weights = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope='agent')
grads = tf.gradients(self.loss, self.weights)
if self.config['truncate_grads']:
grads, _ = tf.clip_by_global_norm(grads, self.grad_norm)
grads = list(zip(grads, self.weights))
self.train_op = self.train_step.apply_gradients(grads)
def update_epoch(self):
return self.sess.run([self.update_epoch_op])[0]
def get_action_values(self, obs):
run_ops = [
self.target_action, self.target_state_values, self.target_neglogp
]
if self.network.is_rnn():
run_ops.append(self.target_lstm_state)
return self.sess.run(
run_ops, {
self.target_obs_ph: obs,
self.target_states_ph: self.states,
self.target_masks_ph: self.dones
})
else:
return (*self.sess.run(run_ops, {self.target_obs_ph: obs}), None)
def get_masked_action_values(self, obs, action_masks):
run_ops = [
self.target_action, self.target_state_values, self.target_neglogp,
self.logits
]
if self.network.is_rnn():
run_ops.append(self.target_lstm_state)
return self.sess.run(
run_ops, {
self.action_mask_ph: action_masks,
self.target_obs_ph: obs,
self.target_states_ph: self.states,
self.target_masks_ph: self.dones
})
else:
return (*self.sess.run(run_ops, {
self.action_mask_ph: action_masks,
self.target_obs_ph: obs
}), None)
def get_values(self, obs):
if self.network.is_rnn():
return self.sess.run(
[self.target_state_values], {
self.target_obs_ph: obs,
self.target_states_ph: self.states,
self.target_masks_ph: self.dones
})
else:
return self.sess.run([self.target_state_values],
{self.target_obs_ph: obs})
def get_weights(self):
return self.variables.get_flat()
def set_weights(self, weights):
return self.variables.set_flat(weights)
def play_steps(self):
# here, we init the lists that will contain the mb of experiences
mb_obs, mb_rewards, mb_actions, mb_values, mb_dones, mb_neglogpacs = [],[],[],[],[],[]
mb_states = []
epinfos = []
# for n in range number of steps
for _ in range(self.steps_num):
if self.network.is_rnn():
mb_states.append(self.states)
if self.use_action_masks:
masks = self.vec_env.get_action_masks()
if self.use_action_masks:
actions, values, neglogpacs, _, self.states = self.get_masked_action_values(
self.obs, masks)
else:
actions, values, neglogpacs, self.states = self.get_action_values(
self.obs)
actions = np.squeeze(actions)
values = np.squeeze(values)
neglogpacs = np.squeeze(neglogpacs)
mb_obs.append(self.obs.copy())
mb_actions.append(actions)
mb_values.append(values)
mb_neglogpacs.append(neglogpacs)
mb_dones.append(self.dones.copy())
self.obs[:], rewards, self.dones, infos = self.vec_env.step(
actions)
self.current_rewards += rewards
self.current_lengths += 1
for reward, length, done, info in zip(
self.current_rewards[::self.num_agents],
self.current_lengths[::self.num_agents],
self.dones[::self.num_agents], infos):
if done:
self.game_rewards.append(reward)
self.game_lengths.append(length)
game_res = info.get('battle_won', 0.5)
self.game_scores.append(game_res)
self.current_rewards = self.current_rewards * (1.0 - self.dones)
self.current_lengths = self.current_lengths * (1.0 - self.dones)
shaped_rewards = self.rewards_shaper(rewards)
epinfos.append(infos)
mb_rewards.append(shaped_rewards)
#using openai baseline approach
mb_obs = np.asarray(mb_obs, dtype=self.obs.dtype)
mb_rewards = np.asarray(mb_rewards, dtype=np.float32)
mb_actions = np.asarray(mb_actions, dtype=np.float32)
mb_values = np.asarray(mb_values, dtype=np.float32)
mb_neglogpacs = np.asarray(mb_neglogpacs, dtype=np.float32)
mb_dones = np.asarray(mb_dones, dtype=np.bool)
mb_states = np.asarray(mb_states, dtype=np.float32)
last_values = self.get_values(self.obs)
last_values = np.squeeze(last_values)
mb_returns = np.zeros_like(mb_rewards)
mb_advs = np.zeros_like(mb_rewards)
lastgaelam = 0
for t in reversed(range(self.steps_num)):
if t == self.steps_num - 1:
nextnonterminal = 1.0 - self.dones
nextvalues = last_values
else:
nextnonterminal = 1.0 - mb_dones[t + 1]
nextvalues = mb_values[t + 1]
delta = mb_rewards[
t] + self.gamma * nextvalues * nextnonterminal - mb_values[t]
mb_advs[
t] = lastgaelam = delta + self.gamma * self.tau * nextnonterminal * lastgaelam
mb_returns = mb_advs + mb_values
if self.network.is_rnn():
result = (*map(swap_and_flatten01,
(mb_obs, mb_returns, mb_dones, mb_actions,
mb_values, mb_neglogpacs, mb_states)), epinfos)
else:
result = (*map(swap_and_flatten01,
(mb_obs, mb_returns, mb_dones, mb_actions,
mb_values, mb_neglogpacs)), None, epinfos)
return result
def save(self, fn):
self.saver.save(self.sess, fn)
def restore(self, fn):
self.saver.restore(self.sess, fn)
def train(self):
self.obs = self.vec_env.reset()
batch_size = self.steps_num * self.num_actors * self.num_agents
batch_size_envs = self.steps_num * self.num_actors
minibatch_size = self.config['minibatch_size']
mini_epochs_num = self.config['mini_epochs']
num_minibatches = batch_size // minibatch_size
last_lr = self.config['learning_rate']
frame = 0
update_time = 0
last_mean_rewards = -100500
play_time = 0
epoch_num = 0
max_epochs = self.config.get('max_epochs', 1e6)
start_time = time.time()
total_time = 0
rep_count = 0
while True:
play_time_start = time.time()
epoch_num = self.update_epoch()
frame += batch_size_envs
obses, returns, dones, actions, values, neglogpacs, lstm_states, _ = self.play_steps(
)
advantages = returns - values
if self.normalize_advantage:
advantages = (advantages -
advantages.mean()) / (advantages.std() + 1e-8)
a_losses = []
c_losses = []
entropies = []
kls = []
play_time_end = time.time()
play_time = play_time_end - play_time_start
update_time_start = time.time()
if self.network.is_rnn():
total_games = batch_size // self.seq_len
num_games_batch = minibatch_size // self.seq_len
game_indexes = np.arange(total_games)
flat_indexes = np.arange(total_games * self.seq_len).reshape(
total_games, self.seq_len)
lstm_states = lstm_states[::self.seq_len]
for _ in range(0, mini_epochs_num):
np.random.shuffle(game_indexes)
for i in range(0, num_minibatches):
batch = range(i * num_games_batch,
(i + 1) * num_games_batch)
mb_indexes = game_indexes[batch]
mbatch = flat_indexes[mb_indexes].ravel()
dict = {}
dict[self.old_values_ph] = values[mbatch]
dict[self.old_logp_actions_ph] = neglogpacs[mbatch]
dict[self.advantages_ph] = advantages[mbatch]
dict[self.rewards_ph] = returns[mbatch]
dict[self.actions_ph] = actions[mbatch]
dict[self.obs_ph] = obses[mbatch]
dict[self.masks_ph] = dones[mbatch]
dict[self.states_ph] = lstm_states[batch]
dict[self.learning_rate_ph] = last_lr
run_ops = [
self.actor_loss, self.critic_loss, self.entropy,
self.kl_approx, self.current_lr,
self.lr_multiplier, self.train_op
]
run_ops.append(
tf.get_collection(tf.GraphKeys.UPDATE_OPS))
a_loss, c_loss, entropy, kl, last_lr, lr_mul, _, _ = self.sess.run(
run_ops, dict)
a_losses.append(a_loss)
c_losses.append(c_loss)
kls.append(kl)
entropies.append(entropy)
else:
for _ in range(0, mini_epochs_num):
permutation = np.random.permutation(batch_size)
obses = obses[permutation]
returns = returns[permutation]
actions = actions[permutation]
values = values[permutation]
neglogpacs = neglogpacs[permutation]
advantages = advantages[permutation]
for i in range(0, num_minibatches):
batch = range(i * minibatch_size,
(i + 1) * minibatch_size)
dict = {
self.obs_ph: obses[batch],
self.actions_ph: actions[batch],
self.rewards_ph: returns[batch],
self.advantages_ph: advantages[batch],
self.old_logp_actions_ph: neglogpacs[batch],
self.old_values_ph: values[batch]
}
dict[self.learning_rate_ph] = last_lr
run_ops = [
self.actor_loss, self.critic_loss, self.entropy,
self.kl_approx, self.current_lr,
self.lr_multiplier, self.train_op
]
run_ops.append(
tf.get_collection(tf.GraphKeys.UPDATE_OPS))
a_loss, c_loss, entropy, kl, last_lr, lr_mul, _, _ = self.sess.run(
run_ops, dict)
a_losses.append(a_loss)
c_losses.append(c_loss)
kls.append(kl)
entropies.append(entropy)
update_time_end = time.time()
update_time = update_time_end - update_time_start
sum_time = update_time + play_time
total_time = update_time_end - start_time
if True:
scaled_time = self.num_agents * sum_time
print('frames per seconds: ', batch_size / scaled_time)
self.writer.add_scalar('performance/fps',
batch_size / scaled_time, frame)
self.writer.add_scalar('performance/upd_time', update_time,
frame)
self.writer.add_scalar('performance/play_time', play_time,
frame)
self.writer.add_scalar('losses/a_loss', np.mean(a_losses),
frame)
self.writer.add_scalar('losses/c_loss', np.mean(c_losses),
frame)
self.writer.add_scalar('losses/entropy', np.mean(entropies),
frame)
self.writer.add_scalar('info/last_lr', last_lr * lr_mul, frame)
self.writer.add_scalar('info/lr_mul', lr_mul, frame)
self.writer.add_scalar('info/e_clip', self.e_clip * lr_mul,
frame)
self.writer.add_scalar('info/kl', np.mean(kls), frame)
self.writer.add_scalar('epochs', epoch_num, frame)
if len(self.game_rewards) > 0:
mean_rewards = np.mean(self.game_rewards)
mean_lengths = np.mean(self.game_lengths)
mean_scores = np.mean(self.game_scores)
self.writer.add_scalar('rewards/mean', mean_rewards, frame)
self.writer.add_scalar('rewards/time', mean_rewards,
total_time)
self.writer.add_scalar('episode_lengths/mean',
mean_lengths, frame)
self.writer.add_scalar('episode_lengths/time',
mean_lengths, total_time)
self.writer.add_scalar('win_rate/mean', mean_scores, frame)
self.writer.add_scalar('win_rate/time', mean_scores,
total_time)
if rep_count % 10 == 0:
self.save("./nn/" + 'last_' + self.config['name'] +
'ep=' + str(epoch_num) + 'rew=' +
str(mean_rewards))
rep_count += 1
if mean_rewards > last_mean_rewards:
print('saving next best rewards: ', mean_rewards)
last_mean_rewards = mean_rewards
self.save("./nn/" + self.config['name'] +
self.env_name)
if last_mean_rewards > self.config['score_to_win']:
print('Network won!')
self.save("./nn/" + self.config['name'] + 'ep=' +
str(epoch_num) + 'rew=' +
str(mean_rewards))
return last_mean_rewards, epoch_num
if epoch_num > max_epochs:
self.save("./nn/" + 'last_' + self.config['name'] + 'ep=' +
str(epoch_num) + 'rew=' + str(mean_rewards))
print('MAX EPOCHS NUM!')
return last_mean_rewards, epoch_num
update_time = 0
class DQNAgent:
def __init__(self,
sess,
base_name,
observation_space,
action_space,
config,
logger,
central_state_space=None):
observation_shape = observation_space.shape
actions_num = action_space.n
self.config = config
self.is_adaptive_lr = config['lr_schedule'] == 'adaptive'
self.is_polynom_decay_lr = config['lr_schedule'] == 'polynom_decay'
self.is_exp_decay_lr = config['lr_schedule'] == 'exp_decay'
self.lr_multiplier = tf.constant(1, shape=(), dtype=tf.float32)
self.learning_rate_ph = tf.placeholder('float32', (), name='lr_ph')
self.games_to_track = tr_helpers.get_or_default(
config, 'games_to_track', 100)
self.max_epochs = tr_helpers.get_or_default(self.config, 'max_epochs',
1e6)
self.game_rewards = deque([], maxlen=self.games_to_track)
self.game_lengths = deque([], maxlen=self.games_to_track)
self.epoch_num = tf.Variable(tf.constant(0, shape=(),
dtype=tf.float32),
trainable=False)
self.update_epoch_op = self.epoch_num.assign(self.epoch_num + 1)
self.current_lr = self.learning_rate_ph
if self.is_adaptive_lr:
self.lr_threshold = config['lr_threshold']
if self.is_polynom_decay_lr:
self.lr_multiplier = tf.train.polynomial_decay(
1.0,
global_step=self.epoch_num,
decay_steps=self.max_epochs,
end_learning_rate=0.001,
power=tr_helpers.get_or_default(config, 'decay_power', 1.0))
if self.is_exp_decay_lr:
self.lr_multiplier = tf.train.exponential_decay(
1.0,
global_step=self.epoch_num,
decay_steps=self.max_epochs,
decay_rate=config['decay_rate'])
self.env_name = config['env_name']
self.network = config['network']
self.state_shape = observation_shape
self.actions_num = actions_num
self.writer = SummaryWriter('runs/' + config['name'] +
datetime.now().strftime("%d, %H:%M:%S"))
self.epsilon = self.config['epsilon']
self.rewards_shaper = self.config['reward_shaper']
self.epsilon_processor = tr_helpers.LinearValueProcessor(
self.config['epsilon'], self.config['min_epsilon'],
self.config['epsilon_decay_frames'])
self.beta_processor = tr_helpers.LinearValueProcessor(
self.config['priority_beta'], self.config['max_beta'],
self.config['beta_decay_frames'])
if self.env_name:
self.env = env_configurations.configurations[
self.env_name]['env_creator'](name=config['name'])
self.sess = sess
self.steps_num = self.config['steps_num']
self.states = deque([], maxlen=self.steps_num)
self.is_prioritized = config['replay_buffer_type'] != 'normal'
self.atoms_num = self.config['atoms_num']
self.is_categorical = self.atoms_num > 1
if self.is_categorical:
self.v_min = self.config['v_min']
self.v_max = self.config['v_max']
self.delta_z = (self.v_max - self.v_min) / (self.atoms_num - 1)
self.all_z = tf.range(self.v_min, self.v_max + self.delta_z,
self.delta_z)
self.categorical = CategoricalQ(self.atoms_num, self.v_min,
self.v_max)
self.n_agents = self.env.env_info['n_agents']
if not self.is_prioritized:
self.exp_buffer = experience.ReplayBuffer(
config['replay_buffer_size'], observation_space, self.n_agents)
else:
self.exp_buffer = experience.PrioritizedReplayBuffer(
config['replay_buffer_size'], config['priority_alpha'],
observation_space, self.n_agents)
self.sample_weights_ph = tf.placeholder(tf.float32,
shape=[
None,
],
name='sample_weights')
self.obs_ph = tf.placeholder(observation_space.dtype,
shape=(None, ) + self.state_shape,
name='obs_ph')
self.actions_ph = tf.placeholder(tf.int32,
shape=[
None,
],
name='actions_ph')
self.rewards_ph = tf.placeholder(tf.float32,
shape=[
None,
],
name='rewards_ph')
self.next_obs_ph = tf.placeholder(observation_space.dtype,
shape=(None, ) + self.state_shape,
name='next_obs_ph')
self.is_done_ph = tf.placeholder(tf.float32,
shape=[
None,
],
name='is_done_ph')
self.is_not_done = 1 - self.is_done_ph
self.name = base_name
self.gamma = self.config['gamma']
self.gamma_step = self.gamma**self.steps_num
self.input_obs = self.obs_ph
self.input_next_obs = self.next_obs_ph
if observation_space.dtype == np.uint8:
print('scaling obs')
self.input_obs = tf.to_float(self.input_obs) / 255.0
self.input_next_obs = tf.to_float(self.input_next_obs) / 255.0
if self.atoms_num == 1:
self.setup_qvalues(actions_num)
else:
self.setup_cat_qvalues(actions_num)
self.reg_loss = tf.losses.get_regularization_loss()
self.td_loss_mean += self.reg_loss
self.learning_rate = self.config['learning_rate']
self.train_step = tf.train.AdamOptimizer(self.learning_rate *
self.lr_multiplier).minimize(
self.td_loss_mean,
var_list=self.weights)
self.saver = tf.train.Saver()
self.assigns_op = [
tf.assign(w_target, w_self, validate_shape=True)
for w_self, w_target in zip(self.weights, self.target_weights)
]
self.variables = TensorFlowVariables(self.qvalues, self.sess)
if self.env_name:
sess.run(tf.global_variables_initializer())
self._reset()
def _get_q(self, probs):
res = probs * self.all_z
return tf.reduce_sum(res, axis=2)
def get_weights(self):
return self.variables.get_flat()
def set_weights(self, weights):
return self.variables.set_flat(weights)
def update_epoch(self):
return self.sess.run([self.update_epoch_op])[0]
def setup_cat_qvalues(self, actions_num):
config = {
'name': 'agent',
'inputs': self.input_obs,
'actions_num': actions_num,
}
self.logits = self.network(config, reuse=False)
self.qvalues_c = tf.nn.softmax(self.logits, axis=2)
self.qvalues = self._get_q(self.qvalues_c)
config = {
'name': 'target',
'inputs': self.input_next_obs,
'actions_num': actions_num,
}
self.target_logits = self.network(config, reuse=False)
self.target_qvalues_c = tf.nn.softmax(self.target_logits, axis=2)
self.target_qvalues = self._get_q(self.target_qvalues_c)
if self.config['is_double'] == True:
config = {
'name': 'agent',
'inputs': self.input_next_obs,
'actions_num': actions_num,
}
self.next_logits = tf.stop_gradient(
self.network(config, reuse=True))
self.next_qvalues_c = tf.nn.softmax(self.next_logits, axis=2)
self.next_qvalues = self._get_q(self.next_qvalues_c)
self.weights = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope='agent')
self.target_weights = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, scope='target')
self.current_action_values = tf.reduce_sum(
tf.expand_dims(tf.one_hot(self.actions_ph, actions_num), -1) *
self.logits,
reduction_indices=(1, ))
if self.config['is_double'] == True:
self.next_selected_actions = tf.argmax(self.next_qvalues, axis=1)
self.next_selected_actions_onehot = tf.one_hot(
self.next_selected_actions, actions_num)
self.next_state_values_target = tf.stop_gradient(
tf.reduce_sum(
tf.expand_dims(self.next_selected_actions_onehot, -1) *
self.target_qvalues_c,
reduction_indices=(1, )))
else:
self.next_selected_actions = tf.argmax(self.target_qvalues, axis=1)
self.next_selected_actions_onehot = tf.one_hot(
self.next_selected_actions, actions_num)
self.next_state_values_target = tf.stop_gradient(
tf.reduce_sum(
tf.expand_dims(self.next_selected_actions_onehot, -1) *
self.target_qvalues_c,
reduction_indices=(1, )))
self.proj_dir_ph = tf.placeholder(tf.float32,
shape=[None, self.atoms_num],
name='best_proj_dir')
log_probs = tf.nn.log_softmax(self.current_action_values, axis=1)
if self.is_prioritized:
# we need to return loss to update priority buffer
self.abs_errors = tf.reduce_sum(-log_probs * self.proj_dir_ph,
axis=1) + 1e-5
self.td_loss = self.abs_errors * self.sample_weights_ph
else:
self.td_loss = tf.reduce_sum(-log_probs * self.proj_dir_ph, axis=1)
self.td_loss_mean = tf.reduce_mean(self.td_loss)
def setup_qvalues(self, actions_num):
config = {
'name': 'agent',
'inputs': self.input_obs,
'actions_num': actions_num,
}
self.qvalues = self.network(config, reuse=False)
config = {
'name': 'target',
'inputs': self.input_next_obs,
'actions_num': actions_num,
}
self.target_qvalues = tf.stop_gradient(
self.network(config, reuse=False))
if self.config['is_double'] == True:
config = {
'name': 'agent',
'inputs': self.input_next_obs,
'actions_num': actions_num,
}
self.next_qvalues = tf.stop_gradient(
self.network(config, reuse=True))
self.weights = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope='agent')
self.target_weights = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, scope='target')
self.current_action_qvalues = tf.reduce_sum(
tf.one_hot(self.actions_ph, actions_num) * self.qvalues,
reduction_indices=1)
if self.config['is_double'] == True:
self.next_selected_actions = tf.argmax(self.next_qvalues, axis=1)
self.next_selected_actions_onehot = tf.one_hot(
self.next_selected_actions, actions_num)
self.next_state_values_target = tf.stop_gradient(
tf.reduce_sum(self.target_qvalues *
self.next_selected_actions_onehot,
reduction_indices=[
1,
]))
else:
self.next_state_values_target = tf.stop_gradient(
tf.reduce_max(self.target_qvalues, reduction_indices=1))
self.reference_qvalues = self.rewards_ph + self.gamma_step * self.is_not_done * self.next_state_values_target
if self.is_prioritized:
# we need to return l1 loss to update priority buffer
self.abs_errors = tf.abs(self.current_action_qvalues -
self.reference_qvalues) + 1e-5
# the same as multiply gradients later (other way is used in different examples over internet)
self.td_loss = tf.losses.huber_loss(
self.current_action_qvalues,
self.reference_qvalues,
reduction=tf.losses.Reduction.NONE) * self.sample_weights_ph
self.td_loss_mean = tf.reduce_mean(self.td_loss)
else:
self.td_loss_mean = tf.losses.huber_loss(
self.current_action_qvalues,
self.reference_qvalues,
reduction=tf.losses.Reduction.MEAN)
self.reg_loss = tf.losses.get_regularization_loss()
self.td_loss_mean += self.reg_loss
self.learning_rate = self.config['learning_rate']
if self.env_name:
self.train_step = tf.train.AdamOptimizer(
self.learning_rate * self.lr_multiplier).minimize(
self.td_loss_mean, var_list=self.weights)
def save(self, fn):
self.saver.save(self.sess, fn)
def restore(self, fn):
self.saver.restore(self.sess, fn)
def _reset(self):
self.states.clear()
if self.env_name:
self.state = self.env.reset()
self.total_reward = 0.0
self.total_shaped_reward = 0.0
self.step_count = 0
def get_qvalues(self, state):
return self.sess.run(self.qvalues, {self.obs_ph: state})
def get_action(self, state, epsilon=0.0):
if np.random.random() < epsilon:
action = self.env.action_space.sample()
else:
qvals = self.get_qvalues([state])
action = np.argmax(qvals)
return action
def play_steps(self, steps, epsilon=0.0):
done_reward = None
done_shaped_reward = None
done_steps = None
steps_rewards = 0
cur_gamma = 1
cur_states_len = len(self.states)
# always break after one
while True:
if cur_states_len > 0:
state = self.states[-1][0]
else:
state = self.state
action = self.get_action(state, epsilon)
new_state, reward, is_done, _ = self.env.step(action)
#reward = reward * (1 - is_done)
self.step_count += 1
self.total_reward += reward
shaped_reward = self.rewards_shaper(reward)
self.total_shaped_reward += shaped_reward
self.states.append([new_state, action, shaped_reward])
if len(self.states) < steps:
break
for i in range(steps):
sreward = self.states[i][2]
steps_rewards += sreward * cur_gamma
cur_gamma = cur_gamma * self.gamma
next_state, current_action, _ = self.states[0]
self.exp_buffer.add(self.state, current_action, steps_rewards,
new_state, is_done)
self.state = next_state
break
if is_done:
done_reward = self.total_reward
done_steps = self.step_count
done_shaped_reward = self.total_shaped_reward
self._reset()
return done_reward, done_shaped_reward, done_steps
def load_weigths_into_target_network(self):
self.sess.run(self.assigns_op)
def sample_batch(self, exp_replay, batch_size):
obs_batch, act_batch, reward_batch, next_obs_batch, is_done_batch = exp_replay.sample(
batch_size)
return {
self.obs_ph: obs_batch,
self.actions_ph: act_batch,
self.rewards_ph: reward_batch,
self.is_done_ph: is_done_batch,
self.next_obs_ph: next_obs_batch
}
def sample_prioritized_batch(self, exp_replay, batch_size, beta):
obs_batch, act_batch, reward_batch, next_obs_batch, is_done_batch, sample_weights, sample_idxes = exp_replay.sample(
batch_size, beta)
batch = {
self.obs_ph: obs_batch,
self.actions_ph: act_batch,
self.rewards_ph: reward_batch,
self.is_done_ph: is_done_batch,
self.next_obs_ph: next_obs_batch,
self.sample_weights_ph: sample_weights
}
return [batch, sample_idxes]
def train(self):
mem_free_steps = 0
last_mean_rewards = -100500
epoch_num = 0
frame = 0
update_time = 0
play_time = 0
start_time = time.time()
total_time = 0
self.load_weigths_into_target_network()
for _ in range(0, self.config['num_steps_fill_buffer']):
self.play_steps(self.steps_num, self.epsilon)
steps_per_epoch = self.config['steps_per_epoch']
num_epochs_to_copy = self.config['num_epochs_to_copy']
batch_size = self.config['batch_size']
lives_reward = self.config['lives_reward']
episodes_to_log = self.config['episodes_to_log']
frame = 0
play_time = 0
update_time = 0
rewards = []
shaped_rewards = []
steps = []
losses = deque([], maxlen=100)
while True:
epoch_num = self.update_epoch()
t_play_start = time.time()
self.epsilon = self.epsilon_processor(frame)
self.beta = self.beta_processor(frame)
for _ in range(0, steps_per_epoch):
reward, shaped_reward, step = self.play_steps(
self.steps_num, self.epsilon)
if reward != None:
self.game_lengths.append(step)
self.game_rewards.append(reward)
#shaped_rewards.append(shaped_reward)
t_play_end = time.time()
play_time += t_play_end - t_play_start
# train
frame = frame + steps_per_epoch
t_start = time.time()
if self.is_categorical:
if self.is_prioritized:
batch, idxes = self.sample_prioritized_batch(
self.exp_buffer, batch_size=batch_size, beta=self.beta)
next_state_vals = self.sess.run(
[self.next_state_values_target], batch)[0]
projected = self.categorical.distr_projection(
next_state_vals, batch[self.rewards_ph],
batch[self.is_done_ph], self.gamma**self.steps_num)
batch[self.proj_dir_ph] = projected
_, loss_t, errors_update, lr_mul = self.sess.run([
self.train_step, self.td_loss_mean, self.abs_errors,
self.lr_multiplier
], batch)
self.exp_buffer.update_priorities(idxes, errors_update)
else:
batch = self.sample_batch(self.exp_buffer,
batch_size=batch_size)
next_state_vals = self.sess.run(
[self.next_state_values_target], batch)[0]
projected = self.categorical.distr_projection(
next_state_vals, batch[self.rewards_ph],
batch[self.is_done_ph], self.gamma**self.steps_num)
batch[self.proj_dir_ph] = projected
_, loss_t, lr_mul = self.sess.run([
self.train_step, self.td_loss_mean, self.lr_multiplier
], batch)
else:
if self.is_prioritized:
batch, idxes = self.sample_prioritized_batch(
self.exp_buffer, batch_size=batch_size, beta=self.beta)
_, loss_t, errors_update, lr_mul = self.sess.run([
self.train_step, self.td_loss_mean, self.abs_errors,
self.lr_multiplier
], batch)
self.exp_buffer.update_priorities(idxes, errors_update)
else:
batch = self.sample_batch(self.exp_buffer,
batch_size=batch_size)
_, loss_t, lr_mul = self.sess.run([
self.train_step, self.td_loss_mean, self.lr_multiplier
], batch)
losses.append(loss_t)
t_end = time.time()
update_time += t_end - t_start
total_time += update_time
if frame % 1000 == 0:
mem_free_steps += 1
if mem_free_steps == 10:
mem_free_steps = 0
tr_helpers.free_mem()
sum_time = update_time + play_time
print('frames per seconds: ', 1000 / (sum_time))
self.writer.add_scalar('performance/fps', 1000 / sum_time,
frame)
self.writer.add_scalar('performance/upd_time', update_time,
frame)
self.writer.add_scalar('performance/play_time', play_time,
frame)
self.writer.add_scalar('losses/td_loss', np.mean(losses),
frame)
self.writer.add_scalar('info/lr_mul', lr_mul, frame)
self.writer.add_scalar('info/lr', self.learning_rate * lr_mul,
frame)
self.writer.add_scalar('info/epochs', epoch_num, frame)
self.writer.add_scalar('info/epsilon', self.epsilon, frame)
if self.is_prioritized:
self.writer.add_scalar('beta', self.beta, frame)
self.logger.log_stat("whirl/performance/fps", 1000 / sum_time,
self.num_env_steps_train)
self.logger.log_stat("whirl/performance/upd_time", update_time,
self.num_env_steps_train)
self.logger.log_stat("whirl/performance/play_time", play_time,
self.num_env_steps_train)
self.logger.log_stat("losses/td_loss", np.mean(losses),
self.num_env_steps_train)
self.logger.log_stat("whirl/info/last_lr",
self.learning_rate * lr_mul,
self.num_env_steps_train)
self.logger.log_stat("whirl/info/lr_mul", lr_mul,
self.num_env_steps_train)
self.logger.log_stat("whirl/epochs", epoch_num,
self.num_env_steps_train)
self.logger.log_stat("whirl/epsilon", self.epsilon,
self.num_env_steps_train)
update_time = 0
play_time = 0
num_games = len(self.game_rewards)
if num_games > 10:
d = num_games / lives_reward
mean_rewards = np.sum(self.game_rewards) / d
mean_lengths = np.sum(self.game_lengths) / d
self.writer.add_scalar('rewards/mean', mean_rewards, frame)
self.writer.add_scalar('rewards/time', mean_rewards,
total_time)
self.writer.add_scalar('episode_lengths/mean',
mean_lengths, frame)
self.writer.add_scalar('episode_lengths/time',
mean_lengths, total_time)
self.logger.log_stat("whirl/rewards/mean",
np.asscalar(mean_rewards),
self.num_env_steps_train)
self.logger.log_stat("whirl/rewards/time", mean_rewards,
total_time)
self.logger.log_stat("whirl/episode_lengths/mean",
np.asscalar(mean_lengths),
self.num_env_steps_train)
self.logger.log_stat("whirl/episode_lengths/time",
mean_lengths, total_time)
if mean_rewards > last_mean_rewards:
print('saving next best rewards: ', mean_rewards)
last_mean_rewards = mean_rewards
self.save("./nn/" + self.config['name'] + 'ep=' +
str(epoch_num) + 'rew=' + str(mean_rewards))
if last_mean_rewards > self.config['score_to_win']:
print('network won!')
return last_mean_rewards, epoch_num
#clear_output(True)
# adjust agent parameters
if frame % num_epochs_to_copy == 0:
self.load_weigths_into_target_network()
if epoch_num >= self.max_epochs:
print('Max epochs reached')
self.save("./nn/" + 'last_' + self.config['name'] + 'ep=' +
str(epoch_num) + 'rew=' + str(
np.sum(self.game_rewards) * lives_reward /
len(self.game_rewards)))
return last_mean_rewards, epoch_num