class ExplorerWorker:
def __init__(self, name, actions):
self.e_model = EncoderModel()
self.act_model = ActorModel(actions)
self.c_model = CriticModel()
self.t_model = A2CGaeTrain(self.e_model,
self.act_model,
self.c_model,
name,
lr=2e-6)
self.name = name
self.steps = 0
self.log_dir = 'logs/{}'.format(name)
self.summary_writer = tf.summary.create_file_writer(self.log_dir)
self.boards = {
'reward': tf.keras.metrics.Mean('reward_board', dtype=tf.float32),
'actor_loss': tf.keras.metrics.Mean('train_loss',
dtype=tf.float32),
'critic_loss': tf.keras.metrics.Mean('train_loss_c',
dtype=tf.float32),
}
def set_weights(self, model):
self.e_model.update_model(model.e_model)
self.act_model.update_model(model.act_model)
self.c_model.update_model(model.c_model)
def get_weights(self):
return self.e_model.trainable_variables + self.act_model.trainable_variables + self.c_model.trainable_variables
def set_board(self, name, value):
self.boards[name](value)
def get_board(self, name):
return self.boards[name].result()
def get_values(self, state):
embeding = self.e_model(
tf.convert_to_tensor(state[None, ...], dtype=tf.float32))
action, p, log_policy = self.act_model.get_action(embeding)
value = np.asarray(self.c_model(embeding))[0][0]
return action, p, log_policy, value
def train(self, states, actions, rewards, next_states, values,
old_log_policies, R, adv, dones):
return self.t_model.get_gradients(states, actions, rewards,
next_states, values,
old_log_policies, R, adv, dones)
def update_model(self, cg, ag, eg):
self.t_model.apply_grads(cg, ag, eg)
def __init__(self, name, actions):
self.e_model = EncoderModel()
self.act_model = ActorModel( actions )
self.c_model = CriticModel()
self.t_model = A2CGaeTrain(self.e_model, self.act_model, self.c_model, name, lr=2e-6)
self.name = name
self.steps = 0
self.log_dir = 'logs/{}'.format(name)
self.summary_writer = tf.summary.create_file_writer(self.log_dir)
self.boards = {
'reward': tf.keras.metrics.Mean('reward_board', dtype=tf.float32),
'actor_loss': tf.keras.metrics.Mean('train_loss', dtype=tf.float32),
'critic_loss': tf.keras.metrics.Mean('train_loss_c', dtype=tf.float32)
}
cv.waitKey(1)
size = ( 84, 84 )
bs = 1024
buffer_memory = 5000
t_steps = 1000
n_episodes = 100000
num_local_steps = 100
num_epochs = 5
gamma = 0.9
tau = 1.0
env = create_train_env( 1, 1, RIGHT_ONLY, random=True )
e_model = EncoderModel()
act_model = ActorModel( env.action_space.n )
c_model = CriticModel()
t_model = A2CGaeTrain( e_model, act_model, c_model )
fig, axs = plt.subplots(1)
fig.suptitle('Vertically stacked subplots')
curr_states, state_ = env.reset()
curr_episode = 0
ct = 0
while True:
curr_episode += 1
actions = []
values = []
states = []
_states = []