def __init__(self, state_size, action_size, random_seed, agent_size=1):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
random_seed (int): random seed
"""
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(random_seed)
self.agent_size = agent_size
self.local_actor = ActorNetwork(state_size, action_size,
random_seed).to(device)
self.target_actor = ActorNetwork(state_size, action_size,
random_seed).to(device)
self.local_critic = CriticNetwork(state_size, action_size,
random_seed).to(device)
self.target_critic = CriticNetwork(state_size, action_size,
random_seed).to(device)
self.opt_actor = optim.Adam(self.local_actor.parameters(), lr=LR_ACTOR)
self.opt_critic = optim.Adam(self.local_critic.parameters(),
lr=LR_CRITIC,
weight_decay=WEIGHT_DECAY)
self.noise = OUNoise(action_size, random_seed)
# Replay memory
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE,
random_seed)
def __init__(self, state_size, action_size, n_agents, seed):
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(seed)
self.stacked_state_size = state_size * n_agents
self.stacked_action_size = action_size * n_agents
# Actor networks
self.actor_local = ActorNetwork(state_size, action_size,
seed).to(device)
self.actor_target = ActorNetwork(state_size, action_size,
seed).to(device)
self.actor_optimizer = optim.Adam(self.actor_local.parameters(),
lr=ACTOR_LR)
# Critic networks
self.critic_local = CriticNetwork(self.stacked_state_size,
self.stacked_action_size,
seed).to(device)
self.critic_target = CriticNetwork(self.stacked_state_size,
self.stacked_action_size,
seed).to(device)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(),
lr=CRITIC_LR)
# OUNoise
self.exploration_noise = OUNoise(action_size, seed)
def __init__(self, state_size, action_size, agent_id):
self.state_size = state_size
self.action_size = action_size
self.seed = args['seed']
self.device = args['device']
#self.args = args
# Q-Network
self.actor_network = ActorNetwork(state_size, action_size).to(self.device)
self.actor_target = ActorNetwork(state_size, action_size).to(self.device)
self.actor_optimizer = optim.Adam(self.actor_network.parameters(), lr=args['LR_ACTOR'])
#Model takes too long to run --> load model weights from previous run (took > 24hours on my machine)
#if not agent_id:
# self.actor_network.load_state_dict(torch.load(args['agent_p0_path']), strict=False)
# self.actor_target.load_state_dict(torch.load(args['agent_p0_path']), strict=False)
#else:
# self.actor_network.load_state_dict(torch.load(args['agent_p1_path']), strict=False)
# self.actor_target.load_state_dict(torch.load(args['agent_p1_path']), strict=False)
# Replay memory
self.memory = ReplayBuffer(action_size, args['BUFFER_SIZE'], args['BATCH_SIZE'], self.device, self.seed)
# Noise process
self.noise = OUNoise(action_size, self.seed)
# Initialize time step (for updating every UPDATE_EVERY steps)
self.t_step = 0
self.mCriticLoss = 0
self.actorLoss = 0
def __init__(self, state_size, obs_size, action_size, num_agents):
super(DDPGAgent, self).__init__()
#self.actor = Network(in_actor, hidden_in_actor, hidden_out_actor, out_actor, actor=True).to(device)
#self.critic = Network(in_critic, hidden_in_critic, hidden_out_critic, 1).to(device)
#self.target_actor = Network(in_actor, hidden_in_actor, hidden_out_actor, out_actor, actor=True).to(device)
#self.target_critic = Network(in_critic, hidden_in_critic, hidden_out_critic, 1).to(device)
self.actor = ActorNetwork(obs_size, action_size).to(device)
self.critic = CriticNetwork(state_size,
action_size * num_agents).to(device)
self.target_actor = ActorNetwork(obs_size, action_size).to(device)
self.target_critic = CriticNetwork(state_size,
action_size * num_agents).to(device)
#self.noise = OUNoise(out_actor, scale=1.0 )
self.noise = OUNoise(action_size, scale=1.0)
# initialize targets same as original networks
hard_update(self.target_actor, self.actor)
hard_update(self.target_critic, self.critic)
self.actor_optimizer = Adam(self.actor.parameters(), lr=LR_ACTOR)
self.critic_optimizer = Adam(self.critic.parameters(),
lr=LR_CRITIC,
weight_decay=WEIGHT_DECAY)
def __init__(self, state_size, action_size, seed):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
seed (int): random seed
"""
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(seed)
# create the local and target actor networks
self.actor_local = ActorNetwork(state_size, action_size,
seed).to(device)
self.actor_target = ActorNetwork(state_size, action_size,
seed).to(device)
# create the local and target critic networks
self.critic_local = CriticNetwork(state_size, action_size,
seed).to(device)
self.critic_target = CriticNetwork(state_size, action_size,
seed).to(device)
# optimizers for local actor and critic
self.actor_optimizer = optim.Adam(self.actor_local.parameters(), lr=LR)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(),
lr=LR,
weight_decay=0.0)
# MSE loss for updating the critic
self.critic_loss_function = nn.MSELoss()
# ensure that the local and target networks are initialized with the same random weights
for target_param, param in zip(self.actor_target.parameters(),
self.actor_local.parameters()):
target_param.data.copy_(param.data)
for target_param, param in zip(self.critic_target.parameters(),
self.critic_local.parameters()):
target_param.data.copy_(param.data)
# Replay memory
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, seed)
# Initialize time step (for updating every UPDATE_EVERY steps)
self.t_step = 0
# init the noise class to sample from
self.noise = GaussianNoise(self.action_size)
def create_model(context):
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
context.sess = tf.InteractiveSession(config=config)
context.actor = ActorNetwork(
context.sess, [len(context.assets), context.n,
len(context.features)],
len(context.assets) + 1, context.actor_learning_rate, context.tau,
context.minibatch_size)
context.critic = CriticNetwork(
context.sess, [len(context.assets), context.n,
len(context.features)],
len(context.assets) + 1, context.critic_learning_rate, context.tau,
context.gamma, context.actor.get_num_trainable_vars())
# Inicializar las variables de Tensorflow
context.sess.run(tf.global_variables_initializer())
context.saver = tf.train.Saver()
# Inicializar los pesos de las redes objetivo
context.actor.update_target_network()
context.critic.update_target_network()
def __init__(self, state_size, action_size, random_seed, learning_rate,
noise, device):
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(random_seed)
self.learning_rate = learning_rate
self.actor_local = ActorNetwork(state_size, action_size,
random_seed).to(device)
self.actor_target = ActorNetwork(state_size, action_size,
random_seed).to(device)
hard_update(self.actor_target, self.actor_local)
self.actor_optimizer = optim.Adam(self.actor_local.parameters(),
lr=self.learning_rate)
self.noise = noise
self.device = device
def __init__(self, state_size, action_size, memory, seed=None):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
seed (int): random seed
"""
self.state_size = state_size
self.action_size = action_size
if seed is not None:
self.seed = seed
# create the local and target actor networks
self.actor_local = ActorNetwork(state_size, action_size, seed).to(device)
self.actor_target = ActorNetwork(state_size, action_size, seed).to(device)
# create the local and target critic networks
self.critic_local = CriticNetwork(state_size, action_size, seed).to(device)
self.critic_target = CriticNetwork(state_size, action_size, seed).to(device)
# optimizers for local actor and critic
self.actor_optimizer = optim.Adam(self.actor_local.parameters(), lr=LR)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(), lr=LR, weight_decay=0.0)
# MSE loss for updating the critic
# self.critic_loss_function = nn.MSELoss()
self.critic_loss_function = nn.SmoothL1Loss()
# copy the local networks weights to the target network
self.copy_weights_from_local_to_target()
# Replay memory
self.memory = memory
# Initialize time step (for updating every UPDATE_EVERY steps)
self.t_step = 0
# init the noise class to sample from
self.noise = GaussianNoise(self.action_size)
def __init__(self,
state_size,
action_size,
num_agents,
hidden_in_actor=512,
hidden_out_actor=256,
lr_actor=1e-4,
hidden_in_critic=512,
hidden_out_critic=256,
lr_critic=3e-4,
weight_decay_critic=0,
seed=1,
device='cpu'):
super(DDPGAgent, self).__init__()
self.device = device
# Actor
self.actor = ActorNetwork(state_size, hidden_in_actor,
hidden_out_actor, action_size,
seed).to(device)
self.target_actor = ActorNetwork(state_size, hidden_in_actor,
hidden_out_actor, action_size,
seed).to(device)
self.actor_optimizer = Adam(self.actor.parameters(), lr=lr_actor)
# Target
self.critic = CriticNetwork(state_size, action_size, num_agents,
hidden_in_critic, hidden_out_critic,
seed).to(device)
self.target_critic = CriticNetwork(state_size, action_size, num_agents,
hidden_in_critic, hidden_out_critic,
seed).to(device)
self.critic_optimizer = Adam(self.critic.parameters(),
lr=lr_critic,
weight_decay=weight_decay_critic)
# Noise
self.noise = OUNoise(action_size, seed, scale=1.0)
# initialize targets same as original networks
hard_update(self.target_actor, self.actor)
hard_update(self.target_critic, self.critic)
def build_actor_critic(sess, env):
with tf.variable_scope("model", reuse=None):
with tf.name_scope("actor"):
actor = ActorNetwork(sess, env, config, is_training=False)
with tf.name_scope("critic"):
critic = CriticNetwork(sess, env, config, is_training=False)
sess.run(tf.global_variables_initializer())
return actor, critic
def build_actor_critic(sess, env, env_eval):
w_init = tflearn.initializations.xavier_initializer()
with tf.variable_scope("model", reuse=None, initializer=w_init):
with tf.name_scope("actor"):
actor = ActorNetwork(sess, env, config, is_training=True)
with tf.name_scope("critic"):
critic = CriticNetwork(sess, env, config, is_training=True)
# if config.noise_std:
# actor_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(env.input_cardin * env.state_cardin),
# sigma=config.noise_std,
# sigma_dec=config.noise_dec)
# else:
# actor_noise = None
with tf.variable_scope("model", reuse=True):
with tf.name_scope("actor"):
actor_eval = ActorNetwork(sess,
env_eval,
config,
is_training=False)
with tf.name_scope("critic"):
critic_eval = CriticNetwork(sess,
env_eval,
config,
is_training=False)
sess.run(tf.global_variables_initializer())
# Initialize target network weights
actor.update_target_network()
critic.update_target_network()
return actor, critic, actor_eval, critic_eval #, actor_noise
def __init__(self,
device,
key,
state_size, action_size, random_seed,
memory, noise,
lr, weight_decay,
checkpoint_folder = './Saved_Model/'):
self.DEVICE = device
self.KEY = key
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(random_seed)
# Hyperparameters
self.LR = lr
self.WEIGHT_DECAY = weight_decay
self.CHECKPOINT_FOLDER = checkpoint_folder
# Actor Network (w/ Target Network)
self.local = ActorNetwork(state_size, action_size, random_seed).to(self.DEVICE)
self.target = ActorNetwork(state_size, action_size, random_seed).to(self.DEVICE)
self.optimizer = optim.Adam(self.local.parameters(), lr=self.LR)
self.checkpoint_full_name = self.CHECKPOINT_FOLDER + 'checkpoint_actor_' + str(self.KEY) + '.pth'
if os.path.isfile(self.checkpoint_full_name):
self.local.load_state_dict(torch.load(self.checkpoint_full_name))
self.target.load_state_dict(torch.load(self.checkpoint_full_name))
# Replay memory
self.memory = memory
# Noise process
self.noise = noise
def build_actor_critic(sess, env):
w_init = tflearn.initializations.xavier_initializer()
with tf.variable_scope("model", reuse=None, initializer=w_init):
with tf.name_scope("actor"):
actor = ActorNetwork(sess, env, config, is_training=True)
with tf.name_scope("critic"):
critic = CriticNetwork(sess, env, config, is_training=True)
sess.run(tf.global_variables_initializer())
# Initialize target network weights
actor.update_target_network()
critic.update_target_network()
return actor, critic
def __init__(self, opt, q_batch):
self.opt = opt
self.q_batch = q_batch
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
self.env = gym.make(self.opt.env)
self.env.seed(self.opt.seed)
self.n_state = self.env.observation_space.shape[0]
self.n_act = self.env.action_space.n
self.actor = ActorNetwork(self.n_state, self.n_act).to(self.device)
self.critic = CriticNetwork(self.n_state).to(self.device)
self.actor.share_memory()
self.critic.share_memory()
self.actor_optimizer = optim.Adam(self.actor.parameters(), lr=opt.lr)
self.critic_optimizer = optim.Adam(self.critic.parameters(), lr=opt.lr)
def __init__(self, path_to_weights, weights_id):
print("Init ANN")
self.init = False
# aviod TF from allocation all GPU mem
# https://stackoverflow.com/questions/34199233/how-to-prevent-tensorflow-from-allocating-the-totality-of-a-gpu-memory
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
#sess = tf.Session()
K.set_session(sess)
self.path_to_weights = path_to_weights
self.next_weights_id = weights_id
#check if path is valid
if not os.path.isdir(self.path_to_weights):
print("Folder with weights does not exist!")
print(self.path_to_weights)
self.init = False
else:
self.init = True
if self.init:
#create actor network
self.actor = ActorNetwork.ActorNetwork(sess, 5, 2)
#load first weights
print("Loading first weights for ANN from " + str(self.path_to_weights) + "/" + str(self.next_weights_id) + "/actormodel.h5")
try:
self.actor.model.load_weights(
str(self.path_to_weights) + "/" + str(self.next_weights_id) + "/actormodel.h5")
self.next_weights_id = self.next_weights_id + 1
except:
print("Cannot find the weight (.h5) file")
print(str(self.path_to_weights) + "/" + str(self.next_weights_id) + "/actormodel.h5")
self.init = False
def __init__(self, opt, actor_id, q_trace, learner):
self.opt = opt
self.q_trace = q_trace
self.learner = learner
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
self.env = gym.make(self.opt.env)
self.env.seed(self.opt.seed + actor_id)
self.n_state = self.env.observation_space.shape[0]
self.n_act = self.env.action_space.n
self.n_episodes = 0
self.n_steps = 0
self.gamma = opt.gamma
# epsilon
self.eps_greedy = 0.4 ** (1 + actor_id * 7 / (opt.n_actors - 1)) \
if opt.n_actors > 1 else 0.4
# モデル
self.actor = ActorNetwork(self.n_state, self.n_act).to(self.device)
self.critic = CriticNetwork(self.n_state).to(self.device)
def __init__(self, opt, actor_id, q_trace, learner):
self.opt = opt
self.q_trace = q_trace
self.learner = learner
self.env = gym.make(self.opt.env)
self.env_state = self.env.reset()
self.n_state = self.env.observation_space.shape[0]
self.n_act = self.env.action_space.n
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
## パラメータ
self.batch_size = opt.batch_size
self.roll_out_n_steps = opt.roll_out_n_steps
self.gamma = opt.gamma
self.eps_greedy = 0.4 ** (1 + actor_id * 7 / (opt.n_actors - 1)) \
if opt.n_actors > 1 else 0.4
self.n_episodes = 0
self.actor = ActorNetwork(self.n_state, self.n_act).to(self.device) # ActorNetwork
self.critic = CriticNetwork(self.n_state).to(self.device) # CriticNetwork
def main(args):
with tf.Session() as sess:
env = gym.make(args['env'])
np.random.seed(int(args['random_seed']))
tf.set_random_seed(int(args['random_seed']))
env.seed(int(args['random_seed']))
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.shape[0]
action_bound = env.action_space.high
# Ensure action bound is symmetric
# assert (env.action_space.high == -env.action_space.low)
actor = ActorNetwork(sess, state_dim, action_dim, action_bound,
float(args['actor_lr']), float(args['tau']),
int(args['minibatch_size']))
critic = CriticNetwork(sess, state_dim, action_dim,
float(args['critic_lr']), float(args['tau']),
float(args['gamma']),
actor.get_num_trainable_vars())
actor_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(action_dim))
if args['train']:
if not os.path.exists(args['save_dir']):
os.makedirs(args['save_dir'])
with open(os.path.join(args['save_dir'], 'config.json'), 'w') as f:
json.dump(args, f, indent=2)
train(sess, env, args, actor, critic, actor_noise)
else:
# ddpg = []
# indexes = [e for e in range(400) if e % 10 == 9]
# indexes = [0] + indexes
indexes = [399]
num_test_tasks = 100
buckets = 1
successes = []
directory = args['to_pickle']
for index in indexes:
# times = []
task_success = []
saver = tf.train.Saver()
saver.restore(
sess, "../final_models/multitask/fixed/{0}/model-{1}.ckpt".
format(directory, index))
for _ in range(buckets):
tasks = env.unwrapped.sample_tasks(num_test_tasks)
# tasks = [{'goal': np.array([0., 0.])} for e in range(num_test_tasks)]
success = 0
for task in tasks:
s = env.reset_task(task)
step = 0
d = False
while not d:
# env.render()
action = actor.predict_target(
np.reshape(s, (1, actor.s_dim)))[0]
step += 1
s, r, d, _ = env.step(action)
if r == 1:
success += 1
# times.append(step)
env.close()
task_success.append(success / num_test_tasks)
successes.append(task_success)
# ddpg.append(times)
# out = [successes, ddpg]
env.close()
if not os.path.exists('./pkls'):
os.makedirs('./pkls')
with open('./pkls/{0}.pkl'.format(args['save_dir']), 'wb') as f:
pickle.dump(successes, f)
import torch
import torch.optim as optim
import torch.nn.functional as F
from torch.distributions import Categorical
from torch.autograd import Variable
import gym
import numpy as np
import matplotlib.pyplot as plt
from model import ActorNetwork, CriticNetwork
actor = ActorNetwork(4, 2)
critic = CriticNetwork(4)
actor_optimizer = optim.Adam(actor.parameters(), lr=1e-4)
critic_optimizer = optim.Adam(critic.parameters(), lr=8e-4)
env = gym.make('CartPole-v0')
GAMMA = 0.99
N_EPISODES = 20000
LOG_STEPS = 100
SAVE_STEPS = 100
def select_action(S):
'''
select action based on currentr state
args:
S: current state
returns:
action to take, log probability of the chosen action
'''
