def test_deepq():
"""
test DeepQ on atari
"""
logger.configure()
set_global_seeds(SEED)
env = make_atari(ENV_ID)
env = bench.Monitor(env, logger.get_dir())
env = wrap_atari_dqn(env)
model = DQN(env=env,
policy=CnnPolicy,
learning_rate=1e-4,
buffer_size=10000,
exploration_fraction=0.1,
exploration_final_eps=0.01,
train_freq=4,
learning_starts=10000,
target_network_update_freq=1000,
gamma=0.99,
prioritized_replay=True,
prioritized_replay_alpha=0.6,
checkpoint_freq=10000)
model.learn(total_timesteps=NUM_TIMESTEPS)
env.close()
del model, env
def train_DQN(env, out_dir, seed=None, **kwargs):
# Logs will be saved in log_dir/monitor.csv
global output_dir
output_dir = out_dir
log_dir = os.path.join(out_dir, 'log')
os.makedirs(log_dir, exist_ok=True)
env = Monitor(env, log_dir + '/', allow_early_resets=True)
global n_steps, best_mean_reward
best_mean_reward, n_steps = -np.inf, 0
policy = kwargs['policy']
n_timesteps = kwargs['n_timesteps']
del kwargs['policy']
del kwargs['n_timesteps']
model = DQN(policy,
env,
verbose=1,
tensorboard_log=os.path.join(log_dir, 'tb'),
full_tensorboard_log=True,
checkpoint_path=log_dir,
seed=seed,
**kwargs)
model.learn(total_timesteps=n_timesteps, callback=log_callback)
return model
def main(args):
"""
Run a trained model for the vrep
:param args: (ArgumentParser) the input arguments
"""
env = gym.make('vrep-v0')
model = DQN.load("./model/cartpole_model3000.pkl", env)
model_index = [x for x in range(3000, 54001, 3000)]
#model = DQN.load("./model/bk1/cartpole_model3000.pkl", env, tensorboard_log='./log')
for index in model_index:
model = DQN.load("./model/cartpole_model" + str(index) + ".pkl", env)
all_espisode_rewards = []
all_try_steps = []
all_success = []
for i in range(100):
obs, done = env.reset(), False
episode_rew = 0
try_steps = 0
while not done:
action, _ = model.predict(obs)
obs, rew, done, _ = env.step(action)
# print('action: %d, reward: %f, counts: %d' %(action, rew, env.counts))
episode_rew += rew
try_steps += 1
if done and rew > 0:
success = 1
else:
success = 0
print("Episode reward: %d, try_steps:%d, if success:%d " %
(episode_rew, try_steps, success))
all_espisode_rewards.append(episode_rew)
if success:
all_try_steps.append(try_steps)
else:
all_try_steps.append(30)
all_success.append(success)
# No render is only used for automatic testing
if args.no_render:
break
# average reward, steps, and success rate
ave_rew = sum(all_espisode_rewards) / len(all_espisode_rewards)
ave_step = sum(all_try_steps) / len(all_try_steps)
success_rate = sum(all_success) / len(all_success)
print("ave_rew:%f, ave_step:%f, success_rate:%f " %
(ave_rew, ave_step, success_rate))
res = ','.join([str(ave_rew), str(ave_step), str(success_rate)])
with open('./log/test_results.txt', 'a') as f:
f.write(res + '\n')
env.reset_simulation()
env.close()
def main():
setupOdrive()
# sanity check
if drive is None:
print("Failed to initialize Odrive. Exiting...")
exit()
_setCurrent(0)
nnBal = DQN.load("deepq_policy_bal.zip.quiet")
nnUp = DQN.load("deepq_policy_up.zip.quiet")
initObs()
obs = getObs()
reward = 0
i = 0
while i < LOOP_COUNT:
i += 1
mark = time.time()
action = 0
if checkArmVel(obs) and checkPoleVel(obs):
policy = None
if abs(obs[2]) > cm.deg2Rad(10):
policy = "spin-up"
action, _ = nnUp.predict(np.array(obs))
else:
policy = "balance"
action, _ = nnBal.predict(np.array(obs))
current = computeCurrent(action)
print("%s\taction: %d\tcurrent: %.2f" % (policy, action, current))
setCurrent(current, obs)
diff = time.time() - mark
render = True
while diff < LOOP_COUNT:
obs = getObs(render)
cur = getCurrent()
render = False
buf.append({"target": current_target, "obs": obs, "i": cur})
diff = time.time() - mark
reward += math.cos(abs(obs[2])) - abs(current / MAX_CURRENT) * 0.001
print("Episode reward: %.1f\tdata len: %d" % (reward, len(buf_current)))
_setCurrent(0)
stamp = int(time.time() / 1)
fname = "data/current_data_3_b_" + str(stamp) + ".json"
with open(fname, 'w') as f:
json.dump(buf, f)
print("Wrote behavioral data to: " + fname)
def main(args):
"""
Run a trained model for the mountain car problem
:param args: (ArgumentParser) the input arguments
"""
env = gym.make("MountainCar-v0")
model = DQN.load("mountaincar_model.zip", env)
while True:
obs, done = env.reset(), False
episode_rew = 0
while not done:
if not args.no_render:
env.render()
# Epsilon-greedy
if np.random.random() < 0.02:
action = env.action_space.sample()
else:
action, _ = model.predict(obs, deterministic=True)
obs, rew, done, _ = env.step(action)
episode_rew += rew
print("Episode reward", episode_rew)
# No render is only used for automatic testing
if args.no_render:
break
def main():
"""
Run the atari test
"""
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--env',
help='environment ID',
default='BreakoutNoFrameskip-v4')
parser.add_argument('--seed', help='RNG seed', type=int, default=0)
parser.add_argument('--prioritized', type=int, default=1)
parser.add_argument('--dueling', type=int, default=1)
parser.add_argument('--prioritized-replay-alpha', type=float, default=0.6)
parser.add_argument('--num-timesteps', type=int, default=int(10e6))
parser.add_argument('--checkpoint-freq', type=int, default=10000)
parser.add_argument('--checkpoint-path', type=str, default=None)
args = parser.parse_args()
logger.configure()
set_global_seeds(args.seed)
env = make_atari(args.env)
env = bench.Monitor(env, logger.get_dir())
env = wrap_atari_dqn(env)
policy = partial(CnnPolicy, dueling=args.dueling == 1)
model = DQN(
env=env,
policy=policy,
learning_rate=1e-4,
buffer_size=10000,
exploration_fraction=0.1,
exploration_final_eps=0.01,
train_freq=4,
learning_starts=10000,
target_network_update_freq=1000,
gamma=0.99,
prioritized_replay=bool(args.prioritized),
prioritized_replay_alpha=args.prioritized_replay_alpha,
checkpoint_freq=args.checkpoint_freq,
checkpoint_path=args.checkpoint_path,
)
model.learn(total_timesteps=args.num_timesteps)
env.close()
def mainHybrid():
env = fed.FurutaEnvTorqueDeepq(cm.RUN, render=True)
#env.setRender(True)
modelBal = DQN.load(POLICY_PATH + "deepq_policy_bal_nn.zip", env)
modelUp = DQN.load(POLICY_PATH + "deepq_policy_up_nn.zip", env)
while True:
obs, done = env.reset(), False
episode_rew = 0
while not done:
if abs(obs[2]) > cm.deg2Rad(cm.ANGLE_TERMINAL_MIN_D):
action, _ = modelUp.predict(obs)
else:
action, _ = modelBal.predict(obs)
obs, rew, done, _ = env.step(action)
speedCheck(obs)
episode_rew += rew
print("Episode reward: %.3f" % (episode_rew))
def __init__(self, env=None):
"""Return only every `skip`-th frame"""
super(FireOtherSkipEnv, self).__init__(env)
# most recent raw observations (for max pooling across time steps)
self._obs_buffer = deque(maxlen=2)
self.env = wrap_atari_dqn(self.env)
model_output='/home/jingjia16/stable-baselines/scripts/deepq_pong.zip'
if os.path.exists(model_output):
self.model = DQN.load(model_output)
else:
print("failed to load the model")
def test_DQN(env, out_dir, seed=None, **kwargs):
# Logs will be saved in log_dir/monitor.csv
log_dir = os.path.join(out_dir, 'log')
os.makedirs(log_dir, exist_ok=True)
model = DQN.load(os.path.join(out_dir, 'best_model.pkl'))
# Evaluate the trained agent
mean_reward = evaluate(env, model, num_steps=10000)
return
def main(self, args):
"""
Train and save the DQN model, for the cartpole problem
:param args: (ArgumentParser) the input arguments
"""
#env = gym.make('CartPole-v1')
#model = DQN(MlpPolicy, env, verbose=1)
#model.load("cartpole_model.pkl")
model = DQN(env=env,
policy=CustomPolicy,
learning_rate=1e-3,
buffer_size=50000,
exploration_fraction=0.01,
exploration_final_eps=0.02,
verbose=1)
model.learn(total_timesteps=args.max_timesteps, callback=self.callback)
print("Saving model to cartpole_model.pkl")
model.save("cartpole_model.pkl")
#if __name__ == '__main__':
#parser = argparse.ArgumentParser(description="Train DQN on cartpole")
#parser.add_argument('--max-timesteps', default=100000000, type=int, help="Maximum number of timesteps")
#args = parser.parse_args()
#main(args)
def test(env, agent, filepath):
model = DQN.load(filepath)
obs = env.reset()
episode_count = 0
while (True):
if(episode_count == settings.testing_nb_episodes_per_model):
exit(0)
else:
action, _states = model.predict(obs)
obs, rewards, dones, info = env.step(action)
if(dones == True):
env.reset()
episode_count += 1
def load_model(nw_type, log_dir, env_name, log_name, save_name, model=None):
log_dir = log_dir + env_name + '/'
if log_name is not None:
log_dir = log_dir + log_name + '/'
if nw_type.lower() == 'deepq' or nw_type.lower(
) == 'deep' or nw_type.lower() == 'dqn':
return DQN.load(log_dir + 'DeepQ/' + save_name + '/' + save_name +
'.pkl')
elif nw_type.lower() == 'ppo' or nw_type.lower(
) == 'ppo2' or nw_type.lower() == 'ppo1':
return PPO2.load(log_dir + 'PPO/' + save_name + '/' + save_name +
'.pkl')
elif nw_type.lower() == 'ddqn':
model.load_weights(log_dir + 'DoubleDQN/' + save_name + '/' +
save_name + '.h5f')
return model
def setup(difficulty_level='default', env_name = "AirSimEnv-v42"):
config = tf.ConfigProto()
config.gpu_options.per_process_gpu_memory_fraction = 0.6
config.gpu_options.allow_growth = True
set_session(tf.Session(config=config))
env = gym.make(env_name)
env.init_again(eval("settings."+difficulty_level+"_range_dic"))
# Vectorized environments allow to easily multiprocess training
# we demonstrate its usefulness in the next examples
vec_env = DummyVecEnv([lambda: env]) # The algorithms require a vectorized environment to run
agent = DQN(MlpPolicy, vec_env, verbose=1)
env.set_model(agent)
return env, agent
def mainUp():
env = fed.FurutaEnvTorqueDeepqUp(cm.RUN, render=True)
#env.setRender(True)
model = DQN.load(POLICY_PATH + "deepq_policy_up.zip", env)
while True:
obs, done = env.reset(), False
episode_rew = 0
count = 0
while not done:
action, _ = model.predict(obs)
obs, rew, done, _ = env.step(action)
speedCheck(obs)
episode_rew += rew
count += 1
print("Episode average reward: %.3f" % (episode_rew / count))
def main():
"""
Run a trained model for the pong problem
"""
env = gym.make("PongNoFrameskip-v4")
env = deepq.wrap_atari_dqn(env)
model = DQN.load("pong_model.pkl", env)
while True:
obs, done = env.reset(), False
episode_rew = 0
while not done:
env.render()
action, _ = model.predict(obs)
obs, rew, done, _ = env.step(action)
episode_rew += rew
print("Episode reward", episode_rew)
def main(args):
"""
Train and save the DQN model, for the cartpole problem
:param args: (ArgumentParser) the input arguments
"""
# env = gym.make("CartPole-v0")
# model = DQN(
# env=env,
# policy=MlpPolicy,
# verbose=1,
# learning_rate=1e-3,
# buffer_size=50000,
# exploration_fraction=0.1,
# exploration_final_eps=0.02,
# tensorboard_log='./log',
# )
# model.learn(total_timesteps=args.max_timesteps, callback=callback)
# print("Saving model to cartpole_model.pkl")
# model.save("cartpole_model.pkl")
# env = Vrep_Env()
env = gym.make('vrep-v0')
model = DQN(
env=env,
gamma=0.95,
policy=MlpPolicy,
#policy=CustomPolicy,
verbose=1,
learning_rate=1e-4,
buffer_size=50000, #5000
train_freq=1,
learning_starts=100,
batch_size=64, # 32
checkpoint_freq=3000,
checkpoint_path='./model/',
target_network_update_freq=300,
prioritized_replay=True,
exploration_fraction=0.1,
exploration_final_eps=0.02,
tensorboard_log='./log',
)
# path = './model/'
# model = DQN.load(path+'bk2_16/cartpole_model6000.pkl', env, tensorboard_log='./log')
model.learn(total_timesteps=args.max_timesteps,
callback=callback,
log_interval=30)
print("Saving model to slab_installing_model.pkl")
model.save("slab_installing_model.pkl")
def main(args):
"""
Run a trained model for the cartpole problem
:param args: (ArgumentParser) the input arguments
"""
env = gym.make("CartPole-v0")
model = DQN.load("cartpole_model.zip", env)
while True:
obs, done = env.reset(), False
episode_rew = 0
while not done:
if not args.no_render:
env.render()
action, _ = model.predict(obs)
obs, rew, done, _ = env.step(action)
episode_rew += rew
print("Episode reward", episode_rew)
# No render is only used for automatic testing
if args.no_render:
break
def train(env, fname):
env.setRender(False)
env.reset()
start = time.time()
model = DQN(
env=env,
policy=CustomPolicy,
learning_rate=1e-3,
buffer_size=50000,
exploration_fraction=0.1,
exploration_final_eps=0.02
)
model.learn(total_timesteps=STEPS, callback=callback)
# save trained model
model.save(fname)
print("Duration: %.1f" % ((time.time() - start)/60))
def main(args):
"""
Train and save the DQN model, for the cartpole problem
:param args: (ArgumentParser) the input arguments
"""
env = gym.make("CartPole-v0")
model = DQN(
env=env,
policy=MlpPolicy,
learning_rate=1e-3,
buffer_size=50000,
exploration_fraction=0.1,
exploration_final_eps=0.02,
)
model.learn(total_timesteps=args.max_timesteps, callback=callback)
print("Saving model to cartpole_model.pkl")
model.save("cartpole_model.pkl")
def main(args):
"""
Train and save the DQN model, for the mountain car problem
:param args: (ArgumentParser) the input arguments
"""
env = gym.make("MountainCar-v0")
# using layer norm policy here is important for parameter space noise!
model = DQN(policy=CustomPolicy,
env=env,
learning_rate=1e-3,
buffer_size=50000,
exploration_fraction=0.1,
exploration_final_eps=0.1,
param_noise=True)
model.learn(total_timesteps=args.max_timesteps)
print("Saving model to mountaincar_model.pkl")
model.save("mountaincar_model.pkl")
if __name__ == '__main__':
env = SumoEnvironment(net_file='nets/2way-single-intersection/single-intersection.net.xml',
route_file='nets/2way-single-intersection/single-intersection-vhvh.rou.xml',
out_csv_name='outputs/2way-single-intersection/dqn-vhvh2-stable-mlp-bs',
single_agent=True,
use_gui=True,
num_seconds=100000,
time_to_load_vehicles=120,
max_depart_delay=0,
phases=[
traci.trafficlight.Phase(32000, 32000, 32000, "GGrrrrGGrrrr"),
traci.trafficlight.Phase(2000, 2000, 2000, "yyrrrryyrrrr"),
traci.trafficlight.Phase(32000, 32000, 32000, "rrGrrrrrGrrr"),
traci.trafficlight.Phase(2000, 2000, 2000, "rryrrrrryrrr"),
traci.trafficlight.Phase(32000, 32000, 32000, "rrrGGrrrrGGr"),
traci.trafficlight.Phase(2000, 2000, 2000, "rrryyrrrryyr"),
traci.trafficlight.Phase(32000, 32000, 32000, "rrrrrGrrrrrG"),
traci.trafficlight.Phase(2000, 2000, 2000, "rrrrryrrrrry")
])
model = DQN(
env=env,
policy=MlpPolicy,
learning_rate=1e-3,
buffer_size=50000,
exploration_fraction=0.1,
exploration_final_eps=0.02
)
model.learn(total_timesteps=100000)
def test(env, agent, filepath):
model = DQN.load(filepath)
obs = env.reset()
for i in range(settings.testing_nb_episodes_per_model):
action, _states = model.predict(obs)
obs, rewards, dones, info = env.step(action)
df = df.reset_index()
total_timesteps = 20_000
iterations = 50_000
env = DummyVecEnv([lambda: SepsisEnv(df)])
models = [
PPO2(MlpPolicy, env, verbose=0),
PPO2(MlpLstmPolicy, env, nminibatches=1, verbose=0),
PPO2(MlpLnLstmPolicy, env, nminibatches=1, verbose=0),
A2C(MlpPolicy, env, lr_schedule='constant'),
A2C(MlpLstmPolicy, env, lr_schedule='constant'),
DQN(
env=env,
policy=DQN_MlpPolicy,
learning_rate=1e-3,
buffer_size=50000,
exploration_fraction=0.1,
exploration_final_eps=0.02,
),
DQN(
env=env,
policy=LnMlpPolicy,
learning_rate=1e-3,
buffer_size=50000,
exploration_fraction=0.1,
exploration_final_eps=0.02,
)
]
for model in models:
env = DummyVecEnv([lambda: SepsisEnv(df)])
env = gym.make('CarRacing-v0')
env = DiscreteCarRacing(env)
env = DummyVecEnv([lambda: env
]) # The algorithms require a vectorized environment to run
"""DDPG Algorithm """
# Add some param noise for exploration
# param_noise = AdaptiveParamNoiseSpec(initial_stddev=0.2, desired_action_stddev=0.2)
# n_actions = env.action_space.shape[-1]
# action_noise = OrnsteinUhlenbeckActionNoise(mean=np.zeros(n_actions), sigma=float(0.5) * np.ones(n_actions))
# model1 = DDPG(policy=LnMlpPolicy, gamma=0.995, actor_lr=1e-4, critic_lr=1e-3, env=env, param_noise=param_noise, verbose=1)
# model1 = DDPG(policy=LnMlpPolicy, gamma=0.995, actor_lr=1e-4, critic_lr=1e-3, env=env, action_noise=action_noise, verbose=1)
"""DQN Algorithm """
model2 = DQN(env=env,
policy=MlpPolicy,
learning_rate=1e-3,
buffer_size=50000,
exploration_fraction=0.1,
exploration_final_eps=0.02,
verbose=1)
"""PPO Algorithm """
# model3 = PPO1(policy=MlpPolicy, gamma=0.995, optim_batchsize=32, env=env, verbose=0)
"""Get result """
# note : All model has same named policy but from different algorithm. Can not run three models at once.
# ddpg_pr_result = evaluate(model1, num_steps=10000)
# ddpg_ou_result = evaluate(model1.5, num_steps=10000)
dqn_result = evaluate(model2, num_steps=10000)
# ppo_result = evaluate(model3, num_steps=10000)
def train_deep(env_name='CartPole-v1',
steps=10000,
lr=5e-4,
exploration_fraction=0.1,
exploration_final_eps=0.02,
log_dir='./Logs/',
log_name=None):
"""
Wrapper for training a network with DQN
:param env_name: The name of the environment to load [String]
:param steps: The number of time-steps to train for [Int]
:param exploration_fraction: The exploration rate for the algorithm [double or whatever]
:param exploration_final_eps: The final exploration rate after decay [double or whatever]
:param lr: The learning rate for the algorithm [double or whatever]
:param log_dir: The base log folder [String]
:param log_name: Puts the logs in a subdir of this name [String]
"""
# Generates a folder hierarchy for the logging:
if log_name is None:
log_dir = log_dir + env_name + '/' + 'DeepQ/deep_{0:.0E}'.format(
lr) + '/'
else:
log_dir = log_dir + env_name + '/' + log_name + '/' + 'DeepQ/deep_{0:.0E}'.format(
lr) + '/'
init_logging(log_dir)
# Generates an environment for the algorithm to train against
env = DummyVecEnv([
lambda: Monitor(gym.make(env_name), log_dir, allow_early_resets=True)
])
# Sets up a modified callback funtion to be able to handle saving etc. (Not really needed)
best_mean_reward, n_steps, hist_rew = -np.inf, 0, 0
def callback(_locals, _globals):
"""
Callback called at each step (for DQN an others) or after n steps (see ACER or PPO2)
:param _locals: (dict)
:param _globals: (dict)
"""
nonlocal n_steps, best_mean_reward, hist_rew
# Print stats every 1000 calls
if (n_steps + 1) % 5 == 0:
# Evaluate policy performance
x, y = ts2xy(load_results(log_dir), 'timesteps')
if len(x) > 0:
# mean_rew_plot(y, len(x))
hist_rew = y.copy()
mean_reward = np.mean(y[-100:])
if (n_steps + 1) % 100 == 0:
print(x[-1], 'timesteps')
print(
"Best mean reward: {:.2f} - Last mean reward per episode: {:.2f}"
.format(best_mean_reward, mean_reward))
# New best model, you could save the agent here
if mean_reward > best_mean_reward:
best_mean_reward = mean_reward
# Example for saving best model
print("Saving new best model")
_locals['self'].save(log_dir +
"/deep_{0:.0E}.pkl".format(lr))
n_steps += 1
return False
# Creates the training model etc.
dqn_nw = DQN('MlpPolicy',
env,
learning_rate=lr,
exploration_fraction=exploration_fraction,
exploration_final_eps=exploration_final_eps,
checkpoint_freq=2000,
learning_starts=1000,
target_network_update_freq=500)
# Starts the training:
dqn_nw.learn(total_timesteps=steps, callback=callback)